Random wavelet feature learning method and system for chip weak defect detection

Abstract

The application discloses a kind of random wavelet feature learning method and system for chip weak defect detection, it is related to signal processing and intelligent detection technical field, the method includes: obtaining chip vibration response signal, constructs and includes the multichannel input of original signal and its first-order, second-order difference;Utilize the wavelet kernel of random generation of multiple types and multiple scales to each channel signal is carried out convolution matching, obtains characteristic response;Extract the statistical characteristics such as positive value proportion, longest positive value continuous interval proportion and amplitude mean of response to realize dimension reduction;Through learnable cosine pooling structure, adaptive nonlinear mapping is carried out to statistical characteristics;Different dynamic hierarchical features are weighted and fused in combination with grouping multi-scale adaptive attention mechanism;Finally input lightweight classification network realizes defect state recognition.The application significantly reduces the model complexity and computing overhead, has high accuracy and strong robustness in strong noise environment, and is suitable for resource-limited online detection scene.

Description

Technical Field

[0001] This invention relates to the fields of signal processing and intelligent detection technology, and in particular to a random wavelet feature learning method and system for detecting weak defects in chips. Background Technology

[0002] With the rapid development of semiconductor manufacturing technology, the packaging structure of advanced chips is becoming increasingly complex, and the integration level is constantly improving. This leads to a greater diversity of defect types during the production process, and the defect size is continuously shrinking to the micro-nano level, placing extremely high demands on the sensitivity and accuracy of detection technologies. In actual production environments, chip vibration response signals acquired through methods such as ultrasonic excitation not only contain defect features reflecting the structural state, but are also inevitably affected by the superposition of intrinsic material noise, sensor noise, and environmental interference, resulting in signals that generally exhibit low signal-to-noise ratios and feature aliasing. The response signals of weak defects are often submerged in strong noise backgrounds, and the features of different types of defects highly overlap in the time and frequency domains, making effective separation and identification difficult. This makes high-precision, high-reliability online defect detection of chips a severe technical challenge.

[0003] Currently, the technical approaches for chip defect detection can be mainly divided into two categories: traditional analysis methods based on signal decomposition and data-driven fusion learning methods.

[0004] Signal decomposition-based methods typically follow a "denoise first, then analyze" paradigm. The core idea is to decompose the observed signal on pre-defined or adaptive basis functions (such as wavelet bases, IMF components of empirical mode decomposition, etc.), and enhance and extract implicit defect features by suppressing or eliminating noise-related components. Typical techniques include wavelet thresholding denoising, variational mode decomposition (VMD), and sparse representation. These methods have good physical interpretability, and their processing is closely related to the intrinsic mechanism of the signal. However, they usually require multi-step serial processing (such as decomposition, component selection, feature calculation, etc.), making the process cumbersome. Furthermore, the parameter settings and component selection for each step often heavily rely on the prior knowledge and experience of domain experts, resulting in limited generalization ability. When dealing with different chip types and packaging processes, it is difficult to achieve stable and efficient end-to-end automated analysis.

[0005] Fusion learning-based methods fall under the data-driven category and typically include two key stages: feature extraction and pattern recognition. Depending on the feature acquisition method, they can be further subdivided into methods based on handcrafted features and methods based on automated features. The former relies on knowledge from the signal processing domain, manually designing and extracting time-domain, frequency-domain, or time-frequency-domain statistical features of the signal (such as peak value, kurtosis, spectral energy, etc.), and then combining them with traditional machine learning models such as Support Vector Machines (SVM) and Random Forests for classification. This method has low automation; the completeness and discriminative power of the feature design directly determine the final performance ceiling, and it struggles to handle complex and ever-changing defect patterns. The latter, represented by deep learning, especially Convolutional Neural Networks (CNNs), automatically learns hierarchical feature representations from the original signal or its transformations through multi-layer nonlinear transformations, and completes end-to-end defect classification. Given sufficiently deep network structures (such as ResNet and DenseNet) and ample training data, deep learning models can learn complex mappings from noisy signals to defect categories, demonstrating powerful feature representation capabilities. However, this powerful capability comes at the cost of high computational costs: deep models typically have a huge number of parameters, requiring massive amounts of data and powerful GPU computing power for training, and also consuming a lot of memory and time during inference. Although techniques such as model compression and pruning can alleviate this problem to some extent, their development and deployment still have high requirements for computing resources, making them difficult to directly adapt to resource-constrained industrial online inspection scenarios.

[0006] In summary, existing technologies exhibit significant contradictions and limitations: traditional signal processing methods, while possessing clear physical meaning, suffer from low automation, complex processes, and poor efficiency; while deep learning-based automated methods, although capable of end-to-end learning, are characterized by complex models, high computational overhead, weak interpretability, and heavy dependence on training data and computing resources. Therefore, how to significantly reduce model complexity and computational resource consumption while maintaining good physical interpretability and adaptability, and achieve efficient and accurate online detection of subtle chip defects, has become a critical technical problem urgently needing to be solved in this field. Summary of the Invention

[0007] To address these issues, this invention provides a random wavelet feature learning method and system for detecting weak defects in chips, which solves the problems of complex parameters, high computational cost, insufficient ability to extract weak defect features, and poor robustness in low signal-to-noise ratio environments in existing deep detection models.

