A method and system for processing performance testing data based on the IR sensor hole of a touch screen

Abstract

This invention relates to a method and system for processing performance testing data based on the IR sensing aperture of a touchscreen. The method includes: firstly, receiving detection signal data transmitted by a performance testing sensor arranged in the area surrounding the IR sensing aperture of the touchscreen within a preset sampling period, including infrared light intensity response signals and channel identification information of different sensing channels; then, extracting features from the detection signal data to generate time-frequency domain features and statistical features from time and statistical dimensions, forming a multi-dimensional feature set; inputting the multi-dimensional feature set into a preset machine learning algorithm model for classification and recognition, obtaining a performance status classification result as the performance testing result; if the testing result indicates an abnormal state, calling historical fault case data from an AI knowledge base to construct an abnormal diagnosis model, performing fault type matching and root cause tracing analysis on the abnormal features, obtaining a fault source label, and realizing efficient detection and fault diagnosis of the performance of the IR sensing aperture of the touchscreen.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method and system for processing performance detection data based on the IR sensing aperture of a touchscreen. Background Technology

[0002] In the manufacturing process of touchscreens, the IR sensor aperture is a crucial component, and its performance directly impacts the overall performance and user experience of the touchscreen. To ensure that the IR sensor aperture meets performance requirements, performance testing is essential. Traditional methods for testing touchscreen IR sensor aperture performance mainly rely on manual inspection or simple electrical parameter measurements. Manual inspection suffers from inefficiency and strong subjectivity; different inspectors may have different judgment standards, making it difficult to guarantee the accuracy and consistency of the test results. Simple electrical parameter measurements, on the other hand, can only obtain limited information and cannot comprehensively and deeply reflect the performance status of the IR sensor aperture.

[0003] With the development of technology, some sensor-based detection methods have been gradually applied. These methods collect relevant data by arranging sensors around the IR sensing hole of the touch screen. However, in terms of data processing, they often only perform simple statistical analysis and lack the ability to mine the deep features of the data. In terms of fault diagnosis, most existing methods are based on experience-based judgment or preset fixed rules, which are difficult to deal with complex and ever-changing fault conditions and cannot quickly and accurately locate the fault type and root cause.

[0004] Therefore, existing technologies suffer from low testing efficiency, poor accuracy, and insufficient fault diagnosis capabilities in the performance testing and fault diagnosis of touch screen IR sensing holes, failing to meet the demands of modern touch screen manufacturing for high-quality and high-efficiency testing. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for processing performance detection data based on the IR sensing hole of a touch screen.

[0006] This invention provides a performance testing data processing method based on a touchscreen IR sensor hole, applied to a performance testing data processing system based on a touchscreen IR sensor hole. The method includes: receiving detection signal data transmitted by a performance testing sensor arranged in the area surrounding the touchscreen IR sensor hole within a preset sampling period, wherein the detection signal data includes infrared light intensity response signals collected by the performance testing sensor under different sensing channels and corresponding channel identification information; performing feature extraction processing on the detection signal data, generating time-frequency domain features by analyzing the change law of signal waveform in the time dimension, and generating statistical features by calculating signal distribution characteristics in the statistical dimension, to obtain a multi-dimensional feature set containing time-frequency domain features and statistical features; inputting the multi-dimensional feature set into a preset machine learning algorithm model for classification and recognition processing, generating a performance status classification result by pattern matching of the correlation between features, and using the performance status classification result as the performance testing result of the touchscreen IR sensor hole; if the performance testing result indicates that the touchscreen IR sensor hole is in an abnormal state, calling historical fault case data stored in a preset AI knowledge base to construct an abnormal diagnosis model, and using the abnormal diagnosis model to perform fault type matching and root cause tracing analysis on the abnormal features contained in the performance testing result to obtain a fault source tracing label for indicating the cause of the fault.

[0007] The present invention also provides a performance testing data processing system based on a touchscreen IR sensing hole, comprising: a memory for storing program instructions and data; and a processor for coupling with the memory to execute the instructions in the memory to implement the method described above.

[0008] The present invention also provides a computer storage medium comprising instructions that, when executed on a processor, implement the above-described method.

[0009] This invention enables efficient and accurate detection and fault diagnosis of the performance of the IR sensing aperture of a touchscreen, greatly improving the quality control level in the touchscreen production process. Specifically, it receives detection signal data transmitted from performance detection sensors arranged around the IR sensing aperture of the touchscreen, covering infrared light intensity response signals and channel identification information for different sensing channels. Feature extraction is performed on the detection signal data, generating time-frequency domain features and statistical features from both time and statistical dimensions, forming a multi-dimensional feature set. This multi-dimensional feature extraction method can more comprehensively capture the intrinsic characteristics of the signal and, compared to single-dimensional analysis, more accurately reflect the performance status of the IR sensing aperture of the touchscreen.

[0010] Furthermore, the multi-dimensional feature set is input into a pre-defined machine learning algorithm model for classification and recognition. By performing pattern matching on the correlation between features, performance status classification results are generated. Utilizing the powerful pattern recognition capabilities of machine learning, the performance status of the touchscreen IR sensor hole can be determined efficiently and accurately, avoiding the subjectivity and inaccuracy of manual judgment. When the performance detection results indicate that the touchscreen IR sensor hole is in an abnormal state, historical fault case data stored in the pre-defined AI knowledge base is called to construct an anomaly diagnosis model. Fault type matching and root cause tracing analysis are performed on the abnormal features to obtain fault source labels. This process fully utilizes the value of historical data, enabling rapid and accurate location of fault types and causes, providing support for timely repair measures, effectively reducing downtime and costs in the production process, improving production efficiency and product quality, and realizing comprehensive monitoring and management of the touchscreen IR sensor hole performance. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the steps of a performance testing data processing method based on a touchscreen IR sensing hole, provided in an embodiment of the present invention.

[0012] Figure 2 This is a structural block diagram of a performance detection data processing system based on a touchscreen IR sensing hole, provided in an embodiment of the present invention. Detailed Implementation

[0013] Please see Figure 1 , Figure 1 This is a flowchart illustrating a performance testing data processing method based on a touchscreen IR sensor hole provided in an embodiment of the present invention. The method is applied to a performance testing data processing system based on a touchscreen IR sensor hole and may further include steps 110-140.

[0014] Step 110: Receive detection signal data transmitted by the performance detection sensor arranged in the area around the IR sensing hole of the touch screen within a preset sampling period. The detection signal data includes infrared light intensity response signals collected by the performance detection sensor under different sensing channels and corresponding channel identification information.

[0015] In this embodiment of the invention, during the production process of the touchscreen IR sensing hole, performance detection sensors are arranged around the sensing hole to monitor its performance in real time. These sensors continuously collect relevant detection signal data according to a preset sampling period. For example, on one touchscreen production line, multiple performance detection sensors are arranged for the IR sensing hole of a specific model of touchscreen. Each sensor is responsible for a different sensing channel. Within the preset sampling period, the sensor transmits back the collected infrared light intensity response signal and the corresponding channel identification information. By receiving this data (including the infrared light intensity response under different sensing channels) and combining it with the channel identification information, the specific sensing channel corresponding to each signal can be identified.

[0016] Step 120: Perform feature extraction processing on the detected signal data. Generate time-frequency domain features by analyzing the changing patterns of the signal waveform in the time dimension, and generate statistical features by calculating the signal distribution characteristics in the statistical dimension, thereby obtaining a multi-dimensional feature set containing time-frequency domain features and statistical features.

[0017] In this embodiment of the invention, after receiving the detected signal data, feature extraction processing is required to extract more valuable information. In the time dimension, the variation patterns of the signal waveform are analyzed in depth. For example, the fluctuations of the signal at different time points are observed, and the distribution patterns of peak points, valley points, and zero-crossing points of the signal waveform are identified. These patterns reflect the dynamic characteristics of the signal over time, thereby generating waveform structure features. Simultaneously, the time interval and amplitude change rate between adjacent extreme points are calculated, yielding time interval features and change rate features, which together constitute the time-frequency domain analysis results. In the statistical dimension, the central tendency, dispersion, and morphological characteristics of the signal sampling points are analyzed, generating center position features, discrete distribution features, and distribution morphology features, which are combined to form statistical features. Finally, the time-frequency domain analysis results and statistical features are aligned in feature dimensions, ensuring that different types of features have the same feature vector length, thus obtaining a multi-dimensional feature set containing time-frequency domain feature vectors and statistical feature vectors.

[0018] Step 121: Perform signal segmentation processing on the detection signal data. Based on the channel identification information, divide the infrared light intensity response signals of different sensing channels into independent signal sequence units. Each signal sequence unit contains the light intensity response value of continuous sampling points and the corresponding timestamp information.

[0019] In this embodiment of the invention, to analyze the detection signal data in more detail, signal segmentation processing is required. Based on channel identification information, the infrared light intensity response signals of different sensing channels are distinguished and divided into independent signal sequence units. For example, for detection signal data containing multiple sensing channels, the channel identification information can be used to split it into multiple independent signal sequences. Each signal sequence unit contains the light intensity response values ​​of consecutive sampling points and the corresponding timestamp information. The timestamp information clearly identifies the acquisition time of each light intensity response value, which can be used to combine the time dimension in subsequent analysis. This segmentation processing makes subsequent feature extraction more accurate and targeted, and can better reflect the signal characteristics of each sensing channel.

[0020] Step 122: Perform waveform morphology analysis on each signal sequence unit in the time dimension. Generate waveform structure features by identifying the distribution patterns of peak points, valley points, and zero-crossing points of the signal waveform. Generate time interval features by calculating the time interval between adjacent extreme points. Generate rate of change features by calculating the amplitude change rate between adjacent extreme points. Combine the waveform structure features, time interval features, and rate of change features as the time-frequency domain analysis results.

[0021] In this embodiment of the invention, after signal segmentation, waveform morphology analysis is performed on each signal sequence unit in the time dimension. First, the distribution patterns of peak points, valley points, and zero-crossing points of the signal waveform are identified. Peak points represent the maximum value reached by the signal at a certain moment, valley points represent the minimum value, and zero-crossing points are the positions where the signal crosses zero. The distribution patterns of these points can reflect the periodicity, stability, and other characteristics of the signal, thereby generating waveform structure features. Next, the time interval between adjacent extreme points is calculated, which reflects the rate of signal change, resulting in time interval features. Simultaneously, the amplitude change rate between adjacent extreme points is calculated, which reflects the drastic change of the signal between adjacent extreme points, generating rate of change features. Finally, the waveform structure features, time interval features, and rate of change features are combined as the time-frequency domain analysis results. These results can provide important time dimension information for subsequent performance testing.

[0022] Step 123: Calculate the distribution characteristics of each signal sequence unit in the statistical dimension. Generate the central location feature by analyzing the central tendency of the signal sampling points, generate the discrete distribution feature by analyzing the discreteness of the signal sampling points, generate the distribution morphology feature by analyzing the morphological features of the signal sampling points, and combine the central location feature, the discrete distribution feature, and the distribution morphology feature into statistical features.

[0023] In this embodiment of the invention, the distribution characteristics of each signal sequence unit are calculated on a statistical dimension. For the central tendency analysis of the signal sampling points, the central position of the signal is determined, for example, by calculating the mean, median, etc., generating a central position feature that reflects the overall level of the signal. For the discreteness analysis of the signal sampling points, the dispersion of the signal around the central position is examined, such as by calculating the variance, standard deviation, etc., generating a discrete distribution feature. This feature reflects the stability of the signal. For the morphological feature analysis of the signal sampling points, the distribution shape of the signal is observed, such as whether it is symmetrical or skewed, generating a distribution morphology feature. Finally, the central position feature, discrete distribution feature, and distribution morphology feature are combined into statistical features, which can describe the distribution characteristics of the signal from different perspectives.

[0024] Step 124: Perform feature dimension alignment processing on the time-frequency domain analysis results and the statistical features. By unifying the feature description dimensions, different types of features have the same feature vector length, resulting in a multi-dimensional feature set containing time-frequency domain feature vectors and statistical feature vectors.

