Strong magnetic chip detection method based on adaptive threshold method

Abstract

The present invention relates to a strong magnetic chip detection method based on an adaptive threshold method, including using historical detection data, based on chip samples under different conditions, similarity relationships between samples and magnetic characteristics of each sample, obtaining a snapshot of magnetic characteristic distribution, and calculating magnetic characteristic distribution; based on the snapshot of magnetic characteristic distribution, establishing a main task of predicting future magnetic characteristic distribution, and obtaining a distribution prediction result; based on the changed magnetic characteristic distribution, establishing an auxiliary task of future magnetic characteristic distribution change, and obtaining a prediction result of future magnetic characteristic distribution change; combining the prediction results of the main task and the auxiliary task with the magnetic characteristic distribution of the current conditions, inputting the neural network to obtain the final magnetic characteristic distribution and the corresponding preliminary detection results of the strong magnetic chip; calculating the multi-task composite loss function to obtain the strong magnetic chip detection result. The present invention can improve the accuracy and efficiency of detection, and improve chip quality control and production efficiency.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of electronic device detection, and in particular to a strong magnetic chip detection method based on an adaptive threshold method. Background Art

[0002] In the field of electronic manufacturing, the detection of strong magnetic chips is a key quality control link. Traditional detection methods usually rely on fixed thresholds and empirical rules to determine whether the magnetic characteristics of the chip meet the standards. These methods are often not flexible enough and difficult to adapt to the variable distribution of magnetic characteristics in the production process. In addition, due to the lack of in-depth analysis and utilization of historical data, these methods have limitations in predicting future changes in the distribution of magnetic characteristics. With the expansion of production scale and the improvement of chip performance requirements, the existing technology has gradually shown its shortcomings in terms of accuracy and efficiency.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: first, the fixed threshold method cannot adapt to the dynamic changes of the magnetic characteristics, resulting in inaccurate detection results under certain conditions; second, there is a lack of effective data-driven models to integrate historical detection data, which makes the ability to predict the distribution of magnetic characteristics limited; finally, the prior art fails to fully utilize the advantages of multi-task learning and fails to simultaneously consider the prediction and change prediction of the distribution of magnetic characteristics, which limits the performance improvement of the detection model. Summary of the invention

[0004] The content of this disclosure is used to introduce concepts in a brief form, which will be described in detail in the detailed implementation section below. The content of this disclosure is not intended to identify the key features or essential features of the technical solution claimed for protection, nor is it intended to limit the scope of the technical solution claimed for protection.

[0005] The purpose of the embodiment of the present invention is to provide a strong magnetic chip detection method based on an adaptive threshold method, which effectively improves the accuracy and efficiency of detection through a series of innovative training methods and technical strategies, and improves chip quality control and production efficiency.

[0006] The present invention provides a strong magnetic chip detection method based on an adaptive threshold method, comprising:

[0007] Using historical test data, based on chip samples under different conditions, similarity relationships between samples and magnetic characteristics of each sample, snapshots of magnetic characteristic distribution under different conditions are obtained, and the changing magnetic characteristic distribution is calculated;

[0008] Based on the snapshot of magnetic feature distribution, the main task of predicting the future magnetic feature distribution is established to obtain the prediction results of the future magnetic feature distribution;

[0009] Based on the changing magnetic feature distribution, an auxiliary task is established to predict the future changes in magnetic feature distribution under different conditions, and the prediction results of the future changes in magnetic feature distribution are obtained;

[0010] The prediction results of the main task and the auxiliary task are combined with the magnetic feature distribution of the current conditions, and input into the neural network for multi-task information aggregation to obtain the final magnetic feature distribution and the corresponding preliminary detection results of the strong magnetic chip;

[0011] The multi-task composite loss function is calculated, and the composite loss function is used to assist in learning and training the network to obtain the trained strong magnetic chip detection model and strong magnetic chip detection results.

[0012] Furthermore, the prediction step of the main task specifically includes: performing sample feature encoding and feature embedding using feature extraction technology based on the magnetic feature distribution snapshot to obtain initial feature embedding;

[0013] Based on the initial feature embedding, an adaptive threshold module of a neural network module is used to extract overall information of the magnetic feature to obtain a low-dimensional feature embedding;

[0014] Based on the low-dimensional feature embedding, an adaptive threshold module of a neural network module is used to extract information on changes in magnetic features generated under different conditions to obtain an optimized feature embedding;

[0015] Embedding the optimized features into a feedforward network to obtain embedding results of each sample, and aggregating the embedding results of each sample using a linear layer of the feedforward network to obtain an overall feature embedding result;

[0016] Based on the overall feature embedding result, a prediction module is used to predict the future magnetic feature distribution and the future magnetic state of each sample.

[0017] Furthermore, the prediction step of the auxiliary task specifically includes: performing sample feature encoding and feature embedding using feature extraction technology based on the changed magnetic feature distribution to obtain auxiliary initial feature embedding;

[0018] Based on the auxiliary initial feature embedding, the adaptive threshold module of the neural network module is used to extract the ripple effect of the local change feature on the overall magnetic feature distribution, so as to obtain the auxiliary optimization feature embedding;

[0019] Embedding the auxiliary optimized features into the input feedforward network to obtain auxiliary embedding results of each sample, and aggregating the auxiliary embedding results using a linear layer of the feedforward network to obtain auxiliary overall feature embedding results;

[0020] Based on the auxiliary overall feature embedding results, a prediction module is used to predict future magnetic feature distribution changes under different conditions.