[0008] To address the aforementioned technical problems, embodiments of the present invention provide a random wavelet feature learning method for detecting weak defects in chips, the method comprising: S1: Obtain the vibration response signal of the chip under test under external excitation and construct a labeled defect signal dataset; S2: Perform multi-order differential processing on the original signal in the dataset to generate a multi-channel input signal containing the original signal and at least a first-order differential signal; S3: Normalize the multi-channel input signal; S4: Random wavelet kernel feature extraction: For the normalized signal of each channel, perform the following operations: S41: Randomly generate multiple wavelet kernels with different scale parameters from a wavelet basis function library containing various types; S42: Perform a convolution operation between the normalized signal and the randomly generated wavelet kernel to obtain the original feature response; S43: Perform statistical feature extraction on the original feature response to obtain a low-dimensional statistical feature vector; S44: The low-dimensional statistical feature vector is input into a learnable cosine pooling layer for nonlinear transformation to generate a high-level feature representation; the learnable cosine pooling layer adaptively adjusts the frequency and phase of the cosine function through trainable parameters. S5: The high-level feature representations obtained from different channels in step S4 are spliced ​​together to form a fused feature; S6: Input the fused features into a lightweight classification network for processing. The classification network includes at least a grouped multi-scale adaptive attention mechanism for differential channel attention weighting of feature groups from different source channels. S7: Based on the output of the classification network, obtain the classification result of chip defects.

[0009] Preferably, the multi-order differential processing in step S2 specifically includes generating 0th-order, 1st-order, and 2nd-order differential signals, wherein the 0th-order differential signal is the original signal.

[0010] Preferably, in step S41, the wavelet kernel type includes at least one of Morlet wavelet, Hermitian wavelet, Laplace wavelet and Harmonic wavelet; the parameters of the wavelet kernel include scaling factor a, translation factor b and attenuation factor ζ, wherein scaling factor a is uniformly and randomly sampled within a preset range, translation factor b is uniformly and randomly sampled within a preset range, and attenuation factor ζ is randomly selected from a preset set of discrete values.

[0011] Preferably, in step S42, the convolution operation is a random convolution with a dilation factor, and its operation is defined as follows: ; in, The wavelet kernel weight vector, For input signal, The wavelet kernel length, It is the expansion factor. For bias terms, For the summation index, The position index of the output sequence. For output position The convolution response value; the dilation factor The wavelet kernel length is adaptively determined based on the input signal length to achieve multi-scale feature extraction.

[0012] Preferably, in step S43, the statistical feature extraction operation is performed only on the portion of the original feature response that is greater than zero, and the extracted statistical features include at least two of the following: positive value proportion, the proportion of the longest continuous positive value interval, the global positive value mean, and the local positive value mean normalized based on the positive value proportion.

[0013] Preferably, in step S44, the transformation function of the learnable cosine pooling layer is: ; in, The input feature vector, and These are the learnable frequency scaling factor vector and the phase offset parameter vector, respectively. This indicates element-wise multiplication. This is the transformation function.

[0014] Preferably, in step S6, the specific implementation of the grouped multi-scale adaptive attention mechanism includes: The fusion features are divided into multiple feature subgroups according to the source channel; For each feature subgroup, the size of the one-dimensional convolution kernel is adaptively determined based on the number of its channels; The determined one-dimensional convolution kernel is used to convolve each feature subgroup, and then activated by the Sigmoid function to generate the attention weight vector of each subgroup. The attention weight vectors of all subgroups are concatenated and multiplied channel-by-channel with the original fused features to achieve feature reweighting.

[0015] Preferably, in step S6, the lightweight classification network further includes a classification head module, which includes a first fully connected layer, a ReLU6 activation function, and a second fully connected layer, wherein the output dimension of the second fully connected layer is equal to the number of defect categories.

[0016] Preferably, the method further includes a model training step: training the lightweight classification network using the defect signal dataset, employing the cross-entropy loss function and the Adam optimizer, and applying a learning rate decay strategy to optimize the network parameters.

[0017] This invention also provides a random wavelet feature learning system for detecting weak defects in chips. This system implements the aforementioned random wavelet feature learning method for detecting weak defects in chips, specifically including: The data acquisition and preprocessing module is used to acquire the vibration response signal of the chip under test under external excitation and construct a labeled defect signal dataset. The signal dynamic hierarchy construction module is used to perform multi-order differential processing on the original signal in the dataset to generate a multi-channel input signal containing the original signal and at least a first-order differential signal. A signal normalization module is used to normalize the multi-channel input signal; The random wavelet kernel feature extraction module is used for random wavelet kernel feature extraction: for the normalized signal of each channel, the following operations are performed: Multiple wavelet kernels with different scale parameters are randomly generated from a wavelet basis function library containing various types. The normalized signal is convolved with the randomly generated wavelet kernel to obtain the original feature response; Statistical feature extraction is performed on the original feature response to obtain a low-dimensional statistical feature vector; The low-dimensional statistical feature vector is input into a learnable cosine pooling layer for nonlinear transformation to generate a high-level feature representation; the learnable cosine pooling layer adaptively adjusts the frequency and phase of the cosine function through trainable parameters. The feature fusion module is used to splice together the high-level feature representations extracted from different channels by the random wavelet kernel feature extraction module to form a fused feature. A lightweight classification network module is used to input the fused features into a lightweight classification network for processing. The classification network includes at least a grouped multi-scale adaptive attention mechanism for differential channel attention weighting of feature groups from different source channels. The defect identification and output module is used to obtain the classification result of chip defects based on the output of the classification network.