[0025] In this embodiment of the invention, after obtaining the time-frequency domain analysis results and statistical features, feature dimension alignment processing is required. Since time-frequency domain features and statistical features are obtained from analyzing signals from different perspectives, their feature description dimensions may differ. To effectively integrate and utilize these features, it is necessary to unify the feature description dimensions, ensuring that different types of features have the same feature vector length. For example, by appropriately adjusting and transforming the time-frequency domain features and statistical features, ensuring consistency in feature vector length, a multi-dimensional feature set containing both time-frequency domain feature vectors and statistical feature vectors can be obtained. This multi-dimensional feature set contains more comprehensive signal feature information.

[0026] Step 130: Input the multi-dimensional feature set into a preset machine learning algorithm model for classification and recognition processing. Generate performance status classification results by performing pattern matching on the correlation between features. Use the performance status classification results as the performance detection results of the touch screen IR sensing hole.

[0027] In this embodiment of the invention, after obtaining the multi-dimensional feature set, it is input into a preset machine learning algorithm model for classification and recognition processing. The model analyzes the correlation between features in the multi-dimensional feature set and determines the performance status of the touchscreen IR sensor hole through pattern matching. For example, the model learns the correlation patterns between features under different performance states and matches the input multi-dimensional feature set with these patterns. Based on the matching results, a performance status classification result is generated, which can clearly determine whether the touchscreen IR sensor hole is in a normal or abnormal state. Using this performance status classification result as the performance detection result of the touchscreen IR sensor hole can provide an important basis for subsequent production decisions. If the detection result shows an abnormal state, further fault diagnosis and processing are required.

[0028] Step 131: Perform feature correlation analysis on the time-frequency domain feature vectors and statistical feature vectors in the multi-dimensional feature set, generate a feature correlation matrix by calculating the correlation coefficients between different feature dimensions, and select a subset of key features with correlation based on the feature correlation matrix.

[0029] In this embodiment of the invention, before inputting the multi-dimensional feature set into the machine learning algorithm model, feature correlation analysis needs to be performed on the time-frequency domain feature vectors and statistical feature vectors. First, the waveform structure feature dimension and dynamic change feature dimension contained in the time-frequency domain feature vectors, and the center position feature dimension, discrete distribution feature dimension, and distribution morphology feature dimension contained in the statistical feature vectors, are extracted to construct a feature dimension list containing all feature dimensions. Then, data standardization is performed on each feature dimension in the feature dimension list to eliminate dimensional differences between different feature dimensions. Next, the correlation coefficient between any two standardized feature dimensions is calculated, generating a feature correlation matrix with feature dimensions as rows and columns. The element values ​​in this matrix represent the degree of linear correlation between the corresponding two feature dimensions. Afterwards, feature importance is evaluated on the feature correlation matrix, calculating the mutual information value between each feature dimension and the performance status classification result to generate a feature importance score. Feature dimensions with mutual information values ​​greater than a preset threshold are marked as important feature dimensions. Finally, a feature selection model is constructed based on the feature correlation matrix and important feature dimensions. A recursive feature elimination algorithm is used to select a subset of key features that are correlated and have a redundancy of less than a set redundancy value from the important feature dimensions. This subset of key features contains fewer feature dimensions than the original feature dimensions and can more effectively represent the information of the multi-dimensional feature set.

[0030] Step 1311: Extract the waveform structure feature dimension and dynamic change feature dimension contained in the time-frequency domain feature vector, extract the center position feature dimension, discrete distribution feature dimension and distribution shape feature dimension contained in the statistical feature vector, and construct a feature dimension list containing all feature dimensions.

[0031] In this embodiment of the invention, to perform feature correlation analysis, it is necessary to first define the specific feature dimensions included in the multi-dimensional feature set. Waveform structure feature dimensions and dynamic change feature dimensions are extracted from the time-frequency domain feature vectors; these dimensions reflect the signal's characteristics in the time dimension. Center position feature dimensions, discrete distribution feature dimensions, and distribution pattern feature dimensions are extracted from the statistical feature vectors; these dimensions reflect the signal's characteristics in the statistical dimension. These extracted feature dimensions are combined to construct a feature dimension list containing all feature dimensions. This list covers all key feature dimensions in the multi-dimensional feature set, providing a foundation for subsequent feature analysis and processing. By constructing the feature dimension list, the composition of the multi-dimensional feature set can be accurately and completely defined.

[0032] Step 1312: Perform data standardization on each feature dimension in the feature dimension list by converting the feature values ​​into standard scores with a mean of zero and a variance of one, so as to eliminate the dimensional differences between different feature dimensions.

[0033] In this embodiment of the invention, since the dimensions of different feature dimensions may differ, it can affect subsequent feature correlation analysis. Therefore, it is necessary to perform data standardization on each feature dimension in the feature dimension list. By converting the feature values ​​of each feature dimension into standard scores with a mean of zero and a variance of one, the dimensional differences between different feature dimensions are eliminated. For example, for the feature values ​​in a certain feature dimension, through appropriate transformation, its mean becomes zero and its variance becomes one. Thus, the feature values ​​of different feature dimensions have the same scale, facilitating subsequent calculations and comparisons. Data standardization can improve the accuracy and reliability of feature correlation analysis and avoid analytical errors caused by dimensional differences.

[0034] Step 1313: Calculate the correlation coefficient between any two feature dimensions after standardization, and generate a feature correlation matrix with the feature dimensions as rows and columns. The element values ​​in the feature correlation matrix represent the degree of linear correlation between the corresponding two feature dimensions.

[0035] In this embodiment of the invention, after data standardization, the correlation coefficient between any two standardized feature dimensions is calculated. The correlation coefficient measures the degree of linear correlation between two feature dimensions. By calculating the pairwise correlation coefficients between all feature dimensions, a feature association matrix is ​​generated with the feature dimensions as rows and columns. Each element value in the matrix represents the degree of linear correlation between the corresponding two feature dimensions. For example, an element value close to 1 indicates a strong positive linear correlation between the two feature dimensions; an element value close to -1 indicates a strong negative linear correlation; and an element value close to 0 indicates a weak linear correlation between the two feature dimensions. This feature association matrix can intuitively display the linear correlation between the various feature dimensions in a multi-dimensional feature set.

[0036] Step 1314: Perform feature importance assessment on the feature association matrix. Generate a feature importance score by calculating the mutual information value between each feature dimension and the performance status classification result. Mark feature dimensions with mutual information values ​​greater than a preset threshold as important feature dimensions.

[0037] In this embodiment of the invention, after obtaining the feature correlation matrix, its feature importance needs to be evaluated. The importance of each feature dimension to the performance status classification is determined by calculating the mutual information value between each feature dimension and the performance status classification result. The mutual information value measures the dependency between two variables; in this embodiment, it is used to measure the degree of association between each feature dimension and the performance status classification result. For each feature dimension, its mutual information value with the performance status classification result is calculated, and a feature importance score is generated. Feature dimensions with mutual information values ​​greater than a preset threshold are marked as important feature dimensions. These important feature dimensions have a strong influence on the performance status classification result and will be given priority in subsequent feature selection and model training. Through feature importance evaluation, the most valuable feature dimensions for performance status classification can be selected, improving the accuracy and efficiency of the model.

[0038] Step 1315: Construct a feature selection model based on the feature correlation matrix and important feature dimensions, and use a recursive feature elimination algorithm to select a subset of key features that are correlated and have a redundancy lower than a set redundancy value from the important feature dimensions. The key feature subset contains fewer feature dimensions than the original feature dimensions.

[0039] In this embodiment of the invention, after determining the important feature dimensions, a feature selection model is constructed based on the feature correlation matrix and the important feature dimensions. This model employs a recursive feature elimination algorithm to filter out a subset of key features from the important feature dimensions. The recursive feature elimination algorithm progressively removes feature dimensions that contribute less to the model until the selected feature subset meets the set conditions. During the filtering process, the correlation and redundancy between features are considered. The selected key feature subset must not only be correlated, i.e., able to reflect important information from the multi-dimensional feature set, but also ensure that the redundancy is below the set redundancy value to avoid containing too much duplicate information. The final key feature subset contains fewer feature dimensions than the original feature dimensions, thus reducing the model's complexity and improving its training efficiency and accuracy.

[0040] Step 132: Input the key feature subset into the feature mapping layer of the machine learning algorithm model, and map the key feature subset from the original feature space to the high-dimensional feature space through nonlinear transformation to generate a high-dimensional feature representation vector.

[0041] In this embodiment of the invention, after selecting a subset of key features, it is input into the feature mapping layer of the machine learning algorithm model. The feature mapping layer performs a nonlinear transformation on the key feature subset, mapping it from the original feature space to a higher-dimensional feature space. In the original feature space, the key feature subset may not fully express the inherent features and patterns of the data. Through nonlinear transformation, the key feature subset can be mapped to a higher-dimensional space, enabling the data to be better separated and distinguished in this higher-dimensional space. For example, by transforming the key feature subset through some nonlinear functions, a high-dimensional feature representation vector is generated. This high-dimensional feature representation vector contains richer feature information, providing more favorable conditions for subsequent feature interaction and classification decisions.

[0042] Step 133: Perform feature combination processing on the high-dimensional feature representation vector through the feature interaction layer of the machine learning algorithm model, and generate an enhanced feature vector by constructing product interaction terms and cross combination terms between features. The enhanced feature vector contains the original features and the interaction relationship information between features.

[0043] In this embodiment of the invention, after obtaining the high-dimensional feature representation vector, the feature interaction layer of the machine learning algorithm model performs feature combination processing on it. First, the high-dimensional feature representation vector is divided into multiple feature groups, each containing feature dimensions with similar physical meaning. The core feature dimension of each group is determined by calculating the variance contribution of each feature dimension within the group. Within each feature group, a product interaction term is constructed based on the core feature dimension and other feature dimensions within the group. This product interaction term represents the synergistic relationship between the core feature and other features. Between different feature groups, the core feature dimensions of each group are extracted to construct cross-combination terms, which represent the correlation between different feature groups. Then, feature scaling processing is performed on the product interaction terms and cross-combination terms, mapping the feature values ​​of the interaction terms to a preset numerical range using a minimum-maximum normalization method. Next, the original high-dimensional feature representation vector, product interaction terms, and cross-combination terms are standardized to eliminate dimensional differences. Finally, the standardized original high-dimensional feature representation vector, product interaction term, and cross combination term are concatenated in the order of feature dimensions to form an enhanced feature vector. This enhanced feature vector contains the original features and the interaction relationship information between features, which can more comprehensively reflect the features and patterns of the data.

[0044] Step 1331: Divide the high-dimensional feature representation vector into multiple feature groups. Each feature group contains feature dimensions with similar physical meaning. Determine the core feature dimension of each group by calculating the variance contribution of each feature dimension within the feature group.

[0045] In this embodiment of the invention, to better perform feature combination processing, it is necessary to group the high-dimensional feature representation vectors. The high-dimensional feature representation vectors are divided into multiple feature groups, where the feature dimensions within each group have similar physical meanings. For example, feature dimensions reflecting the temporal characteristics of a signal are grouped into one group, and feature dimensions reflecting the statistical characteristics of a signal are grouped into another. Then, the variance contribution of each feature dimension within each feature group is calculated. The variance contribution measures the magnitude of each feature dimension's contribution to the entire feature group. By comparing the variance contribution of each feature dimension, the core feature dimension of each group is determined. The core feature dimension is highly representative within the feature group and reflects the main feature information of that group. By dividing the feature groups and determining the core feature dimensions, an accurate and complete structure and direction can be provided for subsequent feature combination processing.

[0046] Step 1332: Within each feature group, construct a product interaction term by combining the core feature dimension with other feature dimensions within the group. The product interaction term represents the synergistic relationship between the core feature and other features.

[0047] In this embodiment of the invention, after determining the core feature dimension of each feature group, a product interaction term is constructed within each feature group. Based on the core feature dimension, it is multiplied with other feature dimensions within the group to construct a product interaction term. This product interaction term can represent the synergistic relationship between the core feature and other features. For example, in a certain feature group, the core feature dimension reflects the main trend of signal change, while other feature dimensions reflect some secondary change features of the signal. By constructing product interaction terms, the mutual influence and synergistic effect between the core feature and other features can be captured, uncovering more data information. Product interaction terms enrich feature representation and provide more comprehensive feature information for subsequent model training.