[0021] Furthermore, the adaptive threshold module extracts information based on the following formula:

[0022]

[0023]

[0024] in, For the sample , its query vector in the adaptive threshold module; represents the key vector; represents a value vector; represents the dimension of the key vector; represents the product of the attention weight and value vector of the adaptive threshold module; Representation sample The output feature vector of Represents the sub-layer operation of the adaptive threshold module.

[0025] Furthermore, the adaptive threshold module extracts information based on the following formula:

[0026]

[0027]

[0028] in, For the sample The query vector in the criss-cross attention module; Representation sample The key vector of ; Representation sample A vector of values ​​of ; Represents the product of the attention weight and value vector of the cross-attention module; Representation sample The output feature vector of Represents the sub-layer operation of the criss-cross attention module.

[0029] Furthermore, the multi-task information aggregation is to combine the prediction results of the future magnetic characteristic distribution and the prediction results of the future magnetic characteristic distribution change with the magnetic characteristic distribution of the current conditions, input them into a multi-layer perceptron for information aggregation, and obtain a preliminary prediction result of the final magnetic characteristic distribution; combine the preliminary prediction result with the prediction result of the future sample magnetic state, input them into another multi-layer perceptron for information aggregation, and obtain a preliminary detection result of the strong magnetic chip corresponding to the final magnetic characteristic distribution. The expression of the preliminary detection result of the strong magnetic chip is:

[0030]

[0031] in, Indicates the preliminary test results of the strong magnetic chip corresponding to the final magnetic feature distribution; It indicates the predicted results of the future distribution of magnetic characteristics; It indicates the predicted results of the future changes in the distribution of magnetic characteristics; and is the parameter matrix; ReLU represents the activation function.

[0032] Furthermore, the composite loss function is expressed as:

[0033]

[0034] in, represents the composite loss function; , , is an adjustable training parameter; represents the loss function for the final magnetic feature distribution prediction; represents the loss function for predicting future changes in magnetic feature distribution; Represents the loss function for strong magnetic chip detection.

[0035] Furthermore, the specific steps of predicting the future magnetic feature distribution and the future sample magnetic state include: performing temporal position encoding on the overall feature embedding result to obtain an initial embedding feature;

[0036] Extracting time series information based on the initial embedding features using an adaptive threshold module and a feedforward neural network of a time domain neural network module;

[0037] Decoding the future magnetic characteristic distribution and the prediction result of the future sample magnetic state by a decoder;

[0038] The calculation formula of the temporal position coding is:

[0039]

[0040]

[0041] in, is the position index; is the dimension index; is the embedding vector dimension.

[0042] Furthermore, the calculation formula for extracting the timing information is:

[0043]

[0044]

[0045] in, Indicates The adaptive threshold calculation results of the layer; and is the parameter matrix; represents the activation function; Indicates The final temporal embedding features of the layer.

[0046] Furthermore, the expression of the magnetic characteristic distribution snapshot is:

[0047] S=[ s 1 , s 2 ,..., s n ]

[0048] The corresponding network change structure is:

[0049] [C=[ c ij ]

[0050] in, Represents a snapshot of the magnetic feature distribution; Indicates the corresponding changed network structure; Representation sample Magnetic characteristics of Representation sample and samples The similarity relationship between them.

[0051] The above-mentioned embodiments of the present disclosure have the following beneficial effects: the present invention can dynamically adjust the threshold to adapt to the magnetic feature distribution under different conditions through the adaptive threshold method and multi-task learning of the neural network, thereby improving the accuracy of strong magnetic chip detection. This method can effectively utilize historical detection data, and by establishing a snapshot of the magnetic feature distribution and predicting future distribution, the detection process is made more accurate and reliable. In addition, through the auxiliary learning of multi-task information aggregation and composite loss function, the generalization ability and detection performance of the model can be further improved, making the detection results of strong magnetic chips more stable and accurate.

[0052] In addition, the present invention can predict the changes in the distribution of magnetic features under different conditions in the future, providing forward-looking guidance for the production process. This predictive ability can help manufacturers adjust production strategies in advance, optimize resource allocation, and reduce cost waste caused by inaccurate detection. At the same time, through the application of adaptive threshold modules and cross-attention mechanisms, the similarities and differences between samples can be more deeply explored, and the detection model's ability to recognize complex magnetic features can be further improved, which is of great significance for improving product quality and production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which:

[0054] Figure 1 A schematic flow chart of a strong magnetic chip detection method based on an adaptive threshold method provided in one embodiment of the present invention.

[0055] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0056] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0057] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure can be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0058] According to an embodiment of the present invention, a strong magnetic chip detection method based on an adaptive threshold method is proposed.

[0059] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0060] The principle and spirit of the present invention are explained in detail below with reference to several representative embodiments of the present invention.

[0061] Reference below Figure 1 , Figure 1This is a flow chart of a strong magnetic chip detection method based on an adaptive threshold method provided by an embodiment of the present invention. It should be noted that the embodiments of the present invention can be applied to any applicable scenario.