[0018] As can be seen from the above technical solutions, this invention application has the following beneficial effects: (1) This invention employs a strategy that combines parameter-randomized wavelet kernel convolution with statistical feature extraction (such as PPV and LSPV) that focuses only on positive responses. Randomized wavelet kernels provide a rich variety of feature matching templates, while statistical feature compression can effectively suppress the interference of random noise and focus on signal components that are significantly related to wavelet patterns. At the same time, learnable cosine pooling and grouped multi-scale attention mechanisms further enhance the discriminative power and stability of feature representation. Experiments show that in a low signal-to-noise ratio and strong noise environment of -5dB, this method can still maintain a high detection accuracy of 98.07%, which is significantly better than traditional methods and is particularly suitable for the challenge of unsatisfactory signal quality in actual industrial scenarios.

[0019] (2) This invention replaces the multi-layer stacked computation in deep neural networks with single-layer random convolution, significantly reducing the number of trainable parameters and the computational cost of forward inference. Most of the parameters in the feature extraction stage are randomly generated and fixed, requiring optimization of only a small number of parameters in the subsequent lightweight network (such as cosine pooling). , The entire model has a streamlined structure, and while ensuring performance, the model inference time is only 6.72 seconds. This low computational load characteristic enables it to run efficiently on edge computing devices or embedded platforms, meeting the urgent needs of production lines for real-time online detection and resource-constrained deployment.

[0020] (3) This invention deeply integrates prior knowledge of signal processing with data-driven learning. The random wavelet kernel itself has a clear time-frequency physical meaning (such as the matching of transient oscillations by the Morlet wavelet), and its statistical features (such as PPV reflecting the matching ratio and LSPV reflecting the persistence characteristics) can directly relate to the physical mechanism of the signal and the defect. This makes the model decision-making process more transparent and helps engineers understand and trust the detection results. In addition, the generation of the random wavelet kernel does not depend on a specific dataset, and the learnable structure (such as cosine pooling and attention) is adaptive, making it relatively easy to transfer this method to other types of mechanical vibration fault diagnosis or weak signal detection tasks, and it has good generalization potential. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Referring to the drawings will make the features and advantages of the present invention clearer. The drawings are illustrative and should not be construed as limiting the present invention in any way. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of a random wavelet feature learning method for detecting weak defects in chips provided by the present invention; Figure 2 This is an overall architecture diagram of the method of the present invention; Figure 3 This is a schematic diagram of cosine pooling; Figure 4 This is a schematic diagram of the classification effect of the method of the present invention at SNR=-5dB, where (a) is the confusion matrix; (b) is the result of various indicators; Figure 5 This is a block diagram of a random wavelet feature learning system for detecting weak defects in chips, provided by the present invention.

[0022] Explanation of reference numerals in the accompanying drawings: 100, Data acquisition and preprocessing module; 200, Signal dynamic hierarchy construction module; 300, Signal normalization module; 400, Random wavelet kernel feature extraction module; 500, Feature fusion module; 600, Lightweight classification network module; 700, Defect identification and output module. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1: To address the problems of complex parameters, high computational cost, insufficient ability to extract weak defect features, and poor robustness in low signal-to-noise ratio environments in existing deep detection models, such as... Figure 1 As shown, this invention proposes a random wavelet feature learning method for detecting weak defects in chips, which includes: S1: Obtain the vibration response signal of the chip under test under external excitation and construct a labeled defect signal dataset; S2: Perform multi-order differential processing on the original signal in the dataset to generate a multi-channel input signal containing the original signal and at least a first-order differential signal; S3: Normalize the multi-channel input signal; S4: Random wavelet kernel feature extraction: For the normalized signal of each channel, perform the following operations: S41: Randomly generate multiple wavelet kernels with different scale parameters from a wavelet basis function library containing various types; S42: Perform a convolution operation between the normalized signal and the randomly generated wavelet kernel to obtain the original feature response; S43: Perform statistical feature extraction on the original feature response to obtain a low-dimensional statistical feature vector; S44: The low-dimensional statistical feature vector is input into a learnable cosine pooling layer for nonlinear transformation to generate a high-level feature representation; the learnable cosine pooling layer adaptively adjusts the frequency and phase of the cosine function through trainable parameters. S5: The high-level feature representations obtained from different channels in step S4 are spliced ​​together to form a fused feature; S6: Input the fused features into a lightweight classification network for processing. The classification network includes at least a grouped multi-scale adaptive attention mechanism for differential channel attention weighting of feature groups from different source channels. S7: Based on the output of the classification network, obtain the classification result of chip defects.

