An integrated circuit characteristic detection system based on deep learning

Abstract

The present invention relates to the technical field of integrated circuit characteristic detection, and provides an integrated circuit characteristic detection system based on deep learning; the system obtains circuit characteristic data through a data acquisition module, and removes noise and irrelevant features in the preprocessing stage; by using the dynamic convolution feature extraction technology based on GAN, the system can adaptively generate convolution kernels and accurately extract key circuit features; the LSTM model is combined to analyze the time series to capture the dynamic changes in the circuit; finally, the HFCN model and MI-MDRE fusion technology are used to realize the deep fusion and decision optimization of multiple features; through the above steps, the system significantly improves the accuracy, real-time and abnormal recognition ability of integrated circuit characteristic detection, and provides an efficient and reliable technical solution for the integrated circuit detection field.

Description

Technical Field

[0001] The present invention relates to the technical field of integrated circuit characteristic detection, and in particular to an integrated circuit characteristic detection system based on deep learning. Background Art

[0002] With the rapid development of integrated circuit design, the complexity and integration of integrated circuits are constantly improving. In practical applications, the characteristic detection of integrated circuits is crucial to ensure their performance and reliability. However, existing intelligent detection models often have certain limitations when facing complex circuit structures and changing working environments. For example, many models have low sensitivity to subtle anomalies or specific working conditions in the circuit, and are prone to ignoring key features, resulting in insufficient detection accuracy. In addition, existing intelligent models usually focus on specific types of circuit detection and lack the ability to conduct a comprehensive analysis of the overall characteristics of integrated circuits, which makes them show performance deficiencies when dealing with diverse application scenarios. In response to these problems, there is an urgent need for an integrated circuit characteristic detection system that can integrate multiple deep learning technologies. The system should have higher detection performance, not only able to accurately capture circuit characteristics, but also maintain a high level of detection accuracy and real-time performance in complex and changing environments. Compared with existing intelligent models, this system has significant advantages in the comprehensive analysis of global and local characteristics of circuits, which improves the comprehensiveness, accuracy and performance of integrated circuit characteristic analysis. Summary of the invention

[0003] The present invention provides an integrated circuit characteristic detection system based on deep learning, aiming to solve the problems of low detection accuracy, slow response speed and weak abnormality recognition ability in the prior art; the present invention extracts feature data through a dynamic convolution feature extraction module and a time series analysis module, and fuses, reduces the dimension and nonlinearly transforms the feature data through a hybrid fusion neural network through MI-MDRE fusion technology to generate an integrated circuit characteristic decision value; the accuracy and real-time performance of integrated circuit characteristic detection are significantly improved, and the ability to recognize abnormal situations is enhanced.

[0004] The present invention provides an integrated circuit characteristic detection system based on deep learning, which includes a data acquisition module, a dynamic convolution feature extraction module, a time series analysis module, an integrated feature decision module, an anomaly detection and alarm module and a data storage module;

[0005] The data acquisition module collects circuit characteristic data and preprocesses the circuit characteristic data, including removing noise and irrelevant features to obtain preprocessed data;

[0006] The dynamic convolution feature extraction module processes the preprocessed data through the dynamic convolution feature extraction technology enhanced by GAN (Generative Adversarial Network) to generate a dynamic feature vector set;

[0007] The time series analysis module: uses LSTM (Long Short-Term Memory Network) to perform time series analysis on the dynamic feature vector set, captures the timing characteristics in the integrated circuit, and generates a timing feature vector set;

[0008] The integrated feature decision module: constructs an HFCN (hybrid fully connected network) model, inputs a dynamic feature vector set and a time series feature vector set into the HFCN model, performs feature fusion, dimensionality reduction and nonlinear transformation through MI-MDRE (mutual information-multidimensional reconstruction error) fusion technology, and generates an integrated circuit characteristic decision value;

[0009] The anomaly detection and alarm module: monitors the operating status of the integrated circuit in real time according to the integrated circuit characteristic decision value, and when an anomaly is detected, issues an alarm to the user through a sound alarm, a visual alarm and a text message notification method;

[0010] The data storage module stores circuit characteristic data, dynamic feature vector sets, time series feature vector sets, and integrated circuit characteristic decision values, and feeds back to the dynamic convolution feature extraction module, the time series analysis module, and the anomaly detection and alarm module;

[0011] MI-MDRE fusion technology is constructed by combining mutual information and multi-dimensional reconstruction error;

[0012] The HFCN model is constructed by combining multi-dimensional feature fusion technology, dimensionality reduction technology and MI-MDRE fusion technology.

