Data stream intelligent reconstruction method, device and intelligent chip for integrated module

Abstract

The present invention relates to the field of intelligent chip technology, and discloses a data stream intelligent reconstruction method, device and intelligent chip for integrated modules, wherein the method comprises: performing dual-stream feature decomposition on the chip integrated module data stream to obtain a first module stream feature extraction vector and a first data stream feature extraction vector; inputting an adversarial network for feature optimization to obtain a second module stream feature extraction vector and a second data stream feature extraction vector; constructing a directed causal graph based on the chip system operating parameters and performing graph neural network encoding to obtain an operating mode embedding vector; inputting a conditional variational autoencoder network for causal decoupling to obtain a decoupled feature vector; performing probability density calculation and boundary distance calculation on the decoupled feature vector to obtain a data stream security score; performing multi-layer Bayesian perception reconstruction to obtain a target reconstructed data stream. The method effectively handles the problem of data distribution changes under different operating conditions and improves the reliability of the reconstructed data stream.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent chips, and in particular to a data stream intelligent reconstruction method and device for integrated modules and an intelligent chip. Background Art

[0002] Data flow reconstruction technology for chip integrated modules plays an increasingly important role in system performance optimization. Traditional data flow reconstruction methods mainly rely on distinguishing module features to identify data flow anomalies. This method performs well when processing data flows of known modules, but often performs poorly when facing data flows generated by unknown chip modules, and is difficult to adapt to complex and changing application scenarios.

[0003] In addition, the complex chip integration system environment leads to frequent changes in data distribution and operating conditions, which makes traditional data-driven methods face severe challenges in data flow reconstruction modeling. In particular, in practical applications, the impact of chip operating conditions on system causality is often ignored, resulting in limited generalization ability of the model under different operating conditions and inability to accurately capture the essential characteristics of the data flow. Summary of the invention

[0004] The present invention provides a data stream intelligent reconstruction method, device and intelligent chip for integrated modules. The present invention effectively handles the problem of data distribution changes under different operating conditions and improves the reliability of reconstructed data streams.

[0005] In a first aspect, the present invention provides a data flow intelligent reconstruction method for an integrated module, the data flow intelligent reconstruction method for an integrated module comprising:

[0006] Performing dual-stream feature decomposition on the chip integrated module data stream to obtain a first module stream feature extraction vector and a first data stream feature extraction vector;

[0007] Inputting the first module flow feature extraction vector and the first data flow feature extraction vector into an adversarial network for feature optimization to obtain a second module flow feature extraction vector and a second data flow feature extraction vector;

[0008] Constructing a directed causal graph based on the chip system operating parameters, and performing graph neural network encoding on the directed causal graph to obtain an operating mode embedding vector;

[0009] Inputting the operation mode embedding vector, the second module flow feature extraction vector and the second data flow feature extraction vector into a conditional variational autoencoder network for causal decoupling to obtain a decoupled feature vector;

[0010] Performing probability density calculation and boundary distance calculation on the decoupled feature vector to obtain a data flow security score;

[0011] Multi-layer Bayesian perception reconstruction is performed based on the data flow security score and the decoupled feature vector to obtain a target reconstructed data flow.

[0012] In a second aspect, the present invention provides a data stream intelligent reconstruction device for an integrated module, the data stream intelligent reconstruction device for an integrated module comprising:

[0013] A feature decomposition module, used for performing dual-stream feature decomposition on the chip integrated module data stream to obtain a first module stream feature extraction vector and a first data stream feature extraction vector;

[0014] A feature optimization module, used for inputting the first module flow feature extraction vector and the first data flow feature extraction vector into an adversarial network for feature optimization to obtain a second module flow feature extraction vector and a second data flow feature extraction vector;

[0015] A construction module is used to construct a directed causal graph based on the chip system operating parameters, and perform graph neural network encoding on the directed causal graph to obtain an operating mode embedding vector;

[0016] a decoupling module, configured to input the operation mode embedding vector, the second module flow feature extraction vector and the second data flow feature extraction vector into a conditional variational autoencoder network for causal decoupling to obtain a decoupled feature vector;

[0017] A calculation module, used to perform probability density calculation and boundary distance calculation on the decoupled feature vector to obtain a data flow security score;

[0018] A reconstruction module is used to perform multi-layer Bayesian perception reconstruction based on the data flow security score and the decoupled feature vector to obtain a target reconstructed data flow.

[0019] The third aspect of the present invention provides a data flow intelligent reconstruction device for an integrated module, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the data flow intelligent reconstruction device for the integrated module executes the above-mentioned data flow intelligent reconstruction method for the integrated module.

[0020] A fourth aspect of the present invention provides an intelligent chip, which is used to execute the above-mentioned data flow intelligent reconstruction method for integrated modules.

[0021] In the technical solution provided by the present invention, by designing a dual-stream feature decomposition strategy, the effective separation of module features and data features is achieved, the limitation of traditional methods that rely too much on specific module features is overcome, and the processing ability of unknown chip module data streams is significantly enhanced. The adversarial learning technology is used to reduce the redundant components related to the module in the data stream, improve the accuracy of feature extraction, and make the reconstruction result more accurate and reliable. The introduction of directed graph encoders and graph neural network structures accurately captures the impact of chip system operating conditions on the data stream structure, and solves the problem that traditional methods ignore the impact of operating conditions. The conditional variational autoencoder network is designed for causal decoupling, which effectively handles the problem of data distribution changes under different operating conditions and improves the adaptability of the model in complex environments. By combining probability density calculation and boundary distance calculation, a complete data stream security assessment mechanism is constructed to effectively prevent potential data leakage risks. A multi-layer Bayesian perception reconstruction strategy is adopted to take uncertainty into account while ensuring reconstruction accuracy, thereby improving the reliability of the reconstruction results.

[0022] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0023] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of an embodiment of a data stream intelligent reconstruction method for an integrated module in an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of an embodiment of a data stream intelligent reconstruction device for an integrated module according to an embodiment of the present invention;

[0026] Figure 3 It is a schematic diagram of an embodiment of a data stream intelligent reconstruction device for an integrated module in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0029] To facilitate understanding of this embodiment, firstly, a data stream intelligent reconstruction method for integrated modules disclosed in an embodiment of the present invention is described in detail. Figure 1 As shown, the method comprises the following steps:

[0030] 101. Perform dual-stream feature decomposition on the chip integrated module data stream to obtain a first module stream feature extraction vector and a first data stream feature extraction vector;