[0025] As can be seen from the above technical solution, this invention proposes a random wavelet feature learning method for detecting weak defects in chips. This method achieves efficient, robust, and interpretable defect identification through the following structured process: First, chip vibration signals are acquired and a labeled dataset is constructed to provide a foundation for supervised learning. Second, the original signals are subjected to multi-order differential processing to construct a multi-channel input containing dynamic information, enhancing the ability to capture signal change features. Next, the multi-channel signals are normalized to improve training stability and convergence speed. Then, multi-scale feature extraction is achieved at low cost by randomly generating multi-type, multi-scale wavelet kernels for convolution. Statistical features with clear physical meaning are then extracted from the convolution response to achieve robust compression of high-dimensional features. A learnable cosine pooling layer is introduced for adaptive nonlinear transformation to enhance feature representation. Afterward, high-level features from each channel are fused to form a comprehensive representation. Finally, a lightweight classification network integrating a grouped multi-scale adaptive attention mechanism is used to differentiate and classify the fused features, significantly reducing model complexity while ensuring accuracy, ultimately achieving accurate and efficient identification of chip defects.

[0026] Further, in step S1, data acquisition and dataset construction are performed. An external ultrasonic excitation device (such as an ultrasonic probe) is applied to the package surface of the flip chip under test, while its vibration response signal is collected using a high-precision vibration sensor (such as a laser Doppler vibrometer or piezoelectric accelerometer). For chip samples with different defect states (such as no defects, missing solder balls, microcracks, interface debonding, etc.), the above excitation and acquisition process is repeated to obtain a sufficient number of signal samples. Each sample is labeled with its corresponding defect category label to construct a chip defect signal dataset for training and verification, and it is divided into a training set and a verification set according to a preset ratio (such as 7:3 or 8:2).

[0027] Further, in step S2, a multi-dynamic-level signal is constructed. To improve the sensitivity of features to different signal change rates, multi-order difference processing is performed on the original vibration signals in the training and validation sets. Specifically, the first-order difference (reflecting the signal change rate) and the second-order difference (reflecting the signal change acceleration) are calculated, and the original signal (0th-order difference), the first-order difference signal, and the second-order difference signal are stacked in the channel dimension to form a three-dimensional multi-channel input signal representation with dimensions of [number of samples, signal length, 3]. This step explicitly introduces the dynamic information of the signal into the subsequent feature learning process.

[0028] Further, in step S3, signal normalization is performed. The mean of all samples for each channel (i.e., the original, first-order, and second-order signals) in the training set is calculated separately. ) and standard deviation ( Subsequently, Z-score normalization is performed on all samples in the training and validation sets using the mean and standard deviation of their corresponding channels. This operation aims to eliminate the influence of signal amplitude dimensions, accelerate the convergence process of model training, and make the model invariant to changes in the overall energy of the signal.

[0029] Further, in step S4, random wavelet kernel feature extraction is performed. This step is the core of the invention, aiming to efficiently extract multi-scale features with strong discriminative power through single-layer wavelet convolution with randomized parameters and diverse structures. The following sub-steps are performed independently for each differential channel of each sample (e.g., Figure 2 (As shown in the left part) S41: Random wavelet kernel generation. Random generation. Each wavelet convolution kernel is generated using the following random selection: 1. Wavelet type: Randomly select one from a predefined wavelet basis function library, which contains at least Morlet wavelet, Hermitian wavelet, Laplace wavelet and Harmonic wavelet.

[0030] 2. Kernel parameters: A set of parameters is randomly sampled from a preset parameter space. .

[0031] scaling factor : Controls the scaling of wavelets to match different frequency components, obeying The uniform distribution, i.e. .

[0032] Translation factor : Controls the position of the wavelet on the time axis, obeying The uniform distribution, i.e. .

[0033] Attenuation factor : Controlling the decay rate of wavelets from discrete sets Random selection from the list.

[0034] 3. Nuclear length From a preset length set (e.g.) Randomly select from the kernel to achieve short kernels to capture transient details and long kernels to capture macro trends.

[0035] Through the above randomization, we can obtain Wavelet kernels with different shapes, scales, and locations It provides a rich set of matching templates for feature extraction.

[0036] S42: Random convolution calculation. This involves processing the normalized input signal... With randomly generated wavelet kernel Convolution operations are performed. To further enhance the multi-scale sensing capability of single-layer convolution, an inflation factor is introduced. The convolution operation is defined as: ; in: For output position The convolution response value.

[0037] To find the summation index, iterate through each weight position within the wavelet kernel (from...). arrive ).

[0038] The position index of the output sequence. The range is determined by the input length. ,filling nuclear length and expansion factor A joint decision.

[0039] The weight vector is the random wavelet kernel.