[0062] Figure 1 The strong magnetic chip detection method based on the adaptive threshold method provided by one embodiment of the present invention shown in the figure, the method 100 comprises:

[0063] Step 101, using historical detection data, based on chip samples under different conditions, similarity relationships between samples and magnetic characteristics of each sample, obtaining snapshots of magnetic characteristic distribution under different conditions, and calculating the changed magnetic characteristic distribution;

[0064] Step 102, based on the snapshot of magnetic characteristic distribution, a main task of predicting future magnetic characteristic distribution is established to obtain a prediction result of future magnetic characteristic distribution;

[0065] Step 103, based on the changed magnetic characteristic distribution, establishing an auxiliary task for predicting the future magnetic characteristic distribution change under different conditions, and obtaining a prediction result of the future magnetic characteristic distribution change;

[0066] Step 104, combining the prediction results of the main task and the auxiliary task with the magnetic characteristic distribution of the current conditions, inputting into the neural network for multi-task information aggregation, and obtaining the final magnetic characteristic distribution and the corresponding preliminary detection results of the strong magnetic chip;

[0067] Step 105, calculate the multi-task composite loss function, use the composite loss function to assist in learning the training network, and obtain the trained strong magnetic chip detection model and strong magnetic chip detection results.

[0068] It should be noted that the strong magnetic chip detection method based on the adaptive threshold method predicts and detects the magnetic characteristics of the strong magnetic chip by using historical detection data and neural network technology. The core of this method lies in the use of adaptive thresholds, which can be dynamically adjusted according to changes in detection data to improve the accuracy and efficiency of detection.

[0069] Specifically, the method includes the following key steps: First, using historical detection data, based on chip samples under different conditions, similarity relationships between samples and magnetic characteristics of each sample, snapshots of magnetic feature distribution under different conditions are obtained. This involves an in-depth analysis of historical data, including sample collection, feature extraction and evaluation of similarity relationships. Then, based on the snapshot of magnetic feature distribution, the main task of predicting future magnetic feature distribution is established and the prediction results are obtained. This requires the construction of a neural network model that can process and learn magnetic feature distribution. Finally, the prediction results of the main task and auxiliary task are combined with the magnetic feature distribution of the current conditions, and input into the neural network for multi-task information aggregation to obtain the final magnetic feature distribution and the corresponding preliminary detection results of the strong magnetic chip.

[0070] Preferably, in order to improve the accuracy and efficiency of the detection method, further refinement can be performed. For example, when using historical detection data, data preprocessing techniques such as data cleaning and feature normalization can be used to improve data quality. When building a prediction model, different types of neural network structures, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs), can be selected to adapt to different data characteristics. In the multi-task information aggregation stage, an attention mechanism can be introduced to enhance the model's ability to recognize key features. In addition, the training process of the model can be optimized and the accuracy of the detection results can be improved by adjusting the parameters in the composite loss function.

[0071] In some embodiments, the prediction step of the main task specifically includes: performing sample feature encoding and feature embedding using feature extraction technology based on the magnetic feature distribution snapshot to obtain initial feature embedding;

[0072] Based on the initial feature embedding, an adaptive threshold module of a neural network module is used to extract overall information of the magnetic feature to obtain a low-dimensional feature embedding;

[0073] Based on the low-dimensional feature embedding, an adaptive threshold module of a neural network module is used to extract information on changes in magnetic features generated under different conditions to obtain an optimized feature embedding;

[0074] Embedding the optimized features into a feedforward network to obtain embedding results of each sample, and aggregating the embedding results of each sample using a linear layer of the feedforward network to obtain an overall feature embedding result;

[0075] Based on the overall feature embedding result, a prediction module is used to predict the future magnetic feature distribution and the future magnetic state of each sample.

[0076] It should be noted that the main task prediction step involves encoding and embedding sample features and extracting and predicting features using a neural network module. The key to the implementation of this part is how to process and optimize features through an adaptive threshold module and a feedforward network to predict future magnetic feature distribution.

[0077] Specifically, the prediction step of the main task includes the following key links: First, based on the snapshot of the magnetic feature distribution, the feature extraction technology is used to encode and embed the samples to obtain the initial feature embedding. This usually involves selecting a suitable feature representation method, such as principal component analysis PCA or autoencoder, and setting corresponding parameters, such as encoding dimension and learning rate. Next, the adaptive threshold module in the neural network module is used to extract the overall information of the magnetic features and obtain a low-dimensional feature embedding. This step requires defining the calculation formula and parameters of the adaptive threshold module, such as the dimensions of the query vector, key vector, and value vector. Then, based on the low-dimensional feature embedding, the adaptive threshold module is used again to extract information about the changes in magnetic features generated under different conditions to obtain the optimized feature embedding. Finally, the optimized feature embedding is input into the feedforward network, and the embedding results of each sample are aggregated through a linear layer to obtain the overall feature embedding result, and the prediction module is used to predict the future magnetic feature distribution and the future magnetic state of each sample.