[0031] It is understandable that the execution subject of the present invention may be a data stream intelligent reconstruction device for an integrated module, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0032] Specifically, the chip integrated module data stream is input into the feature conversion layer for mapping transformation, and the original data stream is converted into a feature expression form more suitable for subsequent processing, so as to obtain a feature matrix after dimension conversion, which retains the main characteristics of the data stream and eliminates redundant dimensions. The dimension conversion feature matrix is ​​labeled with module labels, and a corresponding module label is assigned to each part of the feature matrix by combining the prior knowledge of the module or the monitoring data during operation, so as to construct a labeled training data set and a corresponding module label set. Based on the labeled training data set, a first neural network including a convolution layer, a batch normalization layer and an activation layer is constructed. The first neural network extracts the spatial features of the data through the convolution layer. The convolution layer can identify the local feature patterns in the data stream, generate the first feature map, and capture the interaction relationship between modules and the structural characteristics of the data stream. The first feature map is input into the batch normalization layer, and the data distribution is normalized by the batch normalization layer to eliminate the training instability caused by the difference in input data distribution and accelerate the convergence process to obtain the second feature map. The second feature map is input into the activation layer, and the activation layer enhances the features through nonlinear transformation operations (such as ReLU activation function) to generate a third feature map. Nonlinear transformation can effectively mine the potential complex relationships in the data and help extract the deep features of the module flow. By combining the third feature map and the module label set, the cross entropy loss function value is calculated, and the neural network is back-propagated and optimized based on this loss function value, so as to continuously adjust the weight parameters of the network and obtain the first module flow feature extraction vector. At the same time, in order to extract the features of the data flow, the initial dimension conversion feature matrix is ​​enhanced. The enhanced training data set is generated by random affine transformation and random cropping transformation. The purpose of the data enhancement operation is to simulate the diversity that may appear in the data flow in real scenes to a certain extent, so as to improve the robustness and generalization ability of the feature extraction model. Based on the enhanced training data set, a second neural network including a dimensionality reduction layer, an encoding layer and a reconstruction layer is constructed. The dimension of the feature is reduced by the dimensionality reduction layer to obtain the fourth feature map. This process helps to eliminate redundant information in the feature space and retain the main features of the data flow. The fourth feature map is input into the encoding layer, and the encoding layer compresses and expresses the features to obtain the fifth feature map. The encoding layer can effectively extract the core characteristics of the data flow by learning the potential feature space distribution. The fifth feature map is input into the reconstruction layer to restore the original features and generate the sixth feature map. Through the feature reconstruction process, the intrinsic pattern of the data stream is effectively captured. The sixth feature map is compared with the enhanced training data set, and the Euclidean distance loss value between the two is calculated. This loss value reflects the difference between the reconstructed features and the original features. By performing gradient descent optimization based on the Euclidean distance loss value, the accuracy and robustness of feature extraction are gradually improved.After the optimization is completed, the optimized feature map is input into the fully connected layer. Through the dimensionality reduction mapping of the fully connected layer, the high-dimensional features are mapped to a more compact feature space to obtain the first data stream feature extraction vector.

[0033] 102. Input the first module flow feature extraction vector and the first data flow feature extraction vector into the adversarial network for feature optimization to obtain a second module flow feature extraction vector and a second data flow feature extraction vector;

[0034] Specifically, the first module flow feature extraction vector and the first data flow feature extraction vector are feature-connected. By merging the two feature vectors along the channel dimension or the feature dimension, a mixed feature vector is generated, which contains information about the module flow feature and the data flow feature. A discriminator network is constructed based on the mixed feature vector. The discriminator network consists of a feature extraction layer and a classification layer, wherein the feature extraction layer performs deep feature extraction on the mixed feature vector to generate a discriminant feature map. The feature extraction layer adopts a convolutional neural network or a fully connected network structure, which can capture high-order characteristics in the mixed features and more accurately reflect the distribution of features through the multi-dimensional expression of the feature map. The discriminant feature map is input into the classification layer, the feature map is subjected to a binary classification operation, and the discrimination result is output. Through this process, the discriminator network can distinguish whether the mixed feature vector meets a specific optimization goal, thereby providing necessary feedback for adversarial optimization. By measuring the difference between the discrimination result and the target label, an adversarial loss function is defined. The optimization of the adversarial loss function adopts a stochastic gradient ascent method, aiming to improve the discrimination ability of the discriminator network so that it can more effectively distinguish the authenticity of the feature vector. During the stochastic gradient ascent optimization process, the weight parameters of the discriminator network are gradually adjusted to maximize the adversarial loss function value and obtain the optimized discriminator network. Gaussian random noise is input into the optimized discriminator network to evaluate the robustness of the optimized discriminator network to random features and obtain the discriminant score corresponding to the input noise. The noise discriminant score is used to guide the optimization of the first data stream feature extraction vector. In this stage, the first data stream feature extraction vector is optimized by the stochastic gradient descent method, and the representation of the feature vector is continuously adjusted in combination with the feedback of the noise discriminant score to make it more in line with the optimization goal, and the second data stream feature extraction vector is generated. At the same time, the first module stream feature extraction vector is subjected to residual connection processing to retain the original feature information during the optimization process and provide a basis for subsequent mapping operations. Through the residual connection, the first module stream feature extraction vector is combined with the incremental information in the optimization process to generate a residual feature vector. Subsequently, the residual feature vector is input into the fully connected neural network, and the potential pattern of the module stream features is mined through the feature mapping capability of the fully connected neural network to generate a mapping feature vector. The mapping feature vector is normalized. By standardizing the distribution of the feature vector, the numerical distribution of each dimension is made more uniform, thereby improving the efficiency and accuracy of the subsequent modules in using the feature vector, and obtaining the second module flow feature extraction vector.

[0035] 103. Construct a directed causal graph based on the chip system operating parameters, and perform graph neural network encoding on the directed causal graph to obtain an operating mode embedding vector;

[0036] Specifically, the chip system operating parameters are collected, including temperature parameters, voltage parameters, and load parameters, etc., to reflect the state of the chip system and its operating environment. The chip system operating parameters are normalized to generate a system state parameter matrix to ensure the consistency of parameter input and the stability of model training. An initial vertex set is constructed based on the system state parameter matrix, where each vertex represents an operating state parameter, such as the temperature, voltage, or load value of a specific module. By analyzing the logical relationship or physical dependency between the parameters, the directed connection relationship between the vertices is defined, and a directed causal graph is constructed. The edges in the directed causal graph represent the direct relationship between the vertices and reflect the directionality of the causal influence, so that the causal graph can capture the dependency pattern and information flow characteristics in complex systems. For each vertex in the directed causal graph, a vertex feature extractor including an input transformation layer and an output transformation layer is constructed. The input transformation layer maps the initial parameters of the vertex to a high-dimensional feature space. The mapping process expands the original features of the vertex into more expressive mapping features through linear transformation or nonlinear functions. The mapping features of the vertices are input to the output transformation layer. The output transformation layer compresses the high-dimensional features to a smaller but more meaningful dimension through dimensionality reduction operations to generate vertex embedding features, which effectively characterize the unique properties of the vertices in the causal graph. Based on the edge set of the directed causal graph, the attention weights are calculated for the vertex embedding features. A weight is assigned to each edge through the attention mechanism. The size of the weight reflects the relative importance of the edge. When calculating the attention weights, the relationship between the vertices is evaluated through dot products or other related methods to generate an attention score matrix. The attention score matrix is ​​matrix multiplied with the vertex embedding features to generate a weighted feature matrix. In order to improve the nonlinear expression ability of the model, the weighted feature matrix is ​​nonlinearly activated by using activation functions such as ReLU. The weighted feature matrix is ​​normalized to generate a vertex update feature matrix. The vertex update feature matrix is ​​input into the multi-layer graph convolutional network for feature aggregation. The graph convolutional network generates global features that reflect the entire graph structure by propagating and aggregating the information of the vertex neighborhood layer by layer. The stacking of multiple layers of graph convolution can capture higher-order graph information, making the final feature vector more expressive and global. After feature aggregation is completed, a graph structure feature vector is generated, which represents the comprehensive characteristics of all vertices and edges in the causal graph. Self-attention pooling is performed on the graph structure feature vector. Self-attention pooling generates an operating mode embedding vector by assigning weights to different parts of the feature vector, highlighting the important information, while suppressing irrelevant or minor features.