[0040] The expansion factor is defined as follows: ,in This means that the expansion factor adapts to the kernel length; short kernels may gain larger expansion to perceive macroscopic trends, while long kernels use smaller expansion to focus on microstructures, thus enabling true multi-scale analysis within a single layer.

[0041] This is the bias term, the only parameter in this structure that can be adaptively determined based on prior data. It can be set in two ways: (a) based on the convolution results of the training samples. The calculation of the statistics, among which (a) Use training samples; (b) Sample randomly from a uniform distribution, such as The bias is used to eliminate DC offset of the signal and, as a threshold, retains only significant responses that are highly matched to the wavelet pattern (i.e., (part of the text).

[0042] After convolution, the system can be selectively activated by the ReLU function to further suppress negative responses.

[0043] S43: Statistical Feature Extraction. For the convolutional response generated by each wavelet kernel, extract the statistical features of its positive value portion (effective response) to achieve compression from high-dimensional temporal response to low-dimensional statistics and assign them explicit physical meaning. Specifically, extract the following four types of features: 1. Proportion of Positive Values ​​(PPV): The ratio of effective response points to total response points, reflecting the overall correlation between the wavelet kernel and the transient response. When the bias... At that time, only the valid responses that are highly matched with the wavelet mode are retained. 2. Longest Stretch of Positive Values ​​(LSPV): The ratio of the length of the longest continuous effective response segment to the total number of response points. It characterizes the longest continuous duration of the region where the test signal matches the wavelet kernel and is used to describe the duration of the effective response.

[0044] 3. Global Mean of Positive Values ​​(GMPV): The mean of the sum of all valid response values ​​divided by the total length of the signal, representing the overall average strength of the matched signal over the entire signal period.

[0045] 4. Local Mean of Positive Values ​​(LMPV): The ratio of the sum of all valid response values ​​to the number of valid response points (equivalent to the mean normalized based on PPV), representing the local average intensity of the pure matching region and providing another perspective on intensity information.

[0046] Each wavelet kernel generates a 4-dimensional statistical eigenvector. For N wavelet kernels, a 4N-dimensional statistical eigenvector is obtained. .

[0047] S44: Learnable cosine pooling. This involves transforming statistical eigenvectors... Input a learnable cosine pooling layer and perform adaptive nonlinear transformations (such as...) Figure 3 (as shown) ; in, The input feature vector, and These are the learnable frequency scaling factor vector and the phase offset parameter vector, respectively. This indicates element-wise multiplication. This is the transformation function. While preserving the excellent properties of the cosine function, such as periodicity and boundedness, this layer independently adjusts the shape of the nonlinear transformation for each statistical feature dimension through learnable parameters, greatly enhancing the adaptive and discriminative capabilities of feature representation.

[0048] Further, in step S5, feature fusion and standardization are performed. For a single sample, the feature vectors extracted from the three difference channels (original, first-order, and second-order) in step S4 and cosine-pooled are concatenated to form a more comprehensive fused feature matrix. Then, the mean and standard deviation of the fused features in the training set are calculated, and the fused features in the training and validation sets are standardized to eliminate the impact of differences in feature scales on the classifier.

[0049] Furthermore, in step S6, a lightweight classification network is constructed and processed (e.g., Figure 2 (As shown on the right). The standardized fusion features are then input into a carefully designed lightweight classification network for the final decision.

[0050] S61: Learnable cosine pooling layer. A learnable cosine pooling layer can be used again to perform nonlinear transformations on the fused features to further enhance the expressiveness and separability of the features.

[0051] S62: Group Multi-scale Self-Adaptive Attention (GMSAA) mechanism. This is another key module of the invention, used to differentially enhance features from different dynamic levels.

[0052] 1. Feature Grouping: The input feature vector is re-divided into three feature subgroups according to the difference order from which it originates. These correspond to the original signal, the first-order difference, and the second-order difference, respectively.

[0053] 2. Adaptive convolution kernel calculation: For each feature subgroup Based on its number of channels Adaptively calculate a one-dimensional convolution kernel size .For example, (The result is the nearest odd number). This allows the attention module to automatically adapt to the optimal local cross-channel interaction range on feature sets at different scales.

[0054] 3. Attention weight generation: For each subgroup Using a core size of The output channel number is compressed to 1 using a one-dimensional convolutional layer. Then, an attention weight vector with values ​​between [0,1] is generated by applying a sigmoid activation function. Its length is equal to .

[0055] 4. Feature Reweighting: The attention weight vectors of the three subgroups are reweighted. Concatenate along the channel dimension to obtain the complete attention weight vector. .Will By multiplying the original input features channel by channel, the important feature channels are enhanced while the minor channels are suppressed.

[0056] S63: Classification Header Module. Typically composed of one or more fully connected layers. In a preferred embodiment, the classification header includes: The first fully connected layer maps the feature dimensions to 128 dimensions.

[0057] Activation function: ReLU6 function is used.

[0058] The second fully connected layer: The output dimension is equal to the number of defect categories, and the Softmax function is used to convert the output into the predicted probability of each category.