[0078] Preferably, further refinement can be performed to improve the accuracy and efficiency of the prediction. For example, in the feature extraction stage, a deep learning method can be used to automatically learn feature representation instead of manually selecting features. In the design of the adaptive threshold module, more adaptive mechanisms can be introduced, such as dynamically adjusting the threshold parameters to better adapt to data changes. In the construction of the feedforward network, different types of activation functions, such as ReLU or Leaky ReLU, can be considered to improve the nonlinear expression ability of the network. In addition, regularization techniques such as dropout or weight decay can be used to prevent overfitting and improve the generalization ability of the model. In the design of the prediction module, time series analysis methods, such as long short-term memory networks (LSTMs), can be introduced to better capture time dependencies, thereby improving the prediction accuracy of future magnetic feature distributions.

[0079] In some embodiments, the prediction step of the auxiliary task specifically includes: performing sample feature encoding and feature embedding using feature extraction technology based on the changed magnetic feature distribution to obtain auxiliary initial feature embedding;

[0080] Based on the auxiliary initial feature embedding, the adaptive threshold module of the neural network module is used to extract the ripple effect of the local change feature on the overall magnetic feature distribution, so as to obtain the auxiliary optimization feature embedding;

[0081] Embedding the auxiliary optimized features into the input feedforward network to obtain auxiliary embedding results of each sample, and aggregating the auxiliary embedding results using a linear layer of the feedforward network to obtain auxiliary overall feature embedding results;

[0082] Based on the auxiliary overall feature embedding results, a prediction module is used to predict future magnetic feature distribution changes under different conditions.

[0083] Specifically, the prediction step of the auxiliary task includes the following key links: First, based on the changing magnetic feature distribution, feature extraction technology is used to encode and embed features of the samples to obtain auxiliary initial feature embedding. This may involve selecting an appropriate feature representation method and setting related parameters, such as encoding dimension and learning rate. Next, the adaptive threshold module in the neural network module is used to extract the impact of local change features on the overall magnetic feature distribution to obtain auxiliary optimized feature embedding. This step requires defining the calculation formula and parameters of the adaptive threshold module, such as the dimensions of the query vector, key vector, and value vector. Then, the auxiliary optimized feature embedding is input into the feedforward network, and the auxiliary embedding results of each sample are aggregated through a linear layer to obtain the auxiliary overall feature embedding result. Finally, based on the auxiliary overall feature embedding result, the prediction module is used to predict future changes in the magnetic feature distribution under different conditions.

[0084] Preferably, in order to improve the accuracy and efficiency of the prediction, further refinement or alternatives can be provided. For example, in the feature extraction stage, a convolutional neural network CNN can be used to automatically learn spatial feature representation, which is particularly effective when processing image data. In the design of the adaptive threshold module, an attention mechanism can be introduced to dynamically focus on the features that have the greatest impact on the predicted changes. In the construction of the feedforward network, different types of activation functions, such as LeakyReLU or ELU, can be considered to improve the expressive power of the network. In addition, the training process of the network can be accelerated and the stability of the model can be improved by introducing batch normalization BatchNormalization. In the design of the prediction module, a recurrent neural network RNN ​​or its variants, such as a long short-term memory network LSTM, can be introduced to better process time series data, thereby improving the prediction accuracy of future changes in the distribution of magnetic features.

[0085] In some embodiments, the adaptive threshold module extracts information based on the following formula:

[0086]

[0087]

[0088] in, For the sample , its query vector in the adaptive threshold module; represents the key vector; represents a value vector; represents the dimension of the key vector; represents the product of the attention weight and value vector of the adaptive threshold module; Representation sample The output feature vector of Represents the sub-layer operation of the adaptive threshold module.

[0089] Specifically, the operation steps of the adaptive threshold module include the following key steps: First, for each sample , calculate its query vector in the adaptive threshold module , key vector Sum value vector These vectors are the parameters of the neural network and need to be learned during the training process. Then, through the formula Calculate the product of the attention weight of the adaptive threshold module and the value vector, where is the dimension of the key vector, The function is used to normalize the weights so that they sum to 1. Then, using the formula Update sample The output feature vector ,in Represents the sub-layer operation of the adaptive threshold module, which usually includes a linear transformation and a non-linear activation function.

[0090] Preferably, in order to improve the performance of the adaptive threshold module, the operation steps can be further refined. For example, when calculating the query vector, key vector and value vector, it can be considered to introduce a regularization term, such as L2 regularization, to prevent overfitting. In the function, a temperature parameter can be introduced to control the distribution of weights, making them more concentrated or dispersed. In , different types of activation functions, such as ReLU or GELU, can be considered to improve the nonlinear expression ability of the model. In addition, residual connections can be introduced to facilitate the training of deep networks. and ( ) to facilitate the flow of information and the propagation of gradients. In practical applications, the structure and parameters of the adaptive threshold module can also be adjusted according to the requirements of specific tasks to achieve the best detection effect.

[0091] In some embodiments, the adaptive threshold module extracts information based on the following formula:

[0092]

[0093]

[0094] in, For the sample The query vector in the criss-cross attention module; Representation sample The key vector of ; Representation sample A vector of values ​​of ; Represents the product of the attention weight and value vector of the cross-attention module; Representation sample The output feature vector of Represents the sub-layer operation of the criss-cross attention module.