[0037] 104. Input the operation mode embedding vector, the second module flow feature extraction vector, and the second data flow feature extraction vector into a conditional variational autoencoder network for causal decoupling to obtain a decoupled feature vector;

[0038] Specifically, the operation mode embedding vector, the second module stream feature extraction vector and the second data stream feature extraction vector are tensor-concatenated to unify all input features into a high-dimensional joint feature tensor. The joint feature tensor contains the independent information of each input vector and captures the potential interaction between them. The joint feature tensor is input into the first fully connected layer of the conditional variational autoencoder network for feature dimensionality reduction. The dimensionality reduction process maps the joint feature tensor to a lower-dimensional space through linear transformation to obtain the first hidden layer feature. The first hidden layer feature is batch-normalized. Batch normalization can keep the numerical distribution of the feature consistent in each training batch, thereby avoiding the phenomenon of gradient vanishing or gradient exploding, and obtaining the normalized feature. The normalized feature is processed by the ReLU activation function, and the expression ability of the model is introduced through the effect of the nonlinear activation function to generate the input feature of the encoder. The encoder input feature is input into the second fully connected layer and the third fully connected layer of the conditional variational autoencoder network to complete the tasks of mean mapping and variance mapping, respectively. In the mean mapping process, the model generates the mean vector of the latent space by learning the central distribution of the joint feature tensor; while the role of variance mapping is to generate the variance vector of the latent space by learning the range of data variation. The generation of mean vector and variance vector is a key step in variational inference. These two vectors define the Gaussian distribution of the latent space, allowing the conditional variational autoencoder network to capture the potential characteristics and uncertainty of the input data. Reparameterized sampling is performed based on the latent space mean vector and the latent space variance vector to generate latent variable samples. Randomness is separated from the gradient calculation of the model so that the sampling process can be optimized. The latent variable samples are fused with the operating mode embedding vector. The fusion operation combines the representation of the latent variable with the global information of the operating mode to generate the input features of the decoder. The decoder input features are input to the first fully connected layer of the decoder network, where feature transformation is performed to generate the second hidden layer features. The first fully connected layer of the decoder network remaps the joint features of the latent variable samples and the operating mode embedding vector through linear transformation. The generated second hidden layer features are batch normalized to ensure the numerical stability and consistency of the features. The normalized features are processed by the ReLU activation function. By introducing nonlinear relationships, the model can better capture the complex characteristics of the input data and generate activation features. The activation features are input into the second fully connected layer of the decoder network for reconstruction mapping, which converts the high-dimensional activation features into the representation of the original input feature space and generates a reconstructed feature vector. In order to evaluate the reconstruction quality of the decoder, the reconstructed feature vector is compared with the joint feature tensor. The reconstruction loss and KL divergence loss between the two are calculated to measure the similarity between the reconstruction result and the input features and the difference between the potential distribution and the prior distribution. The reconstruction loss uses mean square error or cross entropy loss, while the KL divergence loss is used to constrain the closeness of the distribution of the latent space to the Gaussian prior.By combining the reconstruction loss and the KL divergence loss, the conditional variational autoencoder network is optimized. The optimization process adjusts the parameters of the network through back propagation, so that the model can achieve a balance between reconstruction quality and potential distribution fitting. The optimized conditional variational autoencoder network can efficiently implement causal decoupling operations, and the generated decoupled feature vector eliminates the redundancy and interference between input features and can more accurately characterize the essential characteristics of the data.

[0039] 105. Perform probability density calculation and boundary distance calculation on the decoupled feature vector to obtain a data flow security score;

[0040] Specifically, the Gaussian kernel density transformation of the decoupled feature vector is used to map the decoupled feature vector to a continuous probability density space. The function of the Gaussian kernel density transformation is to estimate the probability density of the decoupled feature in the potential distribution, thereby generating a feature kernel density matrix that reflects the distribution of the decoupled feature in the multidimensional space. The feature kernel density matrix is ​​input into the conditional probability calculation unit, which calculates the conditional probability distribution of the feature kernel density matrix by combining the information of the operation mode embedding vector. The operation mode embedding vector plays the role of a constraint in this process, ensuring that the conditional probability distribution can reflect the characteristics of the feature distribution under a specific operation mode. By calculating the conditional probability distribution, a conditional probability vector is generated, which represents the probability state of the decoupled feature under the current operation mode. The conditional probability vector is logarithmically transformed to transform the nonlinear distribution of the probability value into a characteristic in a linear space, and a probabilistic safety component is generated. The probabilistic safety component provides a measure that reflects the degree to which the decoupled feature deviates from the normal distribution by negative logarithmic transformation of the conditional probability. While calculating the probabilistic safety component, a safety boundary set is constructed according to a preset safety boundary threshold. The safety boundary set defines the possible value range of the decoupled feature vector under normal conditions, and generates a feature mapping vector by mapping the decoupled feature vector to the safety boundary space. The Euclidean distance between the feature mapping vector and the safety boundary set is calculated to generate a boundary distance matrix. Each value in the Euclidean distance matrix represents the distance between the feature mapping vector and the safety boundary, so that the degree of deviation of the decoupled feature from the safety boundary can be quantified. The nearest boundary point of the decoupled feature in the safety boundary set is found by searching the minimum value of the boundary distance matrix. In order to evaluate the directionality of the boundary distance, the minimum value is inverted to generate the boundary safety component. The boundary safety component can accurately describe the direction and degree of deviation of the decoupled feature in the safety boundary. In order to integrate the information of the probabilistic safety component and the boundary safety component, a weighted combination function is constructed based on these two components. The weighted combination function can dynamically adjust the contribution ratio of the probabilistic safety component and the boundary safety component by designing the combination weight coefficient. The determination of the combination weight coefficient is based on the analysis of historical data or prior knowledge to ensure that the result after weighted summation can reflect the real safety status. The probabilistic security component and the boundary security component are weighted and summed according to the combined weight coefficient to generate a unified metric value, namely the data flow security score.

[0041] 106. Multi-layer Bayesian perception reconstruction is performed based on the data flow security score and decoupled feature vector to obtain the target reconstructed data flow.

[0042] Specifically, a threshold judgment is performed on the data stream security score. By analyzing the data stream security score, it is compared with the preset security threshold to generate a security mask vector, which distinguishes the security features from the potential abnormal features in the data stream. The security mask vector is combined with the decoupled feature vector, and the abnormal or irrelevant features are removed through feature filtering operations, and only the security features are retained to generate a security feature vector. Based on the security feature vector, a Bayesian neural network including a priori layer and a likelihood layer is constructed. The prior layer generates a priori weight matrix by sampling the distribution of the neural network weights, which reflects the initial probability distribution of the network weights. It can introduce prior knowledge of the weight parameters, thereby improving the robustness and generalization ability of the model. At the same time, the likelihood layer models the input security feature vector and generates a likelihood probability vector by capturing the distribution pattern of the features. The modeling result of the likelihood layer reflects the distribution of the features in the input data space. The priori weight matrix and the likelihood probability vector are input into the variational inference module. In the variational inference module, the posterior distribution parameters are generated through sampling operations. The posterior distribution parameters integrate the prior weight information and the likelihood distribution information, and can dynamically adjust the neural network's understanding of the input features. The maximum posterior probability is calculated for the posterior distribution parameters to determine the optimal posterior distribution solution. The optimal reconstruction parameters are calculated to represent the optimal model weight configuration of the Bayesian neural network in the process of reconstructing the data stream. The optimal reconstruction parameters and the security feature vector are input into the reconstruction decoder. The core of the reconstruction decoder is a multi-layer perceptron, which converts the security feature vector into a reconstructed feature vector through layer-by-layer feature mapping operations. The multi-layer perceptron uses the expression power of nonlinear transformation and high-dimensional feature space to gradually restore the structural characteristics and distribution pattern of the input data and generate a high-quality reconstructed feature representation. The reconstructed feature vector is denormalized and dimensionally transformed. The feature vector is converted into the initial reconstructed data stream by restoring the original numerical range and structural form of the input data. The denormalization process is based on the inverse process of the normalization operation, so that the numerical distribution of the data stream conforms to the actual situation of the original input; the dimension transformation ensures that the reconstructed data stream is structurally consistent with the target data stream. The target reconstructed data stream is generated by performing a weighted combination operation on the initial reconstructed data stream and the data stream security score. Combining the reconstruction results and security assessment information, it is ensured that the final target reconstructed data stream has both high security and can accurately reflect the main characteristics and structure of the original data stream.