[0059] Further, in step S7, model training and defect identification are performed. Training phase: Using the features and labels of the training set, the cross-entropy loss function is used as the optimization objective. The Adam optimizer (with an initial learning rate set to 0.001) is applied to all trainable parameters in the classification network (including learnable cosine pooling parameters). The learning rate is trained end-to-end, consisting of the convolutional layer weights of the attention module and the fully connected layer weights of the classification head. During training, an exponential decay strategy (e.g., a decay coefficient of 0.9) can be used to dynamically adjust the learning rate to improve training stability and final performance. The optimal-performing model is saved by monitoring metrics such as accuracy, recall, and F1 score on the validation set.

[0060] Inference / Detection Phase: For a new vibration signal from a chip under test, the same preprocessing as the training set is performed first (S2-S3). Then, feature extraction and transformation are performed using a fixed random wavelet kernel (S4). After feature fusion and standardization (S5), the signal is finally input into the trained lightweight classification network (S6). In the probability vector output by the network, the category corresponding to the highest probability value is the defect type predicted by the model.

[0061] To further illustrate the advantages of the present invention, specific experiments are described below.

[0062] Since the proposed random wavelet convolutional network does not require kernel training during the feature extraction stage, the relevant hyperparameters remained consistent across all experiments. For the original signal and its multi-order difference sequences, the number of randomly generated sequences was... The wavelet kernel, the Laplace wavelet kernel value is... Center frequency Time scale parameters The difference order is set to The convolutional kernel length ranges from 11 to 66. The backend classifier uses a two-layer fully connected structure, with 128 neurons in the hidden layer and the output layer dimension matching the number of classes. Training uses the Adam optimizer (learning rate 0.001, weight decay 1×10⁻⁶). -5 The method employs an exponential decay strategy (decay rate of 0.9), with a batch size of 32 and a total of 40 training epochs. All experiments were performed on an Intel Core i5-12400F six-core processor, and training and inference times were recorded. The proposed method runs on the CPU, while the comparative method was executed on an NVIDIA GTX 3090 GPU-accelerated environment.

[0063] Under noise interference conditions (SNR= The diagnostic performance comparison of various methods on the flip chip solder joint defect vibration dataset is shown in Table 1. The selected benchmark methods include: the WDCNN model with strong noise resistance, ResNet-18 based on deep residual learning, WKN and MWK using trainable preprocessing layers for hybrid feature extraction, MWK-ResNet combining MWK and ResNet, and CWT-CNN using fixed-parameter continuous wavelet transform as a predefined feature extraction layer. The results show that the proposed method significantly outperforms all the comparison models in the three core evaluation metrics of accuracy, recall, and F1-score, with a classification accuracy of 98.07% and a single model inference time of only 6.72 seconds. This indicates that, thanks to the broad coverage of random wavelet kernels across multiple scales and frequency bands, this method can effectively capture response patterns related to weak defects in high-noise backgrounds, exhibiting not only higher diagnostic accuracy but also excellent noise robustness when identifying multiple types of solder joint defects.

[0064] Table 1 Comparison of evaluation indicators for each model at SNR=-5dB

[0065] To further analyze the model's classification performance at specific noise levels, Figure 4Figure (a) shows the confusion matrix under this operating condition. It can be observed that the diagonal elements of the matrix remain at a high level overall, indicating that the model can effectively distinguish between the five types of chip operating conditions, including normal state, missing ball defect, and cold solder joint defect, and the confusion between the categories is relatively limited. Figure 4 Figure (b) visually illustrates the results of the proposed method on the three evaluation metrics of Accuracy, Recall, and F1-score using a bar chart. It can be seen that all three metrics remain at a high level, indicating that the model does not exhibit significant performance bias under high noise conditions, and the prediction results demonstrate good stability and consistency.

[0066] Based on this, in order to systematically evaluate the performance evolution of different methods under varying noise intensity conditions, this invention applies multiple discrete signal-to-noise ratio levels (SNR). 10dB 7dB The classification accuracy of each model was compared and analyzed at 5dB, 0dB, 5dB, and 7dB, and the quantitative results are summarized in Table 2.

[0067] Table 2. Classification accuracy of different methods under different signal-to-noise ratio conditions

[0068] Overall, as the signal-to-noise ratio (SNR) gradually decreases, the diagnostic performance of all comparison methods shows a significant decline. However, different models exhibit varying degrees of sensitivity to noise disturbances. Under high-noise conditions (e.g., SNR = ...), ... At 10dB, the classification accuracy of all methods decreased significantly, with WDCNN and ResNet-18 dropping to 26.22% and 38.52% respectively, and the remaining models generally falling below 45%, making it difficult to maintain stable discrimination ability. This indicates that under extremely low signal-to-noise ratio conditions, noise components dominate, and existing methods generally struggle to extract discriminative and effective features from high-noise backgrounds. From the perspective of model mechanism, pure convolutional neural networks are susceptible to noise accumulation interference during deep feature stacking; CWT-CNN, based on fixed-parameter continuous wavelet transform, lacks scale adaptation capability, limiting its feature representation in complex noise environments; while WKN, MWK-CNN, and MWK-ResNet, which introduce trainable wavelet structures, improve feature adaptation ability to some extent, their parameter optimization process still relies on effective signal support, and is prone to degradation under extremely low signal-to-noise ratio conditions, thus their overall performance is still significantly constrained by noise intensity.