[0048] Step 1333: Construct cross-combination terms between different feature groups by extracting the core feature dimensions of each group. The cross-combination terms represent the correlation between different feature groups.

[0049] In this embodiment of the invention, after constructing the product interaction terms within each feature group, the focus is on the correlation between different feature groups. Core feature dimensions are extracted from each feature group, and cross-combination terms are constructed using these core feature dimensions. Cross-combination terms can represent the correlation between different feature groups. For example, one feature group reflects the temporal characteristics of a signal, while another feature group reflects the statistical characteristics of the signal. By constructing cross-combination terms, the intrinsic connection between these two different feature groups can be captured, and the interaction information between different dimensions of the data can be mined. Cross-combination terms can further enrich the feature representation and improve the model's ability to understand and process data.

[0050] Step 1334: Perform feature scaling on the product interaction terms and cross combination terms, and map the feature values ​​of the interaction terms to a preset numerical range using the maximum and minimum value normalization method.

[0051] In this embodiment of the invention, after constructing the product interaction terms and cross-combination terms, they need to undergo feature scaling. Since the feature value ranges of the product interaction terms and cross-combination terms may differ, this can affect subsequent model training and processing. Therefore, a maximum-minimum normalization method is used to map the feature values ​​of the interaction terms to a preset numerical range. For example, mapping the feature values ​​to a fixed range ensures that the feature values ​​of different interaction terms have the same scale. This avoids model training instability caused by differences in feature value ranges, improving the model's convergence speed and accuracy. Feature scaling is an important step in feature combination processing, providing more suitable feature data for subsequent feature concatenation and model training.

[0052] Step 1335: Standardize the original high-dimensional feature representation vector, product interaction term, and cross combination term to eliminate dimensional differences; concatenate the standardized original high-dimensional feature representation vector, product interaction term, and cross combination term in the order of feature dimensions to form an enhanced feature vector, wherein the number of dimensions of the enhanced feature vector is the sum of the original feature dimension, the product interaction term dimension, and the cross combination term dimension.

[0053] In this embodiment of the invention, to eliminate the dimensional differences between different features, the original high-dimensional feature representation vector, product interaction terms, and cross-combination terms are standardized. Standardization ensures they have the same scale and distribution, avoiding computational errors and model training instability caused by dimensional differences. Then, the standardized original high-dimensional feature representation vector, product interaction terms, and cross-combination terms are concatenated according to feature dimensions to obtain an enhanced feature vector. The dimension of the enhanced feature vector is the sum of the original feature dimensions, the product interaction term dimensions, and the cross-combination term dimensions. This concatenated enhanced feature vector contains information about the original features and the interactions between features, more comprehensively reflecting the characteristics and patterns of the data.

[0054] Step 134: Input the enhanced feature vector into the classification decision layer of the machine learning algorithm model, and use the multi-classifier ensemble voting mechanism to classify the enhanced feature vector into performance status, generate a probability distribution vector containing different performance status categories, and take the category with the highest probability value in the probability distribution vector as the performance status classification result.

[0055] In this embodiment of the invention, after generating the enhanced feature vector, it is input into the classification decision layer of the machine learning algorithm model. The classification decision layer uses a multi-classifier ensemble voting mechanism to classify the enhanced feature vector according to its performance status. The multi-classifier ensemble voting mechanism combines the results of multiple classifiers for a comprehensive judgment. Each classifier predicts the performance status of the touchscreen IR sensor based on the enhanced feature vector and provides the probability of different performance status categories. The results of these classifiers are integrated and voted on to generate a probability distribution vector containing different performance status categories. This probability distribution vector represents the probability that the touchscreen IR sensor is in different performance status categories. Finally, the category with the highest probability value in the probability distribution vector is taken as the performance status classification result. For example, if the probability value of a certain performance status category is the highest in the probability distribution vector, then that category is determined as the performance status classification result of the touchscreen IR sensor.

[0056] Step 135: Generate a performance test result containing a state category identifier and a confidence index based on the performance state classification result. The confidence index is used to represent the reliability of the classification result.

[0057] In this embodiment of the invention, after obtaining the performance status classification result, a performance detection result including a status category identifier and a confidence index is generated based on the result. The status category identifier clarifies the performance status of the touchscreen IR sensing hole, such as a normal state or an abnormal state. The confidence index is used to represent the reliability of the classification result. For example, the confidence index of the classification result obtained through a multi-classifier ensemble voting mechanism can be determined based on the consistency and accuracy of the voting of each classifier. If the voting results of multiple classifiers are consistent, and these classifiers have high accuracy during training, then the confidence index will be high, indicating that the classification result is relatively reliable. The status category identifier and confidence index in the performance detection result can provide important references for subsequent production decisions. If the confidence index is low, it may be necessary to further check and verify the classification result to ensure the accuracy of the decision.

[0058] It should be understood that the machine learning algorithm model in this embodiment of the invention integrates multiple algorithm techniques, and from the overall architecture perspective, it belongs to an integrated deep learning model. In the feature association analysis stage, correlation coefficient calculation and recursive feature elimination algorithms are used.

[0059] The correlation coefficient was calculated using the Pearson correlation coefficient: [r=frac{sum_{i=1}^{n}(x_i-bar{x})(y_i-bar{y})}{sqrt{sum_{i=1}^{n}(x_i-bar{x})^2sum_{i=1}^{n}(y_i-bar{y})^2}}].

[0060] Where (x_i) and (y_i) are the (i)th sample values ​​of the two feature dimensions, (bar{x}) and (bar{y}) are the sample mean values ​​of the two feature dimensions, and (n) is the number of samples.

[0061] The recursive feature elimination algorithm iteratively removes features that contribute the least to the model until a preset number of features is reached. In the feature mapping layer, a nonlinear transformation is used; in this embodiment of the invention, the ReLU (Rectified Linear Unit) activation function is employed, with the formula: [f(x)=max(0,x)].

[0062] The ReLU function has advantages such as simple computation and fast convergence speed, and can effectively solve the gradient vanishing problem. In the feature interaction layer, interaction relationships between features are mined by constructing product interaction terms and cross-combination terms. In the classification decision layer, a multi-classifier ensemble voting mechanism is adopted; in this embodiment of the invention, support vector machines (SVM), decision trees, and neural networks can be selected as classifiers.

[0063] The parameter settings of the machine learning algorithm model in this embodiment of the invention are as follows: (1) Feature association analysis parameters: When calculating the correlation coefficient, the sample size (n) is the number of samples of all collected detection signal data, which is 1000. The preset mutual information value threshold can be set to 0.2, that is, when the mutual information value between a certain feature dimension and the performance status classification result is greater than 0.2, it is marked as an important feature dimension. In the recursive feature elimination algorithm, the number of features removed in each iteration is set to 2, until the number of key feature subsets selected meets the requirements.

[0064] (2) Feature mapping layer parameters: In the feature mapping layer, the learning rate is set to 0.001, which is a commonly used learning rate value. This ensures model convergence while avoiding instability caused by excessively fast learning speed. The batch size is set to 32, meaning that 32 samples are input each time for training, which can improve training efficiency.

[0065] (3) Feature interaction layer parameters: In the feature interaction layer, the high-dimensional feature representation vector is divided into 5 feature groups, each containing feature dimensions with similar physical meaning. When calculating the variance contribution, the core feature dimension is determined by calculating the proportion of the variance of each feature dimension to the total variance of the feature group. When performing feature scaling, the feature values ​​of the product interaction term and the cross combination term are mapped to the [0, 1] interval using the maximum-minimum normalization method.

[0066] (4) Classification decision layer parameters: In the classification decision layer, the support vector machine uses the radial basis function (RBF) kernel function, with parameters (C) set to 10 and (gamma) set to 0.1. The maximum depth of the decision tree is set to 5 to avoid overfitting. The number of hidden layer neurons in the neural network is set to 64, and the activation function is also the ReLU function.

[0067] The network layer settings of the machine learning algorithm model in this embodiment of the invention are as follows: (1) Feature Correlation Analysis Layer: This layer receives time-frequency domain feature vectors and statistical feature vectors from a multi-dimensional feature set, containing 20 feature dimensions. After feature correlation analysis, 10 key feature subsets are selected.

[0068] (2) Feature mapping layer: The feature mapping layer maps the key feature subset from the original 10-dimensional feature space to a high-dimensional feature space. The high-dimensional feature representation vector has a dimension of 50.

[0069] (3) Feature Interaction Layer: After dividing the high-dimensional feature representation vector into 5 feature groups, the number of dimensions in each feature group is different. The first feature group has a dimension of 10, the second feature group has a dimension of 12, the third feature group has a dimension of 8, the fourth feature group has a dimension of 10, and the fifth feature group has a dimension of 10. By constructing product interaction terms and cross combination terms, the generated enhanced feature vector has a dimension of 80.

[0070] (4) Classification Decision Layer: The classification decision layer receives the enhanced feature vector and outputs a probability distribution vector containing different performance state categories. There are three performance state categories: normal state, mildly abnormal state, and severely abnormal state. Therefore, the probability distribution vector has a dimension of 3.

[0071] Based on the above, the following section further explains how to perform classification processing based on machine learning algorithm models.

[0072] (1) Data reception and feature extraction First, the system receives detection signal data transmitted by performance detection sensors located around the IR sensing aperture of the touchscreen within a preset sampling period. The preset sampling period is 10ms, and each sensing channel collects 1000 infrared light intensity response signal samples. The detection signal data is then segmented, dividing the infrared light intensity response signals of different sensing channels into independent signal sequence units based on channel identification information. In the time dimension, waveform morphology analysis is performed on each signal sequence unit to identify the distribution patterns of peak points, valley points, and zero-crossing points. Under normal circumstances, the peak point of the signal waveform appears at time point t_1=5ms, the valley point appears at time point t_2=8ms, the time interval between adjacent extreme points is 3ms, and the amplitude change rate is 0.5V / ms. When an anomaly occurs, these characteristics change; for example, the peak point arrives earlier (t_1=3ms), the valley point is delayed (t_2=10ms), the time interval between adjacent extreme points becomes 7ms, and the amplitude change rate increases to 1V / ms.

[0073] Statistically, the distribution characteristics of each signal sequence unit are calculated. Under normal circumstances, the mean of the signal sampling points is 1.5V, the standard deviation is 0.2V, and the distribution shape approximates a normal distribution. When anomalies occur, the mean may become 2V, the standard deviation may increase to 0.5V, and the distribution shape may become asymmetrical. Finally, the time-frequency domain analysis results and statistical features are aligned in terms of feature dimensions to obtain a multi-dimensional feature set containing time-frequency domain feature vectors and statistical feature vectors. This multi-dimensional feature set has 20 dimensions.

[0074] (2) Feature association analysis Feature correlation analysis is performed on the time-frequency domain feature vectors and statistical feature vectors in the multi-dimensional feature set. First, the waveform structure feature dimension and dynamic change feature dimension contained in the time-frequency domain feature vectors, and the center position feature dimension, discrete distribution feature dimension, and distribution pattern feature dimension contained in the statistical feature vectors, are extracted to construct a feature dimension list containing all feature dimensions, totaling 20 feature dimensions. Each feature dimension in the feature dimension list is then standardized to have a mean of 0 and a variance of 1.

[0075] The Pearson correlation coefficient between any two standardized feature dimensions is calculated to generate a feature correlation matrix. For example, the correlation coefficient between feature dimensions (x_1) and (x_2) is 0.8, indicating a strong positive linear correlation between them. Feature importance is evaluated on the feature correlation matrix by calculating the mutual information value between each feature dimension and the performance status classification result. The mutual information value between feature dimension (x_3) and the performance status classification result is 0.3, which is greater than the preset threshold of 0.2, and it is marked as an important feature dimension. Finally, a feature selection model is constructed based on the feature correlation matrix and the important feature dimensions. A recursive feature elimination algorithm is used to select a subset of key features from the important feature dimensions that are correlated and have redundancy below a set redundancy value. The number of key feature subsets selected is 10.