[0043] In the embodiment of the present invention, by designing a dual-stream feature decomposition strategy, the effective separation of module features and data features is achieved, the limitation of traditional methods that rely too much on specific module features is overcome, and the processing ability of unknown chip module data streams is significantly enhanced. The adversarial learning technology is used to reduce the redundant components related to the module in the data stream, improve the accuracy of feature extraction, and make the reconstruction result more accurate and reliable. The introduction of directed graph encoders and graph neural network structures accurately captures the impact of chip system operating conditions on the data stream structure, and solves the problem that traditional methods ignore the impact of operating conditions. The conditional variational autoencoder network is designed for causal decoupling, which effectively handles the problem of data distribution changes under different operating conditions and improves the adaptability of the model in complex environments. By combining probability density calculation and boundary distance calculation, a complete data stream security assessment mechanism is constructed to effectively prevent potential data leakage risks. A multi-layer Bayesian perception reconstruction strategy is adopted to take uncertainty into account while ensuring reconstruction accuracy, thereby improving the reliability of the reconstruction results.

[0044] In a specific embodiment, the process of executing step 101 may specifically include the following steps:

[0045] The chip integrated module data stream is input into the feature conversion layer for mapping transformation to obtain a dimension conversion feature matrix, and the dimension conversion feature matrix is ​​annotated with module labels to obtain an annotated training data set and a corresponding module label set;

[0046] Based on the labeled training data set, a first neural network including a convolution layer, a batch normalization layer and an activation layer is constructed, and spatial features are extracted through the convolution layer to obtain a first feature map; the first feature map is input into the batch normalization layer for data distribution normalization to obtain a second feature map; the second feature map is input into the activation layer for nonlinear transformation to obtain a third feature map;

[0047] Calculate the cross entropy loss function value according to the third feature map and the module label set, and perform back propagation optimization based on the cross entropy loss function value to obtain the first module flow feature extraction vector;

[0048] Perform random affine transformation and random cropping transformation on the dimension transformation feature matrix to obtain an enhanced training data set;

[0049] Based on the enhanced training data set, a second neural network including a dimensionality reduction layer, an encoding layer and a reconstruction layer is constructed, and the feature dimension is reduced by the dimensionality reduction layer to obtain a fourth feature map; the fourth feature map is input into the encoding layer for feature encoding to obtain a fifth feature map; the fifth feature map is input into the reconstruction layer for feature reconstruction to obtain a sixth feature map;

[0050] The Euclidean distance loss value is calculated for the sixth feature map and the enhanced training data set, and gradient descent optimization is performed based on the Euclidean distance loss value to obtain an optimized feature map, and the optimized feature map is input into the fully connected layer for dimensionality reduction mapping to obtain a first data stream feature extraction vector.

[0051] Specifically, the chip integrated module data stream is input into the feature conversion layer for mapping transformation. The original high-dimensional data stream (such as multi-channel time series data) is converted into a feature matrix form that is more suitable for processing, that is, the dimension conversion feature matrix, recorded as ,in Represents the number of features (e.g., the dimensions of physical parameters such as temperature and voltage), Represents the number of time steps or sampling points. For example, if the data stream contains three parameters (temperature, voltage, load) sampled at 10 time steps, the dimension conversion feature matrix is .right Mark the module label. Assume that the module label is ,in Indicates the category of the module to which each feature belongs, for example, temperature belongs to module 1, voltage belongs to module 2, and load belongs to module 3. The labeled training data set is Based on the labeled training data set, the first neural network consisting of convolution layer, batch normalization layer and activation layer is constructed. The spatial features of the data are extracted through the convolution layer. The convolution operation is assumed to be:

[0052] Conv ;

[0053] in represents the weight matrix of the convolution kernel, Indicates bias, is the first feature map. The convolution layer performs weighted summation of local area features through a sliding window. , used to extract local correlation. The input batch normalization layer performs normalization to reduce the instability of the input data distribution. The normalized feature map is recorded as , and its calculation formula is:

[0054] ;

[0055] in and They are The mean and variance of is a small constant to prevent division by zero. The input activation layer is transformed nonlinearly to obtain the third feature map ReLU , where the ReLU function is defined as ReLU , used to introduce nonlinearity. According to the third characteristic diagram and module tagset , define the cross entropy loss function:

[0056] ;

[0057] in is the predicted probability value of the network, is the true label. By back-propagating and optimizing the loss function, the network weights are continuously adjusted to finally extract the first module flow feature vector . Transform the feature matrix to Perform random affine transformation and random cropping transformation. These enhancement operations simulate the diversity of data, such as adding noise or cropping some time steps, to obtain an enhanced training data set For example, for The random affine transformation is expressed as ,in is a randomly generated affine matrix, is the bias vector. Based on the enhanced training data set, a second neural network consisting of a dimensionality reduction layer, an encoding layer, and a reconstruction layer is constructed. The dimensionality reduction layer is used to reduce high-dimensional features to smaller dimensions to eliminate redundancy. The features after dimensionality reduction are recorded as . Coding layer pair Perform feature encoding to obtain the fifth feature map , the reconstruction layer decodes the encoded features into a structure similar to the input, generating the sixth feature map . By calculating the Euclidean distance loss between the sixth feature map and the enhanced training dataset:

[0058] ;

[0059] Based on this loss, gradient descent optimization is performed to improve the quality of the feature. Input the fully connected layer to complete the dimensionality reduction mapping and generate the first data stream feature extraction vector The vector highly summarizes and refines the complex features of the data stream through the learning network, providing high-quality input for subsequent tasks.

[0060] In a specific embodiment, the process of executing step 102 may specifically include the following steps:

[0061] Performing feature connection on the first module stream feature extraction vector and the first data stream feature extraction vector to obtain a mixed feature vector;

[0062] A discriminator network is constructed based on the mixed feature vector. The discriminator network includes a feature extraction layer and a classification layer. The mixed feature vector is extracted by the feature extraction layer to obtain a discriminant feature map. The discriminant feature map is input into the classification layer for binary classification to obtain a discriminant result.

[0063] Based on the discrimination results, the adversarial loss function is calculated, and the discriminator network is optimized by stochastic gradient ascent to obtain the optimized discriminator network. Gaussian random noise is input into the optimized discriminator network for feature discrimination to obtain the noise discrimination score.

[0064] Performing random gradient descent optimization on the first data stream feature extraction vector according to the noise discrimination score to obtain a second data stream feature extraction vector, and performing residual connection processing on the first module stream feature extraction vector to obtain a residual feature vector;

[0065] The residual feature vector is input into a fully connected neural network for feature mapping to obtain a mapping feature vector, and the mapping feature vector is normalized to obtain a second module flow feature extraction vector.