[0013] Furthermore, the dynamic convolution feature extraction module processes the preprocessed data by using the GAN-enhanced dynamic convolution feature extraction technology, specifically including the following steps:

[0014] Step S1: Feature map generation: perform preliminary convolution operation on the preprocessed data, extract the basic feature map, calculate the global statistical information of the basic feature map to capture macro features; and calculate the local statistical information to capture subtle regional features;

[0015] Step S2: Dynamic convolution kernel generation: Generate adaptive convolution kernels through GAN using global and local statistical information;

[0016] Step S3: dynamic convolution operation: use the adaptive convolution kernel to perform a convolution operation on the basic feature map to obtain a convolution feature map;

[0017] Step S4: Feature map integration: Integrate the convolution feature maps, including merging and smoothing the convolution feature maps, to generate a dynamic feature vector set.

[0018] Furthermore, step S2 specifically includes the following steps:

[0019] Step S21: constructing a GAN model, initializing the generator and the discriminator, including initializing the generator weights and the discriminator weights, the generator receiving random noise, global statistical information and local statistical information as input, generating an initial dynamic convolution kernel, and the discriminator evaluating the output quality of the initial dynamic convolution kernel of the generator;

[0020] Step S22: Define the total loss function of the generator, optimize the output quality of the initial dynamic convolution kernel, the generator calculates the loss value according to the total loss function, adjusts the weight of the generator, and generates an optimized dynamic convolution kernel. The formula is as follows:

[0021] α ， β ， γ= soft max(MLP([ L g lobal , L local , L adv ])) ;

[0022] in, represents a multi-layer perceptron neural network, Represents the activation function for multi-classification problems, represents the global statistical information loss, represents the local statistical information loss, Represents resistance to loss, represents the global statistical information loss weight parameter, represents the local statistical information loss weight parameter, represents the adversarial loss weight parameter;

[0023] ;

[0024] in, represents the total loss function;

[0025] Step S23: The discriminator evaluates the optimized dynamic convolution kernel again, outputs the evaluation result, calculates the discriminator loss value based on the evaluation result, and adjusts the discriminator weight according to the discriminator loss value to obtain the discriminator feedback. The generator further adjusts the generator weight according to the discriminator feedback, further adjusts the optimized dynamic convolution kernel, and outputs the adaptive convolution kernel.

[0026] Furthermore, the time series analysis module uses LSTM to perform time series analysis on the dynamic feature vector set, which specifically includes the following steps:

[0027] Step C1: sequence construction and standardization: organize the dynamic feature vector set into a time series data in chronological order, and perform standardization processing on the time series data, including normalization, denoising and data smoothing, to obtain standardized time series data;

[0028] Step C2: Time series feature learning: Input the standardized time series data into LSTM, capture the time series features existing in the circuit, extract the key patterns and change trends in the standardized time series data, and generate a set of time series feature vectors.

[0029] Furthermore, the integrated feature decision module inputs the dynamic feature vector set and the time series feature vector set into the HFCN model, and performs feature fusion, dimensionality reduction and nonlinear transformation through the MI-MDRE fusion technology, which specifically includes the following steps:

[0030] Step B1: Constructing the HFCN model, which includes an input layer, a feature fusion layer, a fully connected layer, an activation function layer, a Dropout layer, and an output layer; each layer of the HFCN model is composed of neurons, and the HFCN model comprehensively analyzes the dynamic feature vector set and the time series feature vector set to generate the integrated circuit characteristic decision value;

[0031] Step B2: Divide the dynamic feature vector set and the time series feature vector set into a training set and a validation set, initialize the HFCN model weights, train the HFCN model through the training set, update the HFCN model weights through the back propagation algorithm, and evaluate the HFCN model performance through the validation set.

[0032] Furthermore, the feature fusion layer of step B1 fuses the dynamic feature vector set and the temporal feature vector set through MI-MDRE fusion technology to form a fusion feature.

[0033] Furthermore, the process of fusing the dynamic feature vector set and the temporal feature vector set by using the MI-MDRE fusion technology specifically includes the following steps:

[0034] Step M1: Calculate the mutual information between the dynamic feature vector set and the time series feature vector set. The mutual information formula is as follows:

[0035] ;

[0036] in, represents the task dimension, Represents the first The dynamic characteristics of the task dimension, Represents a set of time series feature vectors; express and The mutual information of Represents a collection symbol. express A specific feature of express A specific feature of express and The joint distribution of and express and The marginal distribution of

[0037] ;

[0038] in, Represents the first The temporal characteristics of the task dimension, represents a set of dynamic feature vectors, express and The mutual information of express A specific feature of express A specific feature of express and