[0076] (3) Feature mapping A subset of key features is input into the feature mapping layer of the machine learning algorithm model. The feature mapping layer performs a non-linear transformation on the key feature subset using the ReLU activation function, mapping it from the original 10-dimensional feature space to a higher-dimensional feature space, resulting in a 50-dimensional high-dimensional feature representation vector. During training, the model's weight parameters are updated using the stochastic gradient descent (SGD) algorithm, with a learning rate of 0.001 and a batch size of 32. After multiple iterations of training, the model gradually learns a deep representation of the key feature subset.

[0077] (4) Feature interaction The high-dimensional feature representation vector is divided into 5 feature groups, each containing feature dimensions with similar physical meaning. The variance contribution of each feature dimension within each feature group is calculated to determine the core feature dimension for each group. For example, in the first feature group, feature dimension (x_4) has the largest variance contribution and is therefore selected as the core feature dimension. Within each feature group, product interaction terms are constructed based on the core feature dimension and other feature dimensions within the group. Between different feature groups, the core feature dimensions of each group are extracted to construct cross-combination terms.

[0078] Feature scaling is applied to the product interaction term and the cross combination term, and the feature values ​​of the interaction term are mapped to the [0, 1] interval using the minimax normalization method. The original high-dimensional feature representation vector, product interaction term, and cross combination term are standardized to eliminate dimensional differences. Finally, the standardized original high-dimensional feature representation vector, product interaction term, and cross combination term are concatenated in order of feature dimension to form an enhanced feature vector with 80 dimensions.

[0079] (5) Classification decision The enhanced feature vectors are input into the classification decision layer of the machine learning algorithm model. The classification decision layer employs a multi-classifier ensemble voting mechanism; in this embodiment, Support Vector Machine (SVM), Decision Tree, and Neural Network are selected as classifiers. Each classifier predicts the performance status of the touchscreen IR sensor based on the enhanced feature vectors and provides the probability of different performance status categories. For example, the SVM predicts a normal state with a probability of 0.6, a slightly abnormal state with a probability of 0.3, and a severely abnormal state with a probability of 0.1; the Decision Tree predicts a normal state with a probability of 0.5, a slightly abnormal state with a probability of 0.4, and a severely abnormal state with a probability of 0.1; and the Neural Network predicts a normal state with a probability of 0.7, a slightly abnormal state with a probability of 0.2, and a severely abnormal state with a probability of 0.1.

[0080] The results of these classifiers are integrated and voted on to generate a probability distribution vector containing different performance state categories. The final probability distribution vector is [0.6, 0.3, 0.1]. The category with the highest probability value is taken as the performance state classification result, i.e., the normal state. Based on this result, a performance detection result containing a state category identifier and a confidence index is generated. The confidence index is 0.8, indicating that the classification result is relatively reliable.

[0081] If the detection result indicates an abnormal state, it is necessary to further utilize historical fault case data stored in the pre-set AI knowledge base to construct an anomaly diagnosis model, perform fault type matching and root cause tracing analysis. This design enables accurate classification of the performance status of the touchscreen's IR sensing aperture.

[0082] Step 140: If the performance test result indicates that the touch screen IR sensing hole is in an abnormal state, call the historical fault case data stored in the preset AI knowledge base to construct an abnormal diagnosis model, and use the abnormal diagnosis model to perform fault type matching and root cause tracing analysis on the abnormal features contained in the performance test result to obtain a fault source tracing label used to indicate the cause of the fault.

[0083] In this embodiment of the invention, if the performance test results indicate that the touchscreen IR sensor hole is in an abnormal state, further analysis of the cause of the fault is required. At this time, a fault diagnosis model is constructed by calling historical fault case data stored in a preset AI knowledge base. The AI ​​knowledge base stores a large number of historical fault cases related to the touchscreen IR sensor hole, including historical abnormal feature vectors, corresponding fault type labels, and fault cause analysis reports. The fault diagnosis model is constructed using this historical data, and this model can learn the abnormal feature patterns corresponding to different fault types.

[0084] In the performance testing of the IR sensor hole of a touch screen, sensor failure is one type of failure.

[0085] Taking time-frequency domain anomalies as an example, under normal circumstances, the time-frequency domain characteristics of infrared light intensity response signals exhibit certain regularities. The peak value of a normal signal waveform occurs at time point t1=10ms, the valley value occurs at time point t2=20ms, the time interval between adjacent extreme points is Δt=10ms, and the amplitude change rate is relatively stable, approximately 0.5V / ms. When the sensor malfunctions, these characteristics will change significantly.

[0086] For example, in one fault case, the signal waveform exhibited a sudden spike. The normal peak value was 2V, but after the fault occurred, the peak value instantly reached 5V. Simultaneously, the time intervals between adjacent extreme points became irregular; the previously stable 10ms interval shortened to 5ms in some cases and lengthened to 15ms in others after the fault. The distribution of zero-crossing points also changed. Normally, zero-crossing points are evenly distributed within the signal period, with one zero-crossing point appearing every 10ms. However, during a fault, the distribution of zero-crossing points became chaotic, with potentially three zero-crossing points appearing within a 5ms interval.

[0087] From the perspective of frequency characteristics in the time-frequency domain, the main frequency components of a normal signal are concentrated between 100Hz and 200Hz, and the spectrum graph shows a relatively smooth curve. However, when the sensor malfunctions, some high-frequency noise components will appear in the spectrum graph, with obvious spikes in the 500Hz-1000Hz frequency band. The amplitude of these high-frequency components can even reach half of the amplitude of the normal frequency band.

[0088] In terms of statistical characteristics, the central tendency and dispersion of normal signal sampling points also follow certain patterns. Under normal circumstances, the central location characteristics (such as the mean) of signal sampling points are 1.5V, the discrete distribution characteristics (such as the standard deviation) are 0.2V, and the distribution shape approximates a normal distribution.

[0089] When a sensor malfunctions, the center position characteristic shifts. In one failure case, the mean of the signal sampling points shifted from 1.5V to 2.5V, indicating a change in the overall signal level. Simultaneously, the discrete distribution characteristics increased significantly, with the standard deviation increasing from 0.2V to 0.8V, indicating decreased signal stability and more dispersed data points. The distribution shape also changed from a normal distribution to an asymmetrical, right-skewed distribution, indicating the presence of more large values ​​in the data.

[0090] Furthermore, wiring faults can also affect the performance of the touchscreen's IR sensor. Faulty wiring can lead to unstable signal transmission, resulting in noticeable anomalies in the time-frequency domain characteristics. Normally, the signal waveform is continuous and regular, with zero-crossings evenly distributed at regular time intervals.

[0091] A normal signal's zero-crossing occurs every 10ms, with an 8ms interval between adjacent extreme points. When a line connection malfunctions, the signal waveform exhibits intermittent flat regions. For example, within a certain time period, the signal waveform may show a flat region between t=30ms and t=35ms, with an amplitude almost 0V. Simultaneously, the distribution of zero-crossings is disrupted; the previously uniform distribution becomes chaotic, with only one zero-crossing occurring between t=40ms and t=50ms instead of the normal two.

[0092] In terms of amplitude change rate, under normal circumstances, the amplitude change rate is relatively stable, approximately 0.4V / ms. However, when a fault occurs, the amplitude change rate will fluctuate drastically. At the instant the signal resumes transmission, the amplitude change rate may reach 2V / ms, while in the flat region, the amplitude change rate is 0V / ms.

[0093] Statistically, line connection faults further increase the dispersion of signal sampling points, making the central location characteristics more unstable. Under normal circumstances, the mean of the signal sampling points is 1.2V, and the standard deviation is 0.15V. When a line connection fault occurs, the mean of the signal sampling points fluctuates significantly. Within a certain period, the mean dropped from 1.2V to 0.8V, and then rose to 1.5V. Simultaneously, the standard deviation increased significantly, reaching 0.5V, indicating a substantial increase in data dispersion. The distribution pattern also becomes more complex, no longer exhibiting a clear normal distribution, and may show multiple peaks. For example, in one fault case, the distribution pattern showed a bimodal distribution, with one peak appearing near 0.8V and another peak near 1.5V.

[0094] Furthermore, electromagnetic interference is another important factor affecting the performance of the IR sensor hole in a touchscreen. Electromagnetic interference generates high-frequency noise in the signal, thereby altering the time-frequency domain characteristics of the signal. Under normal circumstances, the main frequency components of the signal are concentrated in the low-frequency band, such as between 100Hz and 200Hz, and the amplitude is relatively stable.

[0095] When subjected to electromagnetic interference, the signal waveform exhibits high-frequency jitter. Normally, the signal amplitude fluctuates between 1V and 2V. After a fault occurs, the signal amplitude, influenced by the high-frequency jitter, fluctuates drastically between 0.5V and 2.5V. The time interval between adjacent extreme points is significantly shortened; the normal 10ms interval is reduced to 2ms-3ms after the interference occurs.

[0096] The spectrum of a normal signal shows a distinct peak in the 100Hz-200Hz frequency band, with an amplitude of approximately 3dB. However, under the influence of electromagnetic interference, multiple high-frequency spikes appear in the spectrum. A spike with an amplitude of 2dB appears in the 500Hz-1000Hz frequency band, and a spike with an amplitude of 1dB also appears in the 1500Hz-2000Hz frequency band.

[0097] In terms of statistical characteristics, electromagnetic interference significantly increases the dispersion of signal sampling points, blurring the central location characteristics. Under normal circumstances, the mean of signal sampling points is 1.8V, and the standard deviation is 0.1V. When subjected to electromagnetic interference, the mean becomes unstable, fluctuating between 1.5V and 2.1V over different time periods. The standard deviation increases significantly, reaching 0.3V, indicating a substantial increase in data dispersion. The distribution pattern becomes more dispersed, no longer exhibiting a clear normal distribution. For example, in one interference case, the distribution pattern shows a relatively broad distribution, with data points evenly distributed between 1V and 2.5V, without a clear central tendency.

[0098] Then, the abnormal features contained in the performance test results are input into the anomaly diagnosis model for fault type matching and root cause tracing analysis. The model calculates the similarity between the abnormal features and historical abnormal feature vectors, identifies the most matching fault type, and combines it with the fault cause analysis report to trace the root cause. Finally, a fault source label is obtained to indicate the cause of the fault, which can clearly identify the type and cause of the fault.

[0099] In this embodiment of the invention, the specific construction method of the anomaly diagnosis model is as follows: In the data preprocessing stage, it is necessary to perform outlier detection, duplicate case processing, data balancing, and dataset partitioning on historical failure case data.

[0100] (1) Outlier detection The AI ​​knowledge base stores 1000 historical fault case data, each case containing 10 time-frequency domain features and 10 statistical features, for a total of 20 feature dimensions. The Isolation Forest algorithm was used for outlier detection, with 100 trees in the Isolation Forest and a sample ratio of 0.2. Anomaly scores were calculated for each sample, and samples with anomaly scores greater than 0.8 were marked as outliers. After detection, 20 samples with anomaly scores greater than 0.8 were found and marked as noise cases, which were then removed.

[0101] (2) Handling of duplicate cases In the remaining 980 cases, the similarity between cases was calculated based on the case's unique identifier and historical anomaly feature vectors. A similarity threshold of 0.9 was set, meaning that if the feature vector similarity between two cases was greater than 0.9, they were considered duplicate cases. After calculation, 50 pairs of duplicate cases were found. The case with the most recent collection timestamp in each pair of duplicate cases was retained, and redundant cases were removed, resulting in a clean case dataset of 930 cases containing a unique and valid case.

[0102] (3) Data balancing process In the cleaning case dataset, the number of cases for different fault types may be imbalanced. There are 300 sensor fault cases, 200 line connection fault cases, and 430 electromagnetic interference fault cases. To achieve a relative balance in the number of cases for different fault types, an oversampling method is used to increase the number of cases for minority fault types. For line connection fault cases, the SMOTE (Synthetic Minority Over-sampling Technique) algorithm is used for oversampling, increasing its number to the same 300 as the sensor fault cases. Ultimately, the cleaning case dataset contains 300 sensor fault cases, 300 line connection fault cases, and 430 electromagnetic interference fault cases.