[0095] Specifically, the operation steps of the adaptive threshold module in the cross-attention mechanism include the following key links: First, for each sample , and calculates its query vector in the criss-cross attention module , and other samples The key vector Sum value vector These vectors are the parameters of the neural network and need to be learned during the training process. Then, through the formula Calculate the product of the attention weights of the criss-cross attention module and the value vector, where dk is the dimension of the key vector, The function is used to normalize the weights so that they sum to 1. Then, using the formula Update sample The output feature vector ,in Represents the sub-layer operation of the criss-cross attention module, which usually includes a linear transformation and a non-linear activation function.

[0096] Preferably, in order to improve the performance of the cross-attention mechanism, the operation steps can be further refined. For example, when calculating the query vector, key vector, and value vector, it is possible to consider introducing a regularization term, such as L2 regularization, to prevent overfitting. In the function, a temperature parameter can be introduced to control the distribution of weights, making them more concentrated or dispersed. In , different types of activation functions, such as ReLU or GELU, can be considered to improve the nonlinear expression ability of the model. In addition, residual connections can be introduced to facilitate the training of deep networks. and Adding direct connections between them helps the flow of information and the propagation of gradients. In practical applications, the structure and parameters of the cross-attention module can also be adjusted according to the requirements of specific tasks to achieve the best detection effect.

[0097] In some embodiments, the multi-task information aggregation is to combine the prediction results of the future magnetic characteristic distribution and the prediction results of the future magnetic characteristic distribution change with the magnetic characteristic distribution of the current conditions, input them into a multi-layer perceptron for information aggregation, and obtain a preliminary prediction result of the final magnetic characteristic distribution; combine the preliminary prediction result with the prediction result of the future sample magnetic state, input them into another multi-layer perceptron for information aggregation, and obtain the preliminary detection result of the strong magnetic chip corresponding to the final magnetic characteristic distribution. The expression of the preliminary detection result of the strong magnetic chip is:

[0098]

[0099] in, Indicates the preliminary test results of the strong magnetic chip corresponding to the final magnetic feature distribution; It indicates the predicted results of the future distribution of magnetic characteristics; It indicates the predicted results of the future changes in the distribution of magnetic characteristics; and is the parameter matrix; ReLU represents the activation function.

[0100] It should be noted that this implementation describes in detail how to combine the prediction results with the magnetic feature distribution of the current conditions, and perform multi-task information aggregation through a multi-layer perceptron to obtain the final magnetic feature distribution prediction results. This step is a key link in realizing the strong magnetic chip detection method, which involves effectively integrating information from different sources to improve the accuracy of detection.

[0101] Specifically, the implementation method includes the following key links: First, the predicted results of the future magnetic characteristic distribution and the predicted results of the future magnetic characteristic distribution changes are combined with the magnetic characteristic distribution of the current conditions. This involves data preprocessing and feature fusion, and the data may need to be normalized to ensure that data from different sources are numerically comparable. Next, these combined data are input into a multi-layer perceptron for information aggregation to obtain preliminary predictions of the final magnetic characteristic distribution. The multi-layer perceptron is a feedforward neural network consisting of multiple layers, each layer containing multiple neurons that can perform nonlinear transformations on the input data. Then, the preliminary prediction results are combined with the predicted results of the future sample magnetic state, and input into another multi-layer perceptron for information aggregation again to obtain preliminary detection results of the strong magnetic chip corresponding to the final magnetic characteristic distribution.

[0102] Preferably, in order to improve the efficiency and accuracy of information aggregation, the above-mentioned operation steps can be further refined or alternatives can be provided. For example, in the data preprocessing stage, different feature selection methods, such as statistical-based methods or model-based methods, can be considered to select the most informative features. In the design of the multilayer perceptron, the number of layers of the network and the number of neurons in each layer can be adjusted to find the best network structure. In addition, different activation functions, such as ReLU or Sigmoid, can be considered to improve the nonlinear expression ability of the network. In the information aggregation stage, regularization techniques, such as L1 or L2 regularization, can be considered to prevent overfitting and improve the generalization ability of the model. In addition, different optimization algorithms, such as stochastic gradient descent SGD or Adam, can be considered to accelerate the training process of the model. In practical applications, the structure and parameters of the multilayer perceptron can also be adjusted according to the needs of specific tasks to achieve the best detection effect.

[0103] In some embodiments, the composite loss function is expressed as:

[0104]

[0105] in, represents the composite loss function; , , is an adjustable training parameter; represents the loss function for the final magnetic feature distribution prediction; represents the loss function for predicting future changes in magnetic feature distribution; Represents the loss function for strong magnetic chip detection.

[0106] It should be noted that this embodiment describes in detail how to use a composite loss function to assist in learning and training the network, which is a key step in implementing the strong magnetic chip detection method. The composite loss function can consider the losses of multiple tasks at the same time, thereby optimizing the overall performance of the network.

[0107] Specifically, the implementation method includes the following key links: First, define a composite loss function, which consists of multiple parts, including a loss function for predicting the final magnetic feature distribution, a loss function for predicting future changes in the magnetic feature distribution, and a loss function for strong magnetic chip detection. These loss functions correspond to different prediction tasks of the model, and they work together in the training process of the network. Next, set adjustable training parameters, which are used to balance the importance of different loss functions to ensure that the model achieves good performance on all tasks. These parameters can be optimized by methods such as cross-validation. The network is then trained by minimizing the composite loss function, which involves using optimization algorithms such as gradient descent to update the weights of the network.