[0069] Unlike the methods described above, the random wavelet convolutional network proposed in this invention exhibits significantly better diagnostic performance than the comparative models across the entire signal-to-noise ratio range. Even under the most stringent conditions... Even at 10dB, its classification accuracy still reaches 89.81%, far exceeding other methods. As the signal-to-noise ratio increases to... 7dB and The accuracy was further improved by 5 dB to 96.59% and 98.07%, and gradually stabilized in the medium-to-high signal-to-noise ratio range. This indicates that the method of the present invention has low sensitivity to changes in noise intensity and can maintain stable and reliable diagnostic capabilities in high-noise environments.

[0070] This advantage is primarily attributed to the large-scale, multi-scale, and multi-band redundant mapping mechanism introduced by random wavelet convolution in the feature extraction stage. By randomly covering wavelet kernels with different time scales and frequency structures, the model can capture local responses related to defects in a noisy environment without relying on precise parameter tuning. Simultaneously, since the convolution kernel parameters are not involved in training, the feature construction process effectively avoids noise interference with parameter updates, thus maintaining stable and discriminative feature representations even under extremely low signal-to-noise ratio conditions.

[0071] Example 2: Figure 5 As shown, this invention provides a random wavelet feature learning system for detecting weak defects in chips. This system is used to implement the random wavelet feature learning method for detecting weak defects in chips described in Embodiment 1 above, and specifically includes: The data acquisition and preprocessing module 100 is used to acquire the vibration response signal of the chip under test under external excitation and construct a labeled defect signal dataset. The signal dynamic hierarchy construction module 200 is used to perform multi-order differential processing on the original signal in the dataset to generate a multi-channel input signal containing the original signal and at least a first-order differential signal. The signal normalization module 300 is used to normalize the multi-channel input signal; The random wavelet kernel feature extraction module 400 is used for random wavelet kernel feature extraction: for the normalized signal of each channel, it performs the following operations: Multiple wavelet kernels with different scale parameters are randomly generated from a wavelet basis function library containing various types. The normalized signal is convolved with the randomly generated wavelet kernel to obtain the original feature response; Statistical feature extraction is performed on the original feature response to obtain a low-dimensional statistical feature vector; The low-dimensional statistical feature vector is input into a learnable cosine pooling layer for nonlinear transformation to generate a high-level feature representation; the learnable cosine pooling layer adaptively adjusts the frequency and phase of the cosine function through trainable parameters. The feature fusion module 500 is used to splice together the high-level feature representations extracted from different channels by the random wavelet kernel feature extraction module 400 to form a fused feature. The lightweight classification network module 600 is used to input the fused features into the lightweight classification network for processing. The classification network includes at least a grouped multi-scale adaptive attention mechanism for differential channel attention weighting of feature groups from different source channels. The defect identification and output module 700 is used to obtain the classification result of chip defects based on the output of the classification network.

[0072] This embodiment provides a random wavelet feature learning system for weak chip defect detection, used to implement the aforementioned random wavelet feature learning method for weak chip defect detection. Therefore, the specific implementation of the random wavelet feature learning system for weak chip defect detection can be found in the previous section on the embodiments of the random wavelet feature learning method for weak chip defect detection. For example, the data acquisition and preprocessing module 100, the signal dynamic hierarchy construction module 200, the signal normalization module 300, the random wavelet kernel feature extraction module 400, the feature fusion module 500, the lightweight classification network module 600, and the defect identification and output module 700 are respectively used to implement steps S1, S2, S3, S4, S5, S6, and S7 in the aforementioned random wavelet feature learning method for weak chip defect detection. Therefore, its specific implementation can be referred to the descriptions of the corresponding embodiments. To avoid redundancy, further details are omitted here.

[0073] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0076] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A random wavelet feature learning method for detecting weak defects in chips, characterized in that, include: S1: Obtain the vibration response signal of the chip under test under external excitation and construct a labeled defect signal dataset; S2: Perform multi-order differential processing on the original signal in the dataset to generate a multi-channel input signal containing the original signal and at least a first-order differential signal; S3: Normalize the multi-channel input signal; S4: Random wavelet kernel feature extraction: For the normalized signal of each channel, perform the following operations: S41: Randomly generate multiple wavelet kernels with different scale parameters from a wavelet basis function library containing various types; S42: Perform a convolution operation between the normalized signal and the randomly generated wavelet kernel to obtain the original feature response; S43: Perform statistical feature extraction on the original feature response to obtain a low-dimensional statistical feature vector; S44: Input the low-dimensional statistical feature vector into a learnable cosine pooling layer for nonlinear transformation to generate a high-level feature representation; The learnable cosine pooling layer adaptively adjusts the frequency and phase of the cosine function through trainable parameters; S5: The high-level feature representations obtained from different channels in step S4 are spliced ​​together to form a fused feature; S6: Input the fused features into a lightweight classification network for processing. The classification network includes at least a grouped multi-scale adaptive attention mechanism for differential channel attention weighting of feature groups from different source channels. S7: Based on the output of the classification network, obtain the classification result of chip defects.