[0066] Specifically, extract the vector of the first module flow feature and the first data stream feature extraction vector Perform feature connection to obtain a mixed feature vector The feature connection is completed through vector concatenation operation, and the formula is expressed as:

[0067] ;

[0068] in and are the dimensions of the module flow feature vector and the data flow feature vector respectively. For example, if Represents the characteristics of the five modules, Represents three data stream features, then the mixed feature vector Based on the mixed feature vector Construct a discriminator network, which includes a feature extraction layer and a classification layer. The feature extraction layer extracts features from the mixed feature vector through a series of linear transformations and nonlinear activation operations. For example, the output of the feature extraction layer is It is expressed as:

[0069] ;

[0070] in is the weight matrix of the feature extraction layer, is the bias vector, is the number of features extracted. For example, if and , then the weight matrix dimension is ReLU is a nonlinear activation function used to introduce nonlinear capabilities of the model. Input classification layer, the classification layer outputs the discrimination result through a binary classification logic function , which is expressed as:

[0071] ;

[0072] in is the sigmoid function, and are the weights and biases of the classification layer. For example, if the discriminant feature map dimension is , the weight matrix of the classification layer is Based on the judgment results and the true label , define the adversarial loss function. The adversarial loss function is the binary cross entropy:

[0073] ;

[0074] in is the batch size, It is The true labels of samples, is the predicted probability value. The loss function is optimized by stochastic gradient ascent to improve the discriminator network's ability to distinguish mixed features and obtain the optimized discriminator network. Input the optimized discriminator network and calculate the noise discrimination score s. The noise discrimination score indicates the network's response to random features and reflects the discriminator's sensitivity to abnormal inputs. Based on the noise discrimination score s, extract the feature vector of the first data stream. Perform stochastic gradient descent optimization, and the update formula is:

[0075] ;

[0076] in is the optimized feature vector of the second data stream, is the learning rate, is the loss function related to the noise discrimination score. At the same time, the first module flow feature extraction vector Perform residual connection processing to generate residual feature vector , the formula is:

[0077] ;

[0078] in is the optimization increment, reflecting the updated information of the module features. Input the fully connected neural network and perform feature mapping to generate mapping feature vector , the formula is:

[0079] ;

[0080] in and are the weights and biases of the fully connected network. Normalize and obtain the second module flow feature extraction vector , the normalized formula is:

[0081] ;

[0082] in and are the mean and variance respectively, is a small constant to prevent division by zero. Through the above steps, high-quality feature representation is generated.

[0083] In a specific embodiment, the process of executing step 103 may specifically include the following steps:

[0084] The chip system operating parameters are collected and normalized to obtain a system state parameter matrix, which includes temperature parameters, voltage parameters and load parameters;

[0085] Construct an initial vertex set based on the system state parameter matrix, and construct directed connection relationships between vertices to obtain a directed causal graph;

[0086] For each vertex in the directed causal graph, a vertex feature extractor including an input transformation layer and an output transformation layer is constructed. The vertex parameters are mapped to the feature space through the input transformation layer to obtain the vertex mapping features. The vertex mapping features are input into the output transformation layer for dimensionality reduction transformation to obtain the vertex embedding features.

[0087] Based on the edge set of the directed causal graph, the attention weights are calculated for the vertex embedding features to obtain the attention score matrix;

[0088] Perform matrix multiplication on the attention score matrix and the vertex embedding feature to obtain a weighted feature matrix, and perform nonlinear activation and normalization on the weighted feature matrix to obtain a vertex update feature matrix;

[0089] The vertex update feature matrix is ​​input into the multi-layer graph convolutional network for feature aggregation to obtain the graph structure feature vector, and the graph structure feature vector is self-attention pooled to obtain the operation mode embedding vector.

[0090] Specifically, the operating parameters are collected from the chip system, including temperature parameters, voltage parameters and load parameters. These parameters are represented as a multidimensional matrix, namely the system state parameter matrix, denoted as ,in is the number of parameter types (such as temperature, voltage, load, etc.), is the sampling time step. In order to eliminate the dimension difference and value range difference of the parameters, Normalization is performed. Normalization is achieved using the following formula:

[0091] ;

[0092] in yes The mean of is the standard deviation. Based on the normalized system state parameter matrix , construct the initial vertex set. Each vertex represents an operating parameter (such as the temperature and voltage of a module), and the vertex set is recorded as In order to capture the causal relationship between operating parameters, we define the directed connection relationship between vertices, such as temperature affecting load, and thus construct the edge set . Combine vertices and edges to form a directed causal graph , and use the adjacency matrix represents the connection strength of the directed edge. For example, if (Temperature) (voltage) has a strong influence, then In order to extract the features of the vertices, A vertex feature extractor is constructed for each vertex in . The vertex feature extractor consists of an input transformation layer and an output transformation layer. The input transformation layer maps vertex parameters to a high-dimensional feature space. The formula is expressed as:

[0093] ;

[0094] in It is The parameter vector of the vertices, is the weight matrix of the input transformation, is the bias vector, is the feature dimension after mapping. The output transformation layer is Perform dimensionality reduction transformation to generate vertex embedding features , the formula is:

[0095] ;

[0096] in is the dimension reduction weight matrix, is the bias. Based on the edge set of the directed causal graph , the attention weight is calculated for the vertex embedding features. The attention mechanism assigns a weight to each edge by learning the importance of adjacent vertices. The calculation formula is:

[0097] ;

[0098] in From the vertex arrive The attention weight, yes The neighbor set of is a learnable weight vector, and || represents the feature concatenation operation. Vertex Embedding Features Perform matrix multiplication to generate the weighted feature matrix:

[0099] ;

[0100] right Perform nonlinear activation and normalization to generate vertex update feature matrix , the formula is:

[0101] ;

[0102] And normalize it through batch normalization. Input a multi-layer graph convolutional network (GCN) to perform global feature aggregation and generate a graph structure feature vector The graph convolution formula is:

[0103] ;

[0104] in It is The characteristics of the layer, It is The weight matrix of the layer, is the activation function. Perform self-attention pooling to generate the operating mode embedding vector , the formula is:

[0105] ;

[0106] in is the attention weight, is the learning vector. Through the above steps, an embedding vector reflecting the system operation status is generated.

[0107] In a specific embodiment, the process of executing step 104 may specifically include the following steps:

[0108] Perform tensor concatenation of the operation mode embedding vector, the second module stream feature extraction vector, and the second data stream feature extraction vector to obtain a joint feature tensor;

[0109] Input the joint feature tensor into the first fully connected layer of the conditional variational autoencoder network to perform feature dimensionality reduction to obtain the first hidden layer feature, perform batch normalization on the first hidden layer feature to obtain the normalized feature, perform ReLU activation on the normalized feature to obtain the encoder input feature;

[0110] The encoder input feature is input into the second fully connected layer of the conditional variational autoencoder network for mean mapping to obtain a latent space mean vector, and the encoder input feature is input into the third fully connected layer of the conditional variational autoencoder network for variance mapping to obtain a latent space variance vector;

[0111] Reparameterized sampling is performed based on the latent space mean vector and the latent space variance vector to obtain latent variable samples, and feature fusion is performed on the latent variable samples and the operating mode embedding vector to obtain the decoder input features;

[0112] Conditionally decode and reconstruct the decoder input features, input the decoder input features into the first fully connected layer of the decoder network for feature transformation to obtain the second hidden layer features, perform batch normalization and ReLU activation processing on the second hidden layer features to obtain activation features;

[0113] Input the activated features into the second fully connected layer of the decoder network for reconstruction mapping to obtain a reconstructed feature vector, and calculate the reconstruction loss and KL divergence loss of the reconstructed feature vector and the joint feature tensor;

[0114] The conditional variational autoencoder network is optimized based on reconstruction loss and KL divergence loss to obtain the decoupled feature vector.