[0108] Preferably, in order to improve the training effect and generalization ability of the model, the above operation steps can be further refined or alternatives can be provided. For example, when defining the loss function, different loss functions, such as mean square error MSE, cross entropy loss, etc., can be considered to adapt to different prediction tasks. When setting adjustable training parameters, hyperparameter optimization techniques, such as grid search or Bayesian optimization, can be considered to automatically find the optimal parameter combination. When training the network, different optimization algorithms, such as Adam or RMSprop, can be considered, which can adaptively adjust the learning rate to accelerate the training process and improve the performance of the model. In addition, regularization techniques, such as dropout or weight decay, can also be considered to prevent overfitting and improve the generalization ability of the model. In practical applications, the composition and parameters of the composite loss function can also be adjusted according to the needs of specific tasks to achieve the best detection effect.

[0109] In some embodiments, the specific steps of predicting the future magnetic feature distribution and the future sample magnetic state include: performing temporal position encoding on the overall feature embedding result to obtain an initial embedding feature;

[0110] Extracting time series information based on the initial embedding features using an adaptive threshold module and a feedforward neural network of a time domain neural network module;

[0111] Decoding the future magnetic characteristic distribution and the prediction result of the future sample magnetic state by a decoder;

[0112] The calculation formula of the temporal position coding is:

[0113]

[0114]

[0115] in, is the position index; is the dimension index; is the embedding vector dimension.

[0116] It should be noted that this embodiment describes in detail how to predict the future magnetic feature distribution and the future sample magnetic state through time series position coding and time domain neural network module. This method utilizes the characteristics of time series data and enhances the model's ability to capture time dependence by encoding time series information.

[0117] Specifically, the implementation method includes the following key links: First, the overall feature embedding result is temporally encoded to obtain the initial embedded feature. Temporal position encoding is a method of incorporating position information into feature representation. It can be a simple position encoding, such as using different frequencies of sine and cosine functions, or a more complex encoding method, such as learnable position encoding. Then, based on the initial embedded features, the adaptive threshold module and feedforward neural network of the time domain neural network module are used to extract the timing information. The time domain neural network module is specially designed to process time series data, and the adaptive threshold module can be dynamically adjusted to adapt to different time characteristics. Then, the decoder is used to decode and obtain the prediction results of the future magnetic feature distribution and the future sample magnetic state. The decoder is usually one or more fully connected layers, which can convert the encoded features into the final prediction output.

[0118] Preferably, in order to improve the accuracy and efficiency of prediction, the above operation steps can be further refined or alternatives can be provided. For example, when performing temporal position encoding, more advanced encoding techniques can be considered, such as relative position encoding in Transformer, which can capture long-distance dependencies. When designing a time-domain neural network module, the use of a recurrent neural network RNN ​​or its variants, such as a long short-term memory network LSTM or a gated recurrent unit GRU, can be considered. These network structures are particularly suitable for processing time series data. In the design of the adaptive threshold module, more adaptive mechanisms can be considered, such as dynamically adjusting the threshold parameters to better adapt to data changes. In the design of the decoder, different types of activation functions, such as ReLU or Sigmoid, can be considered to improve the nonlinear expression ability of the model. In addition, the introduction of batch normalization Batch Normalization can also be considered to accelerate the training process of the network and improve the stability of the model. In practical applications, the structure and parameters of the temporal position encoding and time-domain neural network modules can also be adjusted according to the needs of specific tasks to achieve the best prediction effect.

[0119] In some embodiments, the calculation formula for extracting timing information is:

[0120]

[0121]

[0122] in, Indicates The adaptive threshold calculation results of the layer; and is the parameter matrix; represents the activation function; Indicates The final temporal embedding features of the layer.

[0123] It should be noted that this embodiment describes in detail how to extract time series information through a specific calculation formula, which is a key step in realizing the strong magnetic chip detection method. In this way, time series data can be effectively processed, thereby improving the accuracy of predicting future magnetic feature distribution and sample magnetic state.

[0124] Specifically, the implementation method includes the following key steps: First, define the calculation formula for extracting time series information, which involves updating the adaptive threshold calculation result. In the formula, Indicates The adaptive threshold calculation result of the layer, and is a parameter matrix, which is learned during the training process and used to adjust the weights of the features. Then, through the activation function, such as , to introduce nonlinearity so that the model can capture more complex temporal dependencies. Then, update the The final temporal embedding feature of the layer , which is achieved by combining the computation results of the current layer with the results of the previous layer.

[0125] Preferably, in order to improve the effect of extracting time series information, the above operation steps can be further refined or alternative solutions can be provided. and In the setting, you can consider using different initialization methods, such as random initialization or pre-trained weights, to improve the convergence speed and performance of the model. In the choice of activation function, you can consider using other types of activation functions, such as Leaky ReLU or Swish, which may provide better performance in some cases. In addition, you can consider introducing residual connections, which helps solve the gradient vanishing problem in deep network training and improve the training efficiency of the model. In practical applications, you can also adjust the calculation formula for time series information extraction according to the needs of specific tasks to achieve the best prediction effect. For example, you can prevent overfitting and improve the generalization ability of the model by adding additional regularization terms, such as dropout or weight decay.