2. The random wavelet feature learning method for detecting weak defects in chips according to claim 1, characterized in that, The multi-order differential processing in step S2 specifically includes generating 0th-order, 1st-order, and 2nd-order differential signals, wherein the 0th-order differential signal is the original signal.

3. The random wavelet feature learning method for detecting weak defects in chips according to claim 1, characterized in that, In step S41, the wavelet kernel type includes at least one of Morlet wavelet, Hermitian wavelet, Laplace wavelet, and Harmonic wavelet; the parameters of the wavelet kernel include a scaling factor. Translation factor and attenuation factor , where scaling factor Uniform random sampling within a preset range, shift factor Uniform random sampling within a preset range, attenuation factor Randomly select from a preset set of discrete values.

4. The random wavelet feature learning method for detecting weak defects in chips according to claim 1, characterized in that, In step S42, the convolution operation is a random convolution with a dilation factor, and its operation is defined as follows: ； in, The wavelet kernel weight vector, For input signal, The wavelet kernel length, It is the expansion factor. For bias terms, For the summation index, The position index of the output sequence. For output position The convolution response value; the dilation factor The wavelet kernel length is adaptively determined based on the input signal length to achieve multi-scale feature extraction.

5. The random wavelet feature learning method for detecting weak defects in chips according to claim 1, characterized in that, In step S43, the statistical feature extraction operation is performed only on the portion of the original feature response that is greater than zero. The extracted statistical features include at least two of the following: positive value proportion, the proportion of the longest continuous positive value interval, the global positive value mean, and the local positive value mean normalized based on the positive value proportion.

6. The random wavelet feature learning method for detecting weak defects in chips according to claim 1, characterized in that, In step S44, the transformation function of the learnable cosine pooling layer is: ； in, The input feature vector, and These are the learnable frequency scaling factor vector and the phase offset parameter vector, respectively. This indicates element-wise multiplication. This is the transformation function.

7. The random wavelet feature learning method for detecting weak defects in chips according to claim 1, characterized in that, In step S6, the specific implementation of the grouped multi-scale adaptive attention mechanism includes: The fusion features are divided into multiple feature subgroups according to the source channel; For each feature subgroup, the size of the one-dimensional convolution kernel is adaptively determined based on the number of its channels; The determined one-dimensional convolution kernel is used to convolve each feature subgroup, and then activated by the Sigmoid function to generate the attention weight vector of each subgroup. The attention weight vectors of all subgroups are concatenated and multiplied channel-by-channel with the original fused features to achieve feature reweighting.

8. The random wavelet feature learning method for detecting weak defects in chips according to claim 1, characterized in that, In step S6, the lightweight classification network further includes a classification head module, which includes a first fully connected layer, a ReLU6 activation function, and a second fully connected layer. The output dimension of the second fully connected layer is equal to the number of defect categories.

9. The random wavelet feature learning method for detecting weak defects in chips according to claim 1, characterized in that, It also includes a model training step: training the lightweight classification network using the defect signal dataset, employing the cross-entropy loss function and the Adam optimizer, and applying a learning rate decay strategy to optimize the network parameters.

10. A random wavelet feature learning system for detecting weak defects in chips, characterized in that, The system is used to implement the random wavelet feature learning method for chip weak defect detection as described in any one of claims 1 to 9, specifically including: The data acquisition and preprocessing module is used to acquire the vibration response signal of the chip under test under external excitation and construct a labeled defect signal dataset. The signal dynamic hierarchy construction module is used to perform multi-order differential processing on the original signal in the dataset to generate a multi-channel input signal containing the original signal and at least a first-order differential signal. A signal normalization module is used to normalize the multi-channel input signal; The random wavelet kernel feature extraction module is used for random wavelet kernel feature extraction: for the normalized signal of each channel, the following operations are performed: Multiple wavelet kernels with different scale parameters are randomly generated from a wavelet basis function library containing various types. The normalized signal is convolved with the randomly generated wavelet kernel to obtain the original feature response; Statistical feature extraction is performed on the original feature response to obtain a low-dimensional statistical feature vector; The low-dimensional statistical feature vector is input into a learnable cosine pooling layer for nonlinear transformation to generate a high-level feature representation; the learnable cosine pooling layer adaptively adjusts the frequency and phase of the cosine function through trainable parameters. The feature fusion module is used to splice together the high-level feature representations extracted from different channels by the random wavelet kernel feature extraction module to form a fused feature. A lightweight classification network module is used to input the fused features into a lightweight classification network for processing. The classification network includes at least a grouped multi-scale adaptive attention mechanism for differential channel attention weighting of feature groups from different source channels. The defect identification and output module is used to obtain the classification result of chip defects based on the output of the classification network.

Images