[0103] (4) Dataset partitioning The balanced cleaning case dataset was randomly divided into a training dataset and a validation dataset in a 7:3 ratio. The training dataset contains 721 cases, and the validation dataset contains 209 cases.

[0104] For model architecture selection and parameter settings, a multi-label classification model architecture can be adopted. The input layer dimension is consistent with the dimension of the historical anomaly feature vector, i.e., 20 dimensions. The output layer dimension is consistent with the number of fault type labels; in this embodiment, there are 3 fault types, so the output layer dimension is 3. Two hidden layers are set: the first hidden layer contains 50 neurons, and the second hidden layer contains 30 neurons. During model training, the learning rate is set to 0.001, the batch size to 32, and the number of training epochs to 100. The cross-entropy loss function is used to calculate the loss value between the model's predicted probability and the actual fault type label, and the model weight parameters are updated through the backpropagation algorithm.

[0105] In addition, during model training, historical anomaly feature vectors from the training dataset are input into the model's input layer, and deep representations of case features are extracted through nonlinear transformations in the hidden layers. The output layer generates the predicted probability for each fault type using a sigmoid activation function.

[0106] During training, the model was evaluated using a validation dataset every 10 epochs. At epoch 50, the accuracy on the validation dataset reached 85%, recall 80%, and F1 score 0.82. Training continued, and by epoch 70, the accuracy on the validation dataset improved to 88%, recall 83%, and F1 score 0.85. However, at epoch 80, the accuracy dropped to 87%, recall to 82%, and F1 score to 0.84. Since the early stopping mechanism was set to halt training after 5 consecutive epochs of no further performance improvement, and there had been 2 consecutive epochs of no improvement, observation continued. At epoch 85, the performance metrics still showed no improvement, so training stopped, and the current optimal model parameters were saved.

[0107] This design allows for the accurate definition of different fault types and their corresponding abnormal characteristic patterns, and enables the construction of an abnormality diagnosis model.

[0108] Step 141: When the state category in the performance test result is identified as an abnormal state, extract the abnormal feature vector from the performance test result. The abnormal feature vector includes abnormal fluctuation features in the time-frequency domain features and abnormal distribution features in the statistical features.

[0109] In this embodiment of the invention, when the status category identifier in the performance test result shows an abnormal state, it is necessary to extract an abnormal feature vector from the performance test result. The abnormal feature vector includes abnormal fluctuation features in the time-frequency domain and abnormal distribution features in the statistical features. In the time-frequency domain, abnormal fluctuation features may manifest as abnormal fluctuations in the signal waveform, such as the abnormal occurrence of peak or valley points, or abnormal time intervals and amplitude change rates between adjacent extreme points. In the statistical features, abnormal distribution features may manifest as abnormalities in the central tendency, dispersion, or morphological characteristics of the signal sampling points. For example, the signal's center position may have shifted significantly, or the variance of the discrete distribution feature may be too large.

[0110] Step 142: Call the preset AI knowledge base interface to retrieve historical fault case data related to the touch screen IR sensing hole from the AI ​​knowledge base. The historical fault case data includes historical abnormal feature vectors, corresponding fault type labels, and fault cause analysis reports.

[0111] In this embodiment of the invention, after extracting the abnormal feature vector, a preset AI knowledge base interface is invoked to retrieve relevant historical fault case data from the AI ​​knowledge base. To accurately retrieve useful data, fault case retrieval conditions are constructed. These conditions include the device model information of the touchscreen IR sensor hole, the sensor channel identification information, and the abnormal feature dimension information from the performance test results. A case retrieval request containing the fault case retrieval conditions and data format requirements is sent to the AI ​​knowledge base through the preset AI knowledge base interface. Upon receiving the request, the AI ​​knowledge base performs data retrieval based on the retrieval conditions and returns the case retrieval results. The retrieval results are sorted by relevance, and a case sequence is generated by ranking cases from high to low according to their matching degree with the retrieval conditions. Historical fault case data is extracted from the case sequence and filtered to obtain valid case data containing complete historical abnormal feature vectors, fault type labels, and fault cause analysis reports. This historical fault case data can provide rich training samples for building an abnormal diagnostic model.

[0112] Step 1421: Construct fault case retrieval conditions, which include device model information of the touch screen IR sensing hole, sensing channel identification information, and abnormal feature dimension information in the performance test results.

[0113] In this embodiment of the invention, to accurately retrieve historical fault case data related to the current touchscreen IR sensor hole fault from the AI ​​knowledge base, it is necessary to construct fault case retrieval conditions. These conditions include the device model information of the touchscreen IR sensor hole, the sensing channel identification information, and the abnormal feature dimension information from the performance test results. The device model information ensures that the retrieved historical fault cases are consistent with the currently detected touchscreen IR sensor hole model, as different models of touchscreen IR sensors may have different performance characteristics and fault modes. The sensing channel identification information further refines the retrieval scope, as the fault conditions of different sensing channels may differ. The abnormal feature dimension information from the performance test results can filter out historical fault cases with similar abnormal features based on the currently detected abnormal features. By constructing the above fault case retrieval conditions, the accuracy and efficiency of the retrieval can be improved.

[0114] Step 1422: Send a case retrieval request to the AI ​​knowledge base through the preset AI knowledge base interface. The retrieval request includes the retrieval conditions and data format requirements for the fault cases.

[0115] In this embodiment of the invention, after constructing the fault case retrieval conditions, a case retrieval request is sent to the AI ​​knowledge base through a preset AI knowledge base interface. This request includes not only the fault case retrieval conditions but also data format requirements. Data format requirements ensure that the data returned by the AI ​​knowledge base meets the needs of subsequent processing. For example, it may require the returned data to be stored in a specific file format, or that the field order and format of the data conform to relevant specifications. By sending a case retrieval request containing both fault case retrieval conditions and data format requirements, the AI ​​knowledge base can accurately provide the required historical fault case data, avoiding processing difficulties caused by data format mismatches, thus improving the efficiency of data retrieval and processing.

[0116] Step 1423: Receive the case retrieval results returned by the AI ​​knowledge base, sort the retrieval results by relevance, and generate a case sequence by sorting the cases from high to low according to the matching degree between the cases and the retrieval conditions.

[0117] In this embodiment of the invention, after sending a case retrieval request to the AI ​​knowledge base, the system receives the case retrieval results returned by the AI ​​knowledge base. Since the retrieval results may contain multiple historical fault cases, and these cases may have varying degrees of matching with the retrieval criteria, it is necessary to sort the retrieval results by relevance. Cases are sorted from highest to lowest degree of matching with the retrieval criteria, generating a case sequence. For example, cases that highly match the device model, sensor channel identifier, and abnormal feature dimension information will be ranked higher in the sequence; while cases with lower matching degrees will be ranked lower. Relevance sorting makes subsequent data filtering and analysis more efficient. Prioritizing cases with high matching degrees allows for faster identification of historical fault cases most similar to the current fault situation.

[0118] Step 1424: Extract historical fault case data from the case sequence, filter the extracted historical fault case data to obtain valid case data containing complete historical anomaly feature vectors, fault type labels and fault cause analysis reports.

[0119] In this embodiment of the invention, after generating a case sequence, historical fault case data is extracted from the sequence. Since the case sequence may contain some incomplete or invalid cases, the extracted historical fault case data needs to be filtered. The goal of the filtering is to obtain valid case data that includes complete historical anomaly feature vectors, fault type labels, and fault cause analysis reports. For example, some cases may lack historical anomaly feature vectors or have unclear fault type labels; these cases need to be excluded. Through filtering, the quality and completeness of the historical fault case data used to build the anomaly diagnosis model can be ensured. Valid case data can provide accurate training samples for the anomaly diagnosis model, improving the model's fault diagnosis capability.

[0120] Step 143: Perform data preprocessing on the historical failure case data, generate a cleaning case dataset by removing noisy cases and duplicate cases, and divide the cleaning case dataset into a training dataset and a validation dataset.

[0121] In this embodiment of the invention, after obtaining valid case data, data preprocessing is required. First, outlier detection is performed on the historical abnormal feature vectors in the historical fault case data using the Isolation Forest algorithm to identify outliers, which are then marked as noise cases. Outliers may be caused by data collection errors or other anomalies, which can negatively impact model training and therefore need to be removed. Next, the similarity between cases is calculated based on the unique case identifier and the historical abnormal feature vectors. Duplicate cases are identified by setting a similarity threshold, and the case with the most recent collection timestamp among the duplicate cases is retained as a valid case. Redundant cases in the marked noise cases and duplicate cases are removed to generate a clean case dataset containing only one valid case. To avoid overfitting or underfitting of the model, the clean case dataset needs to be balanced by increasing the number of minority fault types through oversampling. Finally, the balanced clean case dataset is randomly divided into a training dataset and a validation dataset according to a preset partitioning ratio. The training dataset is used to train the anomaly diagnosis model, and the validation dataset is used to evaluate the model's performance.

[0122] Step 1431: Use the Isolation Forest algorithm to detect outliers in the historical abnormal feature vectors of historical fault case data, identify outlier samples, and mark the outlier samples as noise cases.

[0123] In this embodiment of the invention, to improve the quality of historical fault case data, outlier detection is required on the historical abnormal feature vectors. The Isolation Forest algorithm is used to analyze these historical abnormal feature vectors. The Isolation Forest algorithm can quickly identify outliers in the data by constructing decision trees. Outliers are those samples whose characteristics differ significantly from the majority of samples, possibly due to errors in the data acquisition process, sensor malfunctions, or other anomalies. Identified outliers are marked as noisy cases. Noisy cases can interfere with the training of the anomaly diagnosis model, affecting its accuracy and stability. Removing noisy cases makes the training data cleaner and improves the performance of the anomaly diagnosis model.

[0124] Step 1432: Calculate the similarity between cases based on the unique case identifier and historical anomaly feature vector. Identify duplicate cases by setting a similarity threshold and retain the case with the latest collection timestamp among the duplicate cases as a valid case.

[0125] In this embodiment of the invention, after removing noisy cases, it is necessary to identify and process duplicate cases. The similarity between cases is calculated based on the unique case identifier and historical anomaly feature vectors. The unique case identifier ensures the uniqueness of each case, while the historical anomaly feature vectors reflect the case's characteristic information. By calculating the similarity, the degree of similarity between different cases can be determined. A similarity threshold is set; when the similarity between two cases exceeds this threshold, they are considered duplicate cases. For duplicate cases, the case with the most recently collected timestamp is retained as a valid case. This is because the most recently collected case may contain more accurate and up-to-date information, better reflecting the current fault situation. By identifying and processing duplicate cases, data redundancy can be avoided, improving data utilization efficiency and model training effectiveness.

[0126] Step 1433: Remove the labeled noisy cases and redundant cases from the duplicate cases to generate a clean case dataset containing only one valid case.

[0127] In this embodiment of the invention, after outlier detection and duplicate case identification, noisy cases and redundant cases among the marked duplicate cases are removed. Noisy cases interfere with model training, while redundant cases increase data complexity and training time. After removing these invalid cases, a clean case dataset containing only one valid case is generated. The clean case dataset contains only accurate, complete, and non-repeating historical fault case data, providing a good data foundation for subsequent model training and validation. In the clean case dataset, each case is representative and can be used more effectively for the construction and evaluation of anomaly diagnosis models.

[0128] Step 1434: Perform data balancing on the cleaning case dataset and increase the number of cases of minority failure types by oversampling.

[0129] In this embodiment of the invention, after generating the clean case dataset, there may be an imbalance in the number of cases for different fault types. Some fault types may have a large number of cases, while others may have a small number. This data imbalance can cause the anomaly diagnosis model to have a strong ability to identify the majority of fault types during training, but a weak ability to identify the minority fault types, thus affecting the overall performance of the model. To resolve this problem, data balancing processing of the clean case dataset is required. Oversampling methods can be used to increase the number of cases for the minority fault types. For example, by replicating cases of the minority fault types or generating new similar cases, the number of cases for different fault types can be relatively balanced. This can improve the anomaly diagnosis model's ability to identify the minority fault types and enhance the model's generalization ability.