[0126] In some embodiments, the magnetic feature distribution snapshot is expressed as:

[0127] S=[ s 1 , s 2 ,..., s n ]

[0128] The corresponding network change structure is:

[0129] [C=[ c ij ]

[0130] in, Represents a snapshot of the magnetic feature distribution; Indicates the corresponding changed network structure; Representation sample Magnetic characteristics of Representation sample and samples The similarity relationship between them.

[0131] Specifically, the implementation method includes the following key links: First, the snapshot of the magnetic feature distribution is expressed as a set of vectors, where each vector represents the magnetic feature of a sample. This vector set can be obtained by collecting and processing actual detection data, and the dimension and content of each vector depend on the characteristics of the sample and the detection requirements. Next, define the corresponding network change structure, which usually involves building a neural network model that can learn and simulate changes in the distribution of magnetic features. The design of this network structure needs to consider how to effectively represent and process magnetic features, and how to capture the relationship between features. Then, by training this network model, it can accurately predict the distribution changes of magnetic features.

[0132] Preferably, in order to improve the accuracy of the expression of the magnetic feature distribution snapshot and the network change structure, the above-mentioned operation steps can be further refined. For example, when constructing the magnetic feature distribution snapshot, dimensionality reduction techniques such as principal component analysis PCA or t-SNE can be considered to reduce the dimension of the data and highlight important features. When designing the network change structure, different types of neural networks can be considered, such as convolutional neural networks CNN or recurrent neural networks RNN, which are particularly suitable for processing data with spatial or temporal correlations. In addition, it is possible to consider introducing an attention mechanism, which helps the model pay more attention to features that have an important impact on predicting changes in the magnetic feature distribution. When training the network model, different optimization algorithms such as Adam or RMSprop can be considered, which can adaptively adjust the learning rate to accelerate the training process and improve the performance of the model. In practical applications, the expression of the magnetic feature distribution snapshot and the design of the network change structure can also be adjusted according to the needs of specific tasks to achieve the best detection effect.

[0133] The above-mentioned embodiments of the present invention have the following beneficial effects: The strong magnetic chip detection technology based on the adaptive threshold method of the present invention can provide a dynamic and adaptive detection mechanism, which can predict the distribution of magnetic features and their changes in the future by analyzing historical detection data and snapshots of magnetic feature distribution. This method can effectively combine current conditions and prediction results, and use neural networks to perform multi-task information aggregation to obtain more accurate preliminary detection results of strong magnetic chips. In addition, by calculating the multi-task composite loss function, it can assist in learning and training the network, optimize the detection model, and make it have higher accuracy and robustness in practical applications.

[0134] In addition, the present invention can more accurately extract and process magnetic features through the application of adaptive threshold modules and multi-layer perceptrons, thereby improving the reliability of detection results. At the same time, through the use of time series position coding and time domain neural network modules, time series data can be effectively processed, further improving the ability to predict future magnetic feature distribution. The combined use of these technologies can improve the efficiency and accuracy of strong magnetic chip detection, which has important practical significance for improving chip quality control and production efficiency.

[0135] The storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or terminal devices such as a computer, a server, a mobile phone, and a tablet.

[0136] It should be noted that in some embodiments of the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, apparatus, or device. In some embodiments of the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0137] In some embodiments, the client and the server may communicate using any currently known or future developed network protocol such as HTTP (HyperText Transfer Protocol), and may be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or future developed network.

[0138] In some embodiments, the client and the server may communicate using any currently known or future developed network protocol such as HTTP (HyperText Transfer Protocol), and may be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or future developed network.

[0139] The computer-readable medium may be included in the electronic device; or it may exist independently without being assembled into the electronic device. The computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device: adjusts the image resolution to a uniform size, normalizes the pixel value, and applies data enhancement technology; designs a deep learning model architecture, including multiple convolutional layers, batch normalization layers, ReLU activation functions, pooling layers, and fully connected layers; configures training parameters, including setting the initial learning rate to 0.01, the batch size to 32, the total number of iterations to 100, the momentum to 0.9, and uses the Adam optimizer for optimization; trains the model, uses the cross entropy loss function during the training process, and implements an early stopping strategy to avoid overfitting; if the performance on the validation set does not improve within 5 consecutive epochs, the learning rate is reduced to 0.1 of the original learning rate; and uses an independent test set to evaluate the performance of the trained model.