[0115] Specifically, the operation mode is embedded in the vector , the second module stream feature extraction vector and the second data stream feature extraction vector Perform tensor concatenation to generate joint feature tensors The concatenation operation combines the three features into a unified space, and the formula is expressed as:

[0116] ;

[0117] in Represents a concatenation operation. The joint feature tensor contains information about the operation mode, module flow features, and data flow features. Input to the first fully connected layer of the conditional variational autoencoder network to perform feature dimensionality reduction. The first hidden layer feature after dimensionality reduction is recorded as , and its calculation formula is:

[0118] ;

[0119] in is the weight matrix, is the bias vector, is the feature dimension after dimensionality reduction. To improve the stability of training, Perform batch normalization to generate normalized features , and its calculation formula is:

[0120] ;

[0121] in and are the batch mean and variance, is a small value to avoid division by zero. The normalized features are processed by the ReLU activation function to generate the encoder input features , the formula is:

[0122] ;

[0123] Will Input to the second and third fully connected layers of the conditional variational autoencoder network to generate the mean vector of the latent space and variance vector These two vectors define the distribution of the latent space, and the calculation formulas for the mean and variance are:

[0124] ;

[0125] in , , is the dimension of the latent space. The latent space is sampled by reparameterization techniques to generate latent variable samples , the formula is:

[0126] ;

[0127] in represents element-wise multiplication, is noise sampled from a standard normal distribution. Latent variable sample Embedding vector with running mode Perform feature fusion to generate decoder input features , the formula is:

[0128] ;

[0129] in Indicates a concatenation operation. Input to the first fully connected layer of the decoder network, perform feature transformation, and generate the second hidden layer features , and its calculation formula is:

[0130] ;

[0131] in , .right Perform batch normalization and ReLU activation processing to generate activation features The activated features are then input into the second fully connected layer of the decoder network to perform the reconstruction mapping operation and generate the reconstructed feature vector , the formula is:

[0132] ;

[0133] in , . The eigenvector is reconstructed by calculating With the joint feature tensor The reconstruction loss and the KL divergence loss of the latent distribution , optimize the parameters of the conditional variational autoencoder network model, and the sum of the loss functions is:

[0134] ;

[0135] in is a weight factor used to balance reconstruction and distribution fitting. The reconstruction loss uses the mean square error:

[0136] ;

[0137] The KL divergence loss is:

[0138] ;

[0139] By optimizing the above loss function, we finally get the decoupled feature vector , which contains the core information of the input features and retains the modeling ability of the conditional variational autoencoder network for the potential distribution, thereby achieving efficient reconstruction and feature extraction of the input data.

[0140] In a specific embodiment, the process of executing step 105 may specifically include the following steps:

[0141] Perform Gaussian kernel density transformation on the decoupled feature vector to obtain a feature kernel density matrix, and input the feature kernel density matrix into a conditional probability calculation unit to calculate the conditional probability distribution based on the operation mode embedding vector to obtain a conditional probability vector;

[0142] Perform logarithmic transformation on the conditional probability vector to obtain the probabilistic safety component, and construct a safety boundary set according to the preset safety boundary threshold, map the decoupled feature vector to the safety boundary space, and obtain the feature mapping vector;

[0143] Calculate the Euclidean distance matrix for the feature mapping vector and the safety boundary set to obtain the boundary distance vector, and perform minimum value search and sign inversion operation on the boundary distance vector to obtain the boundary safety component;

[0144] A weighted combination function is constructed based on the probabilistic security component and the boundary security component to obtain the combination weight coefficient, and a weighted sum operation is performed on the probabilistic security component and the boundary security component according to the combination weight coefficient to obtain the data flow security score.

[0145] Specifically, the decoupled eigenvector is transformed by Gaussian kernel density to obtain the characteristic kernel density matrix .in, represents the number of sampling points, is the dimension of the feature vector. The formula for Gaussian kernel density estimation is:

[0146] ;

[0147] in is the smoothing parameter, is the eigenvector and sampling points For example, if the decoupled eigenvector Contains 3 dimensions (temperature, voltage, load) and ,but Represents the density value of each sampling point in different dimensions. Input conditional probability calculation unit, and the operating mode embedding vector Combined with, it is used to calculate the conditional probability distribution. The formula for conditional probability distribution is:

[0148] ;

[0149] The joint probability The dot product operation is calculated as:

[0150] ;

[0151] and is a normalization factor that ensures that the conditional probabilities sum to 1. For example, if is the probability vector of the operating mode [0.4, 0.6]. The joint probability distribution adjusts the value weights of the kernel density matrix to obtain the conditional probability vector . For the conditional probability vector Perform logarithmic transformation to generate the probabilistic safety component , the formula is:

[0152] ;

[0153] The logarithmic transformation is used to reflect the degree of deviation of the feature in the low probability area. Construct a safe boundary set. The goal of the safe boundary set is to define the normal range of the feature. Decouple the feature vector Mapping to the safe boundary space generates a feature mapping vector , the formula is:

[0154] ;

[0155] in The operation ensures that the eigenvalue is not lower than the threshold. The Euclidean distance between the feature mapping vector and the safety boundary set is calculated to generate the boundary distance matrix , the formula is:

[0156] ;

[0157] in Represents the first Boundary points, is the Euclidean distance. The nearest boundary point distance of each feature is found by searching for the minimum value, and the boundary distance vector is obtained. , and generate the boundary safety component by negating the sign:

[0158] ;

[0159] Probabilistic safety component and border security component , construct a weighted combination function to generate the combination weight coefficient w, the formula is:

[0160] ;

[0161] in Is a weight parameter that reflects the relative importance of the two security components. The probability security component and the boundary security component are weighted and summed according to the combined weight coefficient to obtain the data flow security score. , the formula is:

[0162] ;

[0163] This security score combines the probability distribution information of the feature and the degree of boundary deviation, thus providing a reliable quantitative indicator for evaluating the security of data streams.

[0164] In a specific embodiment, the process of executing step 106 may specifically include the following steps:

[0165] Perform threshold judgment on the data flow security score to obtain a security mask vector, and perform feature filtering on the decoupled feature vector and the security mask vector to obtain a security feature vector;

[0166] A Bayesian neural network consisting of a priori layer and a likelihood layer is constructed for the security feature vector. The weight distribution is sampled through the priori layer to obtain a priori weight matrix, and the feature distribution is modeled through the likelihood layer to obtain a likelihood probability vector.

[0167] The prior weight matrix and likelihood probability vector are input into the variational inference module for sampling to obtain the posterior distribution parameters, and the maximum posterior probability is calculated for the posterior distribution parameters to obtain the optimal reconstruction parameters;

[0168] The security feature vector and the optimal reconstruction parameter are input into the reconstruction decoder, and feature mapping is performed through the multi-layer perceptron to obtain the reconstructed feature vector;

[0169] The reconstructed feature vector is denormalized and dimensionally transformed to obtain an initial reconstructed data stream, and a weighted combination operation is performed based on the initial reconstructed data stream and the data stream security score to obtain the target reconstructed data stream.