[0130] Step 1435: Randomly divide the balanced cleaning case dataset into a training dataset and a validation dataset according to a preset division ratio. The training dataset is used for model training, and the validation dataset is used for model performance evaluation.

[0131] In this embodiment of the invention, after data balancing, the balanced clean case dataset is randomly divided into a training dataset and a validation dataset according to a preset partitioning ratio. The training dataset is used to train the anomaly diagnosis model. The model learns the feature patterns corresponding to different fault types by studying historical fault cases in the training dataset. The validation dataset is used to evaluate the model's performance. During model training, the model is periodically tested using the validation dataset to check performance metrics such as accuracy and recall. By dividing the training and validation datasets, overfitting can be avoided, ensuring that the model has good generalization ability in practical applications. The preset partitioning ratio can be adjusted according to specific application scenarios and data conditions to achieve the best model training and evaluation results.

[0132] Step 144: Construct an anomaly diagnosis model based on the training dataset. Train a multi-label classification model by using historical anomaly feature vectors as input and fault type labels as output. Adjust model parameters and optimize model performance by using a validation dataset.

[0133] In this embodiment of the invention, after dividing the training dataset and validation dataset, an anomaly diagnosis model is constructed based on the training dataset. The anomaly diagnosis model adopts a multi-label classification model architecture, which includes an input layer, a hidden layer, and an output layer. The dimension of the input layer is consistent with the dimension of the historical anomaly feature vectors to ensure accurate reception of input data. The dimension of the output layer is consistent with the number of fault type labels, used to output prediction results for different fault types. The historical anomaly feature vectors from the training dataset are input into the model's input layer, and a deep representation of the case features is extracted through nonlinear transformations in the hidden layer. Neurons in the hidden layer perform complex calculations and transformations on the input features to uncover the inherent features and patterns of the data. The output layer generates the predicted probability for each fault type using the sigmoid activation function. The cross-entropy loss function is used to calculate the loss value between the model's predicted probability and the actual fault type label, and the model weight parameters are updated using the backpropagation algorithm. An early stopping mechanism is introduced during model training. By monitoring the model performance metrics on the validation dataset, training stops when the performance metrics no longer improve after a preset number of rounds, and the current optimal model parameters are saved. Finally, the trained model is evaluated using the validation dataset, and the anomaly diagnosis model is completed when the evaluation metrics meet preset conditions.

[0134] Step 1441: Construct a multi-label classification model architecture, which includes an input layer, a hidden layer, and an output layer. The dimension of the input layer is consistent with the dimension of the historical anomaly feature vector, and the dimension of the output layer is consistent with the number of fault type labels.

[0135] In this embodiment of the invention, to construct an anomaly diagnosis model, the model architecture must first be determined. A multi-label classification model architecture is adopted, which includes an input layer, a hidden layer, and an output layer. The dimension of the input layer is set to be consistent with the dimension of the historical anomaly feature vectors, ensuring that the input layer can accurately receive historical anomaly feature vectors from the training dataset. The dimension of the output layer is set to be consistent with the number of fault type labels, because the output layer needs to output prediction results for different fault types. The hidden layer is used to perform nonlinear transformations and feature extraction on the input features, mining deep features and patterns in the data. By reasonably setting the dimensions of the input layer, hidden layer, and output layer, the multi-label classification model can effectively process historical fault case data, improving the model's fault diagnosis capability.

[0136] Step 1442: Input the historical anomaly feature vectors from the training dataset into the input layer of the model, extract the deep representation of case features through the nonlinear transformation of the hidden layer, and generate the predicted probability of each fault type through the sigmoid activation function in the output layer.

[0137] In this embodiment of the invention, after constructing a multi-label classification model architecture, historical anomaly feature vectors from the training dataset are input into the model's input layer. The input layer then passes these feature vectors to the hidden layer. The hidden layer performs a non-linear transformation on the input feature vectors, processing the features using non-linear functions (such as the ReLU function) to extract a deep representation of the case features. This deep representation more comprehensively reflects the inherent characteristics and patterns of the data, improving the model's fault diagnosis accuracy. After receiving the feature representations from the hidden layer, the output layer generates a predicted probability for each fault type using the sigmoid activation function. The sigmoid activation function maps the output value to a probability value between 0 and 1, facilitating the determination of the likelihood of each fault type. In this way, the model can learn and predict historical fault cases from the training dataset.

[0138] Step 1443: Calculate the loss value between the model's predicted probability and the actual fault type label using the cross-entropy loss function, and update the model weight parameters using the backpropagation algorithm.

[0139] In this embodiment of the invention, during model training, it is necessary to evaluate the difference between the model's prediction results and the actual fault type labels. The cross-entropy loss function is used to calculate this difference, which measures the degree of inconsistency between the model's predicted probabilities and the actual labels. A smaller loss value indicates that the model's prediction results are close to the actual labels; conversely, a larger loss value indicates a larger error in the model's predictions. To reduce the loss value and improve the model's prediction accuracy, a backpropagation algorithm is used to update the model's weight parameters. The backpropagation algorithm calculates the gradient of each weight parameter based on the magnitude of the loss value and adjusts the values ​​of the weight parameters accordingly. By continuously iterating and updating the weight parameters, the model's prediction results gradually approach the actual labels, improving the model's performance.

[0140] Step 1444: Introduce an early stopping mechanism during model training. By monitoring the model performance metrics on the verification dataset, training is stopped when the performance metrics no longer improve after a preset number of rounds, and the current optimal model parameters are saved.

[0141] In this embodiment of the invention, to avoid overfitting, an early stopping mechanism is introduced during model training. This mechanism monitors model performance metrics, such as accuracy and recall, on the validation dataset. During training, the model is periodically tested on the validation dataset to evaluate its performance. When the performance metrics no longer improve after a preset number of rounds, it indicates that the model has reached a good state, and continued training may lead to overfitting. At this point, training is stopped, and the current optimal model parameters are saved. The model corresponding to the optimal parameters exhibits good performance on the validation dataset and can more accurately diagnose faults in practical applications. The early stopping mechanism improves the model's generalization ability and avoids overtraining.

[0142] Step 1445: Evaluate the performance of the trained model using the validation dataset. When the evaluation metrics meet the preset conditions, the construction of the anomaly diagnosis model is completed.

[0143] In this embodiment of the invention, after completing model training and saving the optimal model parameters, the trained model needs to be evaluated using a validation dataset. The validation dataset contains a portion of historical fault case data that was not used for training, which can be used to test the model's generalization ability. Evaluation metrics include accuracy, recall, and F1 score. When these evaluation metrics meet preset conditions, it indicates that the model has good performance on the validation dataset and can accurately perform fault type matching and root cause analysis. At this point, the construction of the anomaly diagnosis model is complete. If the evaluation metrics do not meet the preset conditions, the model parameters need to be adjusted or retrained until the model performance meets the requirements. Evaluation using the validation dataset ensures that the anomaly diagnosis model has high reliability and accuracy in practical applications.

[0144] Step 145: Input the abnormal feature vector into the trained abnormal diagnosis model, match the fault type by calculating the similarity between the abnormal feature vector and the historical abnormal feature vector, perform root cause tracing analysis in conjunction with the fault cause analysis report, and generate a fault tracing label containing fault type code and cause description.

[0145] In this embodiment of the invention, after constructing the anomaly diagnosis model, the anomaly feature vectors extracted from the performance detection results are input into the trained anomaly diagnosis model. First, the anomaly feature vectors are standardized to have the same scale as the historical anomaly feature vectors stored in the model. Then, the cosine similarity between the transformed anomaly feature vectors and the historical anomaly feature vectors is calculated to generate a similarity score list. Candidate historical cases with similarity scores higher than a preset score are selected from the similarity score list, and the corresponding fault type labels and fault cause analysis reports are extracted. The fault type labels of the candidate historical cases are voted on, and the fault type label with the highest frequency is selected as the matching fault type. Based on the fault cause analysis report corresponding to the matching fault type, key cause description terms are extracted from the report, and a structured cause description text is generated by combining the anomaly feature dimensions in the anomaly feature vectors. Finally, the fault type code of the matching fault type is combined with the cause description text to generate a fault tracing label. The fault tracing label includes a type code field and a cause description field. The type code field is used to store the unique identifier code of the fault type, and the cause description field is used to store the fault cause description in natural language form.

[0146] Step 1451: Standardize the abnormal feature vector, calculate the cosine similarity between the transformed abnormal feature vector and the historical abnormal feature vector stored in the abnormal diagnosis model, and generate a similarity score list.

[0147] In this embodiment of the invention, after the abnormal feature vector is input into the anomaly diagnosis model, it is first subjected to a standardization transformation. The standardization transformation ensures that the abnormal feature vector and the historical abnormal feature vectors stored in the anomaly diagnosis model have the same scale, facilitating similarity calculation. Then, the cosine similarity between the transformed abnormal feature vector and the historical abnormal feature vectors is calculated. Cosine similarity measures the cosine of the angle between two vectors; a value closer to 1 indicates greater similarity. By calculating the cosine similarity between all historical abnormal feature vectors and the current abnormal feature vector, a similarity score list is generated, recording the degree of similarity between the current abnormal feature vector and each historical abnormal feature vector.

[0148] Step 1452: Select candidate historical cases with similarity scores higher than the preset score from the similarity score list, and extract the fault type label and fault cause analysis report corresponding to the candidate historical cases.

[0149] In this embodiment of the invention, after generating a similarity score list, candidate historical cases with similarity scores higher than a preset score are selected from the list. The preset score is a pre-defined threshold used to select historical cases with a high similarity to the current abnormal feature vector. After selecting candidate historical cases, the corresponding fault type labels and fault cause analysis reports are extracted. The fault type labels clearly identify the fault type corresponding to the historical case, while the fault cause analysis report records the cause of the fault in detail. By filtering and extracting this information, the focus can be placed on historical cases most similar to the current fault situation, providing a more targeted reference for fault type matching and root cause tracing analysis.

[0150] Step 1453: Perform a voting statistics on the fault type tags of the candidate historical cases, and select the fault type tag with the highest frequency as the matching fault type.

[0151] In this embodiment of the invention, after extracting fault type labels from candidate historical cases, a voting statistics process is performed on these labels. The frequency of each fault type label appearing in the candidate historical cases is counted. The fault type label with the highest frequency is selected as the matching fault type. This voting statistics method comprehensively considers information from multiple candidate historical cases, improving the accuracy of fault type matching. For example, if a fault type label appears frequently in multiple candidate historical cases, it indicates that the fault type has a high correlation with the current abnormal feature vector, making it a more reliable matching fault type. Through voting statistics, the most likely fault type can be determined from numerous candidate historical cases.

[0152] Step 1454: Based on the fault cause analysis report corresponding to the matched fault type, extract the key cause description terms from the fault cause analysis report, and generate a structured cause description text by combining the abnormal feature dimension in the abnormal feature vector.

[0153] In this embodiment of the invention, after determining the matching fault type, key causal descriptive terms are extracted from the fault cause analysis report corresponding to that fault type. Key causal descriptive terms refer to keywords or phrases that accurately describe the cause of the fault. Combined with the anomaly feature dimension in the anomaly feature vector, a structured causal descriptive text is generated. For example, the anomaly feature dimension might show abnormal fluctuations in the signal of a certain sensing channel, while the fault cause analysis report mentions that the sensor in that sensing channel may be faulty. Combining this information generates a causal descriptive text in natural language, such as "Abnormal fluctuations in the signal of a certain sensing channel may be caused by a fault in the sensor of that channel." The structured causal descriptive text can more accurately and completely explain the cause of the fault.

[0154] Step 1455: Combine the fault type code matching the fault type with the cause description text to generate a fault tracing label. The fault tracing label includes a type code field and a cause description field. The type code field is used to store the unique identifier code of the fault type, and the cause description field is used to store the fault cause description in natural language form.