[0140] Computer program code for performing the operations of some embodiments of the present disclosure may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0141] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0142] The units described in some embodiments of the present disclosure may be implemented in software or in hardware. The units described may also be provided in a processor, and the functions described above herein may be at least partially performed by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0143] The above descriptions are only some preferred embodiments of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A strong magnetic chip detection method based on an adaptive threshold method, characterized in that: The method comprises the following steps: using historical detection data, based on chip samples under different conditions, similarity relationships between samples and magnetic characteristics of each sample, obtaining snapshots of magnetic characteristic distribution under different conditions, and calculating the changing magnetic characteristic distribution; Based on the snapshot of magnetic characteristic distribution, a main task of predicting the future magnetic characteristic distribution is established to obtain the prediction result of the future magnetic characteristic distribution; wherein the snapshot of magnetic characteristic distribution is expressed as: ； The corresponding network change structure is: ； in, Represents a snapshot of the magnetic feature distribution; Indicates the corresponding changed network structure; Representation sample Magnetic characteristics of Representation sample and samples Similarity relationship between Based on the changing magnetic feature distribution, an auxiliary task is established to predict the future changes in magnetic feature distribution under different conditions, and the prediction results of the future changes in magnetic feature distribution are obtained; The prediction results of the main task and the auxiliary task are combined with the magnetic feature distribution of the current conditions, and input into the neural network for multi-task information aggregation to obtain the final magnetic feature distribution and the corresponding preliminary detection results of the strong magnetic chip; Calculate the multi-task composite loss function, use the composite loss function to assist in learning and training the network, and obtain the trained strong magnetic chip detection model and strong magnetic chip detection results; The prediction step of the main task specifically includes: performing sample feature encoding and feature embedding based on the magnetic feature distribution snapshot using feature extraction technology to obtain initial feature embedding; Based on the initial feature embedding, an adaptive threshold module of a neural network module is used to extract overall information of the magnetic feature to obtain a low-dimensional feature embedding; Based on the low-dimensional feature embedding, an adaptive threshold module of a neural network module is used to extract information on changes in magnetic features generated under different conditions to obtain an optimized feature embedding; Embedding the optimized features into a feedforward network to obtain embedding results of each sample, and aggregating the embedding results of each sample using a linear layer of the feedforward network to obtain an overall feature embedding result; Based on the overall feature embedding result, using a prediction module to predict future magnetic feature distribution and future sample magnetic state; The prediction step of the auxiliary task specifically includes: performing sample feature encoding and feature embedding using feature extraction technology based on the changed magnetic feature distribution to obtain auxiliary initial feature embedding; Based on the auxiliary initial feature embedding, the adaptive threshold module of the neural network module is used to extract the ripple effect of the local change feature on the overall magnetic feature distribution, so as to obtain the auxiliary optimization feature embedding; Embedding the auxiliary optimized features into the input feedforward network to obtain auxiliary embedding results for each sample, and aggregating the auxiliary embedding results using a linear layer of the feedforward network to obtain auxiliary overall feature embedding results; Based on the auxiliary overall feature embedding results, a prediction module is used to predict future magnetic feature distribution changes under different conditions; The adaptive threshold module extracts information based on the following formula: ； in, For the sample , its query vector in the adaptive threshold module; represents the key vector; represents a value vector; represents the dimension of the key vector; represents the product of the attention weight and value vector of the adaptive threshold module; represents the output feature vector of sample i; Represents the sub-layer operation of the adaptive threshold module; The adaptive threshold module further extracts information based on the following formula: ； in, For the sample The query vector in the criss-cross attention module; Representation sample The key vector of ; Representation sample A vector of values ​​of ; Represents the product of the attention weight and value vector of the cross-attention module; Representation sample The output feature vector of Represents the sub-layer operation of the criss-cross attention module.

2. A strong magnetic chip detection method based on an adaptive threshold method according to claim 1, characterized in that: The multi-task information aggregation is to combine the prediction results of the future magnetic characteristic distribution and the prediction results of the future magnetic characteristic distribution change with the magnetic characteristic distribution of the current conditions, input them into a multi-layer perceptron for information aggregation, and obtain a preliminary prediction result of the final magnetic characteristic distribution; combine the preliminary prediction result with the prediction result of the future sample magnetic state, input them into another multi-layer perceptron for information aggregation, and obtain a preliminary detection result of the strong magnetic chip corresponding to the final magnetic characteristic distribution. The expression of the preliminary detection result of the strong magnetic chip is: ； in, Indicates the preliminary test results of the strong magnetic chip corresponding to the final magnetic feature distribution; It indicates the prediction results of the future distribution of magnetic characteristics; It indicates the predicted results of the future changes in the distribution of magnetic characteristics; and is the parameter matrix; ReLU represents the activation function.

3. A strong magnetic chip detection method based on an adaptive threshold method according to claim 2, characterized in that: The composite loss function is expressed as: ； in, represents the composite loss function; , , is an adjustable training parameter; represents the loss function for the final magnetic feature distribution prediction; represents the loss function for predicting future changes in magnetic feature distribution; Represents the loss function for strong magnetic chip detection.

4. The method for detecting a strong magnetic chip based on an adaptive threshold method according to claim 3, characterized in that: The specific steps of predicting the future magnetic feature distribution and the future sample magnetic state include: performing temporal position encoding on the overall feature embedding result to obtain an initial embedding feature; Extracting time series information based on the initial embedding features using an adaptive threshold module and a feedforward neural network of a time domain neural network module; Decoding the future magnetic characteristic distribution and the prediction result of the future sample magnetic state by a decoder; The calculation formula of the temporal position coding is: ； ； in, is the position index; is the dimension index; is the embedding vector dimension.

5. The method for detecting a strong magnetic chip based on an adaptive threshold method according to claim 4, characterized in that: The calculation formula for extracting the timing information is: ； ； in, Indicates The adaptive threshold calculation results of the layer; and is the parameter matrix; represents the activation function; Indicates The final temporal embedding features of the layer.

Images