[0170] Specifically, score the data flow security Perform threshold judgment to distinguish normal features from potential abnormal features. Assume that the preset safety score threshold is , the generated security mask vector The formula is:

[0171] ;

[0172] in Is the security mask An element is used to mark whether a feature complies with safety standards. For example, if [0.2,0.8,0.5] and ,but . Set the security mask vector With the decoupled eigenvector Combined, perform feature filtering operations to obtain security feature vectors :

[0173] ;

[0174] in represents element-by-element multiplication. For example, if and ,but Based on the security feature vector Construct a Bayesian neural network consisting of a prior layer and a likelihood layer. The prior layer models the network weight distribution and generates a prior weight matrix Assuming that the weights follow a normal distribution, the prior sampling formula is:

[0175] ;

[0176] in and are the prior mean and standard deviation, initialized to zero mean and unit variance respectively. At the same time, the likelihood layer models the feature distribution and generates a likelihood probability vector , the formula is:

[0177] ;

[0178] in is the sigmoid function, is the bias vector. The prior weight matrix and the likelihood vector Input variational inference module and obtain the posterior distribution parameters through sampling , the formula is:

[0179] ;

[0180] The posterior mean and standard deviation are determined by the optimization process of variational inference. The optimal reconstruction parameters of the posterior distribution are calculated by the maximum a posteriori probability method. :

[0181] ;

[0182] The security feature vector and the optimal reconstruction parameters Input reconstruction decoder, perform feature mapping through multi-layer perceptron, and generate reconstructed feature vector The calculation formula of the multi-layer perceptron is:

[0183] ;

[0184] in and are the weights and biases of the perceptron. Perform anti-normalization processing, the formula is:

[0185] ;

[0186] in and are the standard deviation and mean of the initial normalization. The denormalized feature vector undergoes dimension transformation to generate the initial reconstructed data stream. Based on the initial reconstructed data stream and data stream security score Perform weighted combination operations to generate target reconstruction data stream :

[0187] ;

[0188] in is the weight coefficient used to balance the data flow reconstruction result and the security score.

[0189] The above describes the data flow intelligent reconstruction method for the integrated module in the embodiment of the present invention. The following describes the data flow intelligent reconstruction device for the integrated module in the embodiment of the present invention. Figure 2 , an embodiment of the data stream intelligent reconstruction device for integrated modules in an embodiment of the present invention includes:

[0190] The feature decomposition module 201 is used to perform dual-stream feature decomposition on the chip integrated module data stream to obtain a first module stream feature extraction vector and a first data stream feature extraction vector;

[0191] A feature optimization module 202, used for inputting the first module flow feature extraction vector and the first data flow feature extraction vector into the adversarial network for feature optimization to obtain a second module flow feature extraction vector and a second data flow feature extraction vector;

[0192] A construction module 203 is used to construct a directed causal graph based on the chip system operating parameters, and perform graph neural network encoding on the directed causal graph to obtain an operating mode embedding vector;

[0193] A decoupling module 204 is used to input the operation mode embedding vector, the second module flow feature extraction vector and the second data flow feature extraction vector into a conditional variational autoencoder network for causal decoupling to obtain a decoupled feature vector;

[0194] A calculation module 205 is used to perform probability density calculation and boundary distance calculation on the decoupled feature vector to obtain a data flow security score;

[0195] The reconstruction module 206 is used to perform multi-layer Bayesian perception reconstruction based on the data flow security score and the decoupled feature vector to obtain a target reconstructed data flow.

[0196] Through the collaboration of the above components and the design of a dual-stream feature decomposition strategy, the module features and data features are effectively separated, overcoming the limitation of traditional methods that rely too much on specific module features, and significantly enhancing the processing ability of unknown chip module data streams. The adversarial learning technology is used to reduce the redundant components related to the module in the data stream, improve the accuracy of feature extraction, and make the reconstruction results more accurate and reliable. The directed graph encoder and graph neural network structure are introduced to accurately capture the impact of the chip system operating conditions on the data stream structure, solving the problem that the traditional method ignores the impact of operating conditions. The conditional variational autoencoder network is designed for causal decoupling, which effectively handles the problem of data distribution changes under different operating conditions and improves the adaptability of the model in complex environments. By combining probability density calculation and boundary distance calculation, a complete data stream security assessment mechanism is constructed to effectively prevent potential data leakage risks. The multi-layer Bayesian perception reconstruction strategy is adopted to take into account uncertainty while ensuring the reconstruction accuracy, thereby improving the reliability of the reconstruction results.

[0197] above Figure 2 The data flow intelligent reconstruction device for integrated modules in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The data flow intelligent reconstruction device for integrated modules in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0198] Figure 3 3 is a structural diagram of a data stream intelligent reconstruction device for an integrated module provided by an embodiment of the present invention. The data stream intelligent reconstruction device 300 for an integrated module may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 may be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the data stream intelligent reconstruction device 300 for the integrated module. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the data stream intelligent reconstruction device 300 for the integrated module to implement the steps of the above-mentioned data stream intelligent reconstruction method for the integrated module.

[0199] The data stream intelligent reconstruction device 300 for integrated modules may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 3 The structure of the data flow intelligent reconstruction device for integrated modules shown does not constitute a limitation on the data flow intelligent reconstruction device for integrated modules provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0200] The present invention also provides an intelligent chip, which is used to execute the steps of the data stream intelligent reconstruction method oriented to the integrated module.

[0201] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0202] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0203] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data flow intelligent reconstruction method for integrated modules, characterized in that: The method comprises: Performing dual-stream feature decomposition on the chip integrated module data stream to obtain a first module stream feature extraction vector and a first data stream feature extraction vector; Inputting the first module flow feature extraction vector and the first data flow feature extraction vector into an adversarial network for feature optimization to obtain a second module flow feature extraction vector and a second data flow feature extraction vector; A directed causal graph is constructed based on the chip system operating parameters, and the directed causal graph is encoded by a graph neural network to obtain an operating mode embedding vector; specifically, the method includes: collecting and normalizing the chip system operating parameters to obtain a system state parameter matrix, wherein the system state parameter matrix includes temperature parameters, voltage parameters, and load parameters; an initial vertex set is constructed based on the system state parameter matrix, and a directed connection relationship between vertices is constructed to obtain a directed causal graph; a vertex feature extractor including an input transformation layer and an output transformation layer is constructed for each vertex in the directed causal graph, and vertex parameters are mapped to a feature space through the input transformation layer to obtain a vertex mapping Features, input the vertex mapping features into the output transformation layer for dimensionality reduction transformation to obtain vertex embedding features; based on the edge set of the directed causal graph, the attention weights of the vertex embedding features are calculated to obtain an attention score matrix; the attention score matrix is ​​matrix multiplied with the vertex embedding features to obtain a weighted feature matrix, and the weighted feature matrix is ​​nonlinearly activated and normalized to obtain a vertex update feature matrix; the vertex update feature matrix is ​​input into a multi-layer graph convolutional network for feature aggregation to obtain a graph structure feature vector, and the graph structure feature vector is self-attention pooled to obtain an operation mode embedding vector; The operation mode embedding vector, the second module flow feature extraction vector and the second data flow feature extraction vector are input into the conditional variational autoencoder network for causal decoupling to obtain a decoupled feature vector; specifically including: tensor splicing the operation mode embedding vector, the second module flow feature extraction vector and the second data flow feature extraction vector to obtain a joint feature tensor; inputting the joint feature tensor into the first fully connected layer of the conditional variational autoencoder network for feature dimension reduction to obtain a first hidden layer feature, performing batch normalization processing on the first hidden layer feature to obtain a normalized feature, performing ReLU activation processing on the normalized feature to obtain an encoder input feature; inputting the encoder input feature into the second fully connected layer of the conditional variational autoencoder network for mean mapping to obtain a latent space mean vector, inputting the encoder input feature into the third fully connected layer of the conditional variational autoencoder network for Variance mapping to obtain a latent space variance vector; reparameterization sampling is performed based on the latent space mean vector and the latent space variance vector to obtain a latent variable sample, and feature fusion is performed on the latent variable sample and the operating mode embedding vector to obtain a decoder input feature; conditional decoding reconstruction is performed on the decoder input feature, the decoder input feature is input into the first fully connected layer of the decoder network for feature transformation to obtain a second hidden layer feature, the second hidden layer feature is batch normalized and ReLU activated to obtain an activation feature; the activation feature is input into the second fully connected layer of the decoder network for reconstruction mapping to obtain a reconstructed feature vector, and the reconstruction loss and KL divergence loss of the reconstructed feature vector and the joint feature tensor are calculated; the conditional variational autoencoder network is optimized based on the reconstruction loss and the KL divergence loss to obtain a decoupled feature vector; Performing probability density calculation and boundary distance calculation on the decoupled feature vector to obtain a data flow security score; Multi-layer Bayesian perception reconstruction is performed based on the data flow security score and the decoupled feature vector to obtain a target reconstructed data flow.