[0155] In this embodiment of the invention, after generating structured cause description text, the fault type code matching the fault type is combined with the cause description text to generate a fault tracing tag. The fault tracing tag includes a type code field and a cause description field. The type code field stores a unique identifier for the fault type, facilitating fault type management and identification. The cause description field stores a description of the fault cause in natural language, allowing the fault cause to be presented in an easily understandable way. By generating fault tracing tags, fault type and cause information can be integrated. For example, production managers can quickly identify the type and cause of a fault based on the fault tracing tag and take corresponding measures for repair and improvement.

[0156] In a scalable embodiment, the method further includes: Step 210: Extract the key process parameter dimensions contained in the cause description field from the fault tracing label to obtain a list of key process parameter dimensions.

[0157] In this embodiment of the invention, after generating a fault tracing label, key process parameter dimensions are extracted from its cause description field. The cause description field may contain process parameter information related to the fault's cause, such as temperature and pressure parameters in a certain production stage. By analyzing the cause description field, key process parameter dimensions are extracted, and a list of key process parameter dimensions is constructed. This list contains process parameter dimensions closely related to the current fault's cause, providing a foundation for subsequent process parameter analysis and adjustment. For example, if the cause description field mentions that "excessive temperature in a certain production stage may cause a fault," then "the temperature of that production stage" is extracted as a key process parameter dimension and included in the list of key process parameter dimensions.

[0158] Step 220: Based on the list of key process parameter dimensions, call the production line process parameter database interface to retrieve historical process parameter datasets related to the key process parameter dimensions. The historical process parameter datasets contain different combinations of process parameter values ​​and the corresponding IR sensor hole performance test results.

[0159] In this embodiment of the invention, after obtaining the list of key process parameter dimensions, the production line process parameter database interface is called based on this list. The production line process parameter database stores a large amount of process parameter data related to the production process of the touchscreen IR sensor hole. Through the database interface, historical process parameter datasets related to the key process parameter dimensions are retrieved. The historical process parameter datasets contain different combinations of process parameter values ​​and corresponding IR sensor hole performance test results. For example, for a certain key process parameter dimension, there may be multiple different values, and each combination of values ​​corresponds to an IR sensor hole performance test result. By retrieving the historical process parameter datasets, the impact of different process parameter values ​​on the IR sensor hole performance can be clearly identified.

[0160] Step 230: Perform parameter impact correlation analysis on the historical process parameter dataset to generate a parameter impact correlation matrix with key process parameter dimensions as rows and performance test result status categories as columns. The elements in the parameter impact correlation matrix represent the impact weight of the corresponding process parameter dimension on the performance status. The impact weight is determined through the correlation analysis between parameter value changes and performance status changes in historical data.

[0161] In this embodiment of the invention, after retrieving the historical process parameter dataset, a parameter influence correlation analysis is performed. A parameter influence correlation matrix is ​​generated, with key process parameter dimensions as rows and performance test result status categories as columns. Each element in the matrix represents the influence weight of the corresponding process parameter dimension on the performance status. The influence weight is determined through correlation analysis of parameter value changes and performance status changes in historical data. For example, if a change in the value of a certain process parameter dimension is closely related to a change in the performance status of the IR sensor aperture, then the influence weight of that process parameter dimension on the performance status will be large; conversely, the influence weight will be small. Through parameter influence correlation analysis and the generation of the parameter influence correlation matrix, the degree of influence of each key process parameter dimension on the performance status of the IR sensor aperture can be accurately and completely determined.

[0162] Step 240: Based on the parameter influence correlation matrix and the cause description in the fault tracing label, identify the process parameter dimensions that need to be adjusted and the adjustment direction, and generate a process parameter adjustment plan that includes parameter dimension identifier, current value, suggested adjustment range and expected performance improvement target. The process parameter adjustment plan is used to guide the process parameter update of the touch screen production line.

[0163] In this embodiment of the invention, after obtaining the parameter influence correlation matrix, the process parameter dimensions that need adjustment and the adjustment direction are identified by combining the cause descriptions in the fault tracing tags. Based on the influence weight of each process parameter dimension on the performance status in the parameter influence correlation matrix, and the fault causes indicated in the fault tracing tags, it is determined which process parameter dimensions need adjustment and how to adjust them. For example, if the parameter influence correlation matrix shows that a certain process parameter dimension has a significant negative impact on the performance status, and the fault tracing tags also indicate that this process parameter may be the cause of the fault, then this process parameter dimension needs to be adjusted. After determining the adjustment direction, a process parameter adjustment scheme is generated, including the parameter dimension identifier, the current value, the suggested adjustment range, and the expected performance improvement target. The parameter dimension identifier clarifies the process parameter dimension that needs adjustment; the current value records the existing value of the process parameter; the suggested adjustment range gives the reasonable adjustment range for the process parameter; and the expected performance improvement target describes the expected performance improvement effect of the IR sensing hole after adjusting the process parameter. The process parameter adjustment scheme can provide specific guidance for updating the process parameters of the touch screen production line to improve production quality and efficiency.

[0164] In a scalable embodiment, the method further includes: Step 310: Extract the device component identifier corresponding to the fault type code from the fault tracing label to obtain a list of associated device component identifiers.

[0165] In this embodiment of the invention, after obtaining the fault tracing tag, the device component identifier corresponding to the fault type code is extracted from it. The fault type code is associated with a specific device component; by extracting the corresponding device component identifier, the device component related to the current fault can be clearly identified. These device component identifiers are combined to obtain a list of associated device component identifiers. For example, if the fault type code indicates a fault in a certain sensing channel, the device component identifier corresponding to that sensing channel, such as a sensor or wiring, can be extracted by querying the relevant mapping relationship. The list of associated device component identifiers provides a basis for subsequent equipment status assessment and maintenance, used to identify device components that require focused attention and inspection.

[0166] Step 320: Based on the list of associated equipment component identifiers, collect the current operating status data and historical maintenance record data of each associated equipment component in the touch screen production line. The current operating status data includes component runtime, cumulative number of failures, and real-time performance monitoring indicators.

[0167] In this embodiment of the invention, after obtaining the list of associated equipment component identifiers, the current operating status data and historical maintenance record data of each associated equipment component in the touch screen production line are collected based on this list. The current operating status data includes information such as component runtime, cumulative number of failures, and real-time performance monitoring indicators. Component runtime reflects the usage time of the equipment component, the cumulative number of failures reflects the reliability of the equipment component, and real-time performance monitoring indicators clearly show the operating status of the equipment component in real time. Historical maintenance record data includes past maintenance information of the equipment component, such as maintenance time and maintenance content. By collecting this data, the health status of associated equipment components can be comprehensively assessed. For example, for a certain associated equipment component, if it is known that its runtime is long and the cumulative number of failures is high, and the real-time performance monitoring indicators show abnormalities, combined with historical maintenance record data, the health status of the equipment component can be more accurately determined.

[0168] Step 330: Conduct a comprehensive evaluation of the current operating status data and historical maintenance record data of the associated equipment components to generate health assessment indicators for each equipment component. The health assessment indicators include component failure risk coefficient, performance degradation rate and maintenance urgency score. The health assessment indicators are generated through multi-dimensional feature analysis of operating status data and maintenance record data.

[0169] In this embodiment of the invention, after collecting the current operating status data and historical maintenance record data of the associated equipment components, a comprehensive evaluation is performed on these data. Through multi-dimensional feature analysis of the operating status data and maintenance record data, health assessment indicators for each equipment component are generated. These health assessment indicators include a component failure risk coefficient, performance degradation rate, and maintenance urgency score. The component failure risk coefficient assesses the likelihood of failure based on factors such as the cumulative number of failures and operating time of the equipment component; the performance degradation rate determines the rate of performance decline by analyzing the changing trends of real-time performance monitoring indicators; and the maintenance urgency score comprehensively considers both the component failure risk coefficient and the performance degradation rate to assess the urgency of maintenance required for the equipment component. For example, for a equipment component with a long operating time, a large number of cumulative failures, and a significant decline in real-time performance monitoring indicators, its component failure risk coefficient will be higher, its performance degradation rate will be greater, and its maintenance urgency score will also be higher. By generating health assessment indicators, the health status of associated equipment components can be quantitatively evaluated.

[0170] Step 340: Prioritize the maintenance of related equipment components based on health assessment indicators, and generate an equipment maintenance plan that includes equipment component identification, maintenance priority level, suggested maintenance time period and maintenance work content by combining the production line production plan scheduling data. The equipment maintenance plan is used to guide the allocation of equipment maintenance resources for the touch screen production line.

[0171] In this embodiment of the invention, after generating health assessment indicators for each equipment component, the maintenance priorities of related equipment components are ranked according to these indicators. Factors such as component failure risk coefficient, performance degradation rate, and maintenance urgency score in the health assessment indicators are comprehensively considered to rank the related equipment components and determine the maintenance priority of each equipment component. Combining the production line scheduling data and considering the production tasks and time arrangements of the production line, an equipment maintenance plan is generated, including equipment component identification, maintenance priority level, suggested maintenance time period, and maintenance operation content. The equipment component identification clearly identifies the equipment component that needs maintenance; the maintenance priority level reflects the urgency of the equipment component's maintenance; the suggested maintenance time period is determined based on the production line scheduling and the health status of the equipment component; and the maintenance operation content details the specific operations to be performed on the equipment component. By generating an equipment maintenance plan, equipment maintenance resources for the touch screen production line can be rationally allocated, ensuring the normal operation of the equipment and improving production efficiency and quality.

[0172] This invention enables efficient and accurate detection and fault diagnosis of the performance of the IR sensing aperture of a touchscreen, greatly improving the quality control level in the touchscreen production process. Specifically, it receives detection signal data transmitted from performance detection sensors arranged around the IR sensing aperture of the touchscreen, covering infrared light intensity response signals and channel identification information for different sensing channels. Feature extraction is performed on the detection signal data, generating time-frequency domain features and statistical features from both time and statistical dimensions, forming a multi-dimensional feature set. This multi-dimensional feature extraction method can more comprehensively capture the intrinsic characteristics of the signal and, compared to single-dimensional analysis, more accurately reflect the performance status of the IR sensing aperture of the touchscreen.

[0173] Furthermore, the multi-dimensional feature set is input into a pre-defined machine learning algorithm model for classification and recognition. By performing pattern matching on the correlation between features, performance status classification results are generated. Utilizing the powerful pattern recognition capabilities of machine learning, the performance status of the touchscreen IR sensor hole can be determined efficiently and accurately, avoiding the subjectivity and inaccuracy of manual judgment. When the performance detection results indicate that the touchscreen IR sensor hole is in an abnormal state, historical fault case data stored in the pre-defined AI knowledge base is called to construct an anomaly diagnosis model. Fault type matching and root cause tracing analysis are performed on the abnormal features to obtain fault source labels. This process fully utilizes the value of historical data, enabling rapid and accurate location of fault types and causes, providing support for timely repair measures, effectively reducing downtime and costs in the production process, improving production efficiency and product quality, and realizing comprehensive monitoring and management of the touchscreen IR sensor hole performance.

[0174] Furthermore, Figure 2A structural block diagram of a performance detection data processing system 300 based on a touchscreen IR sensor hole is shown, including: a memory 310 for storing program instructions and data; and a processor 320 for coupling with the memory 310 to execute the instructions in the memory 310 to implement the above-described method.

[0175] Furthermore, a computer storage medium is also provided, comprising instructions that, when executed on a processor, implement the above-described method.

[0176] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for processing performance testing data based on a touchscreen IR sensor aperture, characterized in that, The method includes: The system receives detection signal data transmitted by a performance detection sensor located around the IR sensing hole of the touch screen within a preset sampling period. The detection signal data includes infrared light intensity response signals collected by the performance detection sensor under different sensing channels and corresponding channel identification information. The detected signal data is subjected to feature extraction processing. Time-frequency domain features are generated by analyzing the changing patterns of the signal waveform in the time dimension, and statistical features are generated by calculating the signal distribution characteristics in the statistical dimension, resulting in a multi-dimensional feature set containing time-frequency domain features and statistical features. The multi-dimensional feature set is input into a preset machine learning algorithm model for classification and recognition processing. The performance status classification result is generated by pattern matching of the correlation between features. The performance status classification result is used as the performance detection result of the touch screen IR sensing hole. If the performance test result indicates that the touch screen IR sensing hole is in an abnormal state, the historical fault case data stored in the preset AI knowledge base is called to construct an abnormal diagnosis model. The abnormal diagnosis model is used to perform fault type matching and root cause tracing analysis on the abnormal features contained in the performance test result to obtain a fault source tag used to indicate the cause of the fault. The process of inputting the multi-dimensional feature set into a preset machine learning algorithm model for classification and recognition, generating a performance status classification result by pattern matching of the correlation between features, and using the performance status classification result as the performance detection result of the touch screen IR sensing hole includes: Feature correlation analysis is performed on the time-frequency domain feature vectors and statistical feature vectors in the multi-dimensional feature set. A feature correlation matrix is ​​generated by calculating the correlation coefficients between different feature dimensions. Based on the feature correlation matrix, a subset of key features with correlation is selected. The key feature subset is input into the feature mapping layer of the machine learning algorithm model, and the key feature subset is mapped from the original feature space to a high-dimensional feature space through nonlinear transformation to generate a high-dimensional feature representation vector. The high-dimensional feature representation vector is processed by feature interaction layer of machine learning algorithm model. Enhanced feature vector is generated by constructing product interaction terms and cross combination terms between features. The enhanced feature vector contains the original features and interaction relationship information between features. The enhanced feature vector is input into the classification decision layer of the machine learning algorithm model. The enhanced feature vector is classified and judged for performance status through a multi-classifier ensemble voting mechanism, generating a probability distribution vector containing different performance status categories. The category with the highest probability value in the probability distribution vector is taken as the performance status classification result. Based on the performance status classification results, a performance test result containing a status category identifier and a confidence index is generated, wherein the confidence index is used to indicate the reliability of the classification result.

2. The method according to claim 1, characterized in that, The process of feature extraction of the detected signal data involves generating time-frequency domain features by analyzing the changing patterns of the signal waveform in the time dimension, and generating statistical features by calculating the signal distribution characteristics in the statistical dimension, resulting in a multi-dimensional feature set containing both time-frequency domain features and statistical features, including: The detected signal data is processed by signal segmentation. Based on the channel identification information, the infrared light intensity response signals of different sensing channels are divided into independent signal sequence units. Each signal sequence unit contains the light intensity response value of continuous sampling points and the corresponding timestamp information. Waveform morphology analysis is performed on each signal sequence unit in the time dimension. Waveform structure features are generated by identifying the distribution patterns of peak points, valley points, and zero-crossing points of the signal waveform. Time interval features are generated by calculating the time interval between adjacent extreme points. Rate of change features are generated by calculating the amplitude change rate between adjacent extreme points. The waveform structure features, time interval features, and rate of change features are combined as the time-frequency domain analysis results. The distribution characteristics of each signal sequence unit are calculated in the statistical dimension. The central location feature is generated by analyzing the central tendency of the signal sampling points, the discrete distribution feature is generated by analyzing the discreteness of the signal sampling points, and the distribution morphology feature is generated by analyzing the morphological features of the signal sampling points. The central location feature, the discrete distribution feature and the distribution morphology feature are combined into statistical features. The time-frequency domain analysis results and the statistical features are aligned in terms of feature dimensions. By unifying the feature description dimensions, different types of features have the same feature vector length, resulting in a multi-dimensional feature set containing time-frequency domain feature vectors and statistical feature vectors.

3. The method according to claim 1, characterized in that, The step involves performing feature correlation analysis on the time-frequency domain feature vectors and statistical feature vectors in the multi-dimensional feature set, generating a feature correlation matrix by calculating the correlation coefficients between different feature dimensions, and selecting a subset of key features with correlation based on the feature correlation matrix, including: Extract the waveform structure feature dimension and dynamic change feature dimension contained in the time-frequency domain feature vector, extract the center position feature dimension, discrete distribution feature dimension and distribution shape feature dimension contained in the statistical feature vector, and construct a feature dimension list containing all feature dimensions; Data standardization is performed on each feature dimension in the feature dimension list by converting the feature values ​​into standard scores with a mean of zero and a variance of one, in order to eliminate the dimensional differences between different feature dimensions. Calculate the correlation coefficient between any two feature dimensions after standardization, and generate a feature correlation matrix with the feature dimensions as rows and columns. The element values ​​in the feature correlation matrix represent the degree of linear correlation between the corresponding two feature dimensions. The feature association matrix is ​​evaluated for feature importance. A feature importance score is generated by calculating the mutual information value between each feature dimension and the performance status classification result. Feature dimensions with mutual information values ​​greater than a preset threshold are marked as important feature dimensions. Based on the feature correlation matrix and important feature dimensions, a feature selection model is constructed. A recursive feature elimination algorithm is used to select a subset of key features from the important feature dimensions that are correlated and have a redundancy lower than a set redundancy value. The key feature subset contains fewer feature dimensions than the original feature dimensions.

4. The method according to claim 1, characterized in that, The high-dimensional feature representation vector is processed by a feature interaction layer of a machine learning algorithm model, which performs feature combination processing. An enhanced feature vector is generated by constructing product interaction terms and cross-combination terms between features. This enhanced feature vector contains the original features and information about the interaction relationships between features, including: The high-dimensional feature representation vector is divided into multiple feature groups, each containing feature dimensions with similar physical meanings. The core feature dimension of each group is determined by calculating the variance contribution of each feature dimension within the feature group. Within each feature group, a product interaction term is constructed by the core feature dimension and other feature dimensions within the group. The product interaction term represents the synergistic relationship between the core feature and other features. Between different feature groups, cross-combination terms are constructed by extracting the core feature dimensions of each group, and the cross-combination terms represent the correlation between different feature groups; The product interaction terms and cross combination terms are subjected to feature scaling processing, and the feature values ​​of the interaction terms are mapped to a preset numerical range by the maximum and minimum value normalization method. The original high-dimensional feature representation vector, product interaction term, and cross combination term are standardized to eliminate dimensional differences. The standardized original high-dimensional feature representation vector, product interaction term, and cross combination term are then concatenated in order of feature dimension to form an enhanced feature vector. The number of dimensions of the enhanced feature vector is the sum of the original feature dimension, the product interaction term dimension, and the cross combination term dimension.

5. The method according to claim 1, characterized in that, If the performance test result indicates that the touchscreen IR sensor hole is in an abnormal state, an anomaly diagnosis model is constructed by calling historical fault case data stored in a preset AI knowledge base. The anomaly diagnosis model is then used to perform fault type matching and root cause tracing analysis on the abnormal features contained in the performance test result to obtain fault source tracing labels indicating the cause of the fault, including: When the state category in the performance test result is identified as an abnormal state, an abnormal feature vector is extracted from the performance test result. The abnormal feature vector includes abnormal fluctuation features in the time-frequency domain features and abnormal distribution features in the statistical features. Call the preset AI knowledge base interface to retrieve historical fault case data related to the IR sensing hole of the touch screen from the AI ​​knowledge base. The historical fault case data includes historical abnormal feature vectors, corresponding fault type labels and fault cause analysis reports. The historical failure case data is preprocessed to generate a cleaning case dataset by removing noisy and duplicate cases. The cleaning case dataset is then divided into a training dataset and a validation dataset. An anomaly diagnosis model is constructed based on the training dataset. A multi-label classification model is trained by using historical anomaly feature vectors as input and fault type labels as output. The model parameters are adjusted and the model performance is optimized by using a validation dataset. The abnormal feature vector is input into the trained abnormal diagnosis model. Fault type matching is performed by calculating the similarity between the abnormal feature vector and the historical abnormal feature vector. Root cause analysis is performed in conjunction with the fault cause analysis report to generate fault source tracing labels that include fault type codes and cause descriptions.

6. The method according to claim 5, characterized in that, The process involves calling a preset AI knowledge base interface to retrieve historical fault case data related to the touchscreen IR sensor hole from the AI ​​knowledge base. This historical fault case data includes historical anomaly feature vectors, corresponding fault type labels, and fault cause analysis reports, including: Construct fault case retrieval conditions, which include device model information of the touch screen IR sensing hole, sensing channel identification information, and abnormal feature dimension information in the performance test results; A case retrieval request is sent to the AI ​​knowledge base through a preset AI knowledge base interface. The retrieval request includes the retrieval conditions and data format requirements for the fault cases. Receive case retrieval results returned by the AI ​​knowledge base, sort the retrieval results by relevance, and generate a case sequence by sorting the cases from high to low according to the matching degree between the cases and the retrieval conditions; Historical fault case data is extracted from the case sequence, and the extracted historical fault case data is filtered to obtain valid case data containing complete historical anomaly feature vectors, fault type labels, and fault cause analysis reports.

7. The method according to claim 5, characterized in that, The historical failure case data is preprocessed to generate a cleaning case dataset by removing noisy and duplicate cases. This cleaning case dataset is then divided into a training dataset and a validation dataset, including: The isolated forest algorithm is used to detect outliers in the historical abnormal feature vectors of historical failure case data, and outlier samples are identified and marked as noise cases. The similarity between cases is calculated based on the unique identifier of the case and the historical anomaly feature vector. Duplicate cases are identified by setting a similarity threshold, and the case with the latest collection timestamp among the duplicate cases is retained as a valid case. Remove noisy and redundant cases from the labeled cases to generate a clean case dataset containing only one valid case. The cleaning case dataset is subjected to data balancing processing, and the number of cases of minority failure types is increased by oversampling. The balanced cleaning case dataset is randomly divided into a training dataset and a validation dataset according to a preset division ratio. The training dataset is used for model training, and the validation dataset is used for model performance evaluation.

8. The method according to claim 5, characterized in that, The construction of the anomaly diagnosis model based on the training dataset involves training a multi-label classification model by using historical anomaly feature vectors as input and fault type labels as output, and adjusting model parameters to optimize model performance using a validation dataset. A multi-label classification model architecture is constructed, which includes an input layer, a hidden layer, and an output layer. The dimension of the input layer is consistent with the dimension of the historical anomaly feature vector, and the dimension of the output layer is consistent with the number of fault type labels. The historical anomaly feature vectors from the training dataset are input into the model input layer. The deep representation of case features is extracted through the nonlinear transformation of the hidden layer. The output layer generates the predicted probability of each fault type through the sigmoid activation function. The cross-entropy loss function is used to calculate the loss value between the model's predicted probability and the actual fault type label, and the model weight parameters are updated through the backpropagation algorithm. An early stopping mechanism is introduced during model training. By monitoring the model performance metrics on the validation dataset, training is stopped when the performance metrics no longer improve after a preset number of consecutive rounds, and the current optimal model parameters are saved. The trained model is evaluated using a validation dataset. When the evaluation metrics meet the preset conditions, the anomaly diagnosis model is completed. The process involves inputting the abnormal feature vector into a trained anomaly diagnosis model, matching the fault type by calculating the similarity between the abnormal feature vector and historical abnormal feature vectors, performing root cause tracing analysis in conjunction with the fault cause analysis report, and generating a fault tracing label containing fault type coding and cause description, including: The abnormal feature vectors are standardized and transformed, and the cosine similarity between the transformed abnormal feature vectors and the historical abnormal feature vectors stored in the abnormal diagnosis model is calculated to generate a similarity score list. Candidate historical cases with similarity scores higher than the preset score are selected from the similarity score list, and the fault type tags and fault cause analysis reports corresponding to the candidate historical cases are extracted. The failure type labels of candidate historical cases are voted on and the failure type label with the highest frequency is selected as the matching failure type. Based on the fault cause analysis report corresponding to the matched fault type, the key cause description terms in the fault cause analysis report are extracted, and a structured cause description text is generated by combining the abnormal feature dimension in the abnormal feature vector. The fault type code matching the fault type is combined with the cause description text to generate a fault tracing label. The fault tracing label includes a type code field and a cause description field. The type code field is used to store the unique identifier code of the fault type, and the cause description field is used to store the fault cause description in natural language form.

9. A performance testing data processing system based on a touchscreen IR sensing aperture, characterized in that, include: Memory is used to store program instructions and data; A processor, coupled to a memory, for executing instructions in the memory to implement the method as described in any one of claims 1-8.

Images