2. The data flow intelligent reconstruction method for integrated modules according to claim 1 is characterized in that: The step of performing dual-stream feature decomposition on the chip integrated module data stream to obtain a first module stream feature extraction vector and a first data stream feature extraction vector includes: Input the chip integrated module data stream into the feature conversion layer for mapping transformation to obtain a dimension conversion feature matrix, and annotate the dimension conversion feature matrix with module labels to obtain an annotated training data set and a corresponding module label set; Based on the labeled training data set, a first neural network including a convolution layer, a batch normalization layer and an activation layer is constructed, and spatial features are extracted through the convolution layer to obtain a first feature map; the first feature map is input into the batch normalization layer for data distribution normalization to obtain a second feature map; the second feature map is input into the activation layer for nonlinear transformation to obtain a third feature map; Calculating a cross entropy loss function value according to the third feature map and the module label set, and performing back propagation optimization based on the cross entropy loss function value to obtain a first module flow feature extraction vector; Performing random affine transformation and random cropping transformation on the dimension conversion feature matrix to obtain an enhanced training data set; Based on the enhanced training data set, a second neural network including a dimensionality reduction layer, an encoding layer and a reconstruction layer is constructed, and the feature dimension is reduced by the dimensionality reduction layer to obtain a fourth feature map; the fourth feature map is input into the encoding layer for feature encoding to obtain a fifth feature map; the fifth feature map is input into the reconstruction layer for feature reconstruction to obtain a sixth feature map; The Euclidean distance loss value is calculated for the sixth feature map and the enhanced training data set, and gradient descent optimization is performed based on the Euclidean distance loss value to obtain an optimized feature map, and the optimized feature map is input into a fully connected layer for dimensionality reduction mapping to obtain a first data stream feature extraction vector.

3. The data flow intelligent reconstruction method for integrated modules according to claim 2 is characterized in that: The step of inputting the first module flow feature extraction vector and the first data flow feature extraction vector into an adversarial network for feature optimization to obtain a second module flow feature extraction vector and a second data flow feature extraction vector comprises: Performing feature connection on the first module stream feature extraction vector and the first data stream feature extraction vector to obtain a mixed feature vector; A discriminator network is constructed based on the mixed feature vector, wherein the discriminator network comprises a feature extraction layer and a classification layer, wherein the mixed feature vector is subjected to feature extraction by the feature extraction layer to obtain a discriminant feature map; the discriminant feature map is input into the classification layer for binary classification to obtain a discriminant result; Calculate the adversarial loss function based on the discrimination result, perform stochastic gradient ascent optimization on the discriminator network to obtain an optimized discriminator network, and input Gaussian random noise into the optimized discriminator network for feature discrimination to obtain a noise discrimination score; Performing random gradient descent optimization on the first data stream feature extraction vector according to the noise discrimination score to obtain a second data stream feature extraction vector, and performing residual connection processing on the first module stream feature extraction vector to obtain a residual feature vector; The residual feature vector is input into a fully connected neural network for feature mapping to obtain a mapping feature vector, and the mapping feature vector is normalized to obtain a second module flow feature extraction vector.

4. The data flow intelligent reconstruction method for integrated modules according to claim 1 is characterized in that: The performing probability density calculation and boundary distance calculation on the decoupled feature vector to obtain a data flow security score includes: Performing Gaussian kernel density transformation on the decoupled feature vector to obtain a feature kernel density matrix, and inputting the feature kernel density matrix into a conditional probability calculation unit, calculating the conditional probability distribution based on the operation mode embedding vector, and obtaining a conditional probability vector; Performing a logarithmic transformation operation on the conditional probability vector to obtain a probabilistic safety component, constructing a safety boundary set according to a preset safety boundary threshold, and mapping the decoupled feature vector to a safety boundary space to obtain a feature mapping vector; Calculating a Euclidean distance matrix for the feature mapping vector and the safety boundary set to obtain a boundary distance vector, and performing a minimum value search and a sign inversion operation on the boundary distance vector to obtain a boundary safety component; A weighted combination function is constructed based on the probabilistic security component and the boundary security component to obtain a combination weight coefficient, and a weighted sum operation is performed on the probabilistic security component and the boundary security component according to the combination weight coefficient to obtain a data flow security score.

5. The data flow intelligent reconstruction method for integrated modules according to claim 4 is characterized in that: The multi-layer Bayesian perception reconstruction based on the data flow security score and the decoupled feature vector to obtain the target reconstructed data flow includes: Performing a threshold judgment on the data flow security score to obtain a security mask vector, and performing feature filtering on the decoupling feature vector and the security mask vector to obtain a security feature vector; Constructing a Bayesian neural network including a priori layer and a likelihood layer for the security feature vector, sampling the weight distribution through the priori layer to obtain a priori weight matrix, and modeling the feature distribution through the likelihood layer to obtain a likelihood probability vector; Inputting the prior weight matrix and the likelihood probability vector into a variational inference module for sampling to obtain posterior distribution parameters, and performing maximum posterior probability calculation on the posterior distribution parameters to obtain optimal reconstruction parameters; Inputting the security feature vector and the optimal reconstruction parameter into a reconstruction decoder, performing feature mapping through a multi-layer perceptron, and obtaining a reconstructed feature vector; The reconstructed feature vector is subjected to denormalization and dimension transformation processing to obtain an initial reconstructed data stream, and a weighted combination operation is performed based on the initial reconstructed data stream and the data stream security score to obtain a target reconstructed data stream.

6. A data stream intelligent reconstruction device for integrated modules, characterized in that: The device is used to execute the data flow intelligent reconstruction method for integrated modules according to any one of claims 1 to 5, comprising: A feature decomposition module, used for performing dual-stream feature decomposition on the chip integrated module data stream to obtain a first module stream feature extraction vector and a first data stream feature extraction vector; A feature optimization module, used for inputting the first module flow feature extraction vector and the first data flow feature extraction vector into an adversarial network for feature optimization to obtain a second module flow feature extraction vector and a second data flow feature extraction vector; A construction module is used to construct a directed causal graph based on the chip system operating parameters, and perform graph neural network encoding on the directed causal graph to obtain an operating mode embedding vector; a decoupling module, configured to input the operation mode embedding vector, the second module flow feature extraction vector and the second data flow feature extraction vector into a conditional variational autoencoder network for causal decoupling to obtain a decoupled feature vector; A calculation module, used to perform probability density calculation and boundary distance calculation on the decoupled feature vector to obtain a data flow security score; A reconstruction module is used to perform multi-layer Bayesian perception reconstruction based on the data flow security score and the decoupled feature vector to obtain a target reconstructed data flow.

7. A data stream intelligent reconstruction device for integrated modules, characterized in that: The data stream intelligent reconstruction device for integrated modules comprises: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instruction in the memory so that the data flow intelligent reconstruction device for the integration module executes the data flow intelligent reconstruction method for the integration module according to any one of claims 1-5.

8. A smart chip, characterized in that: The intelligent chip is used to implement the data flow intelligent reconstruction method for integrated modules as described in any one of claims 1-5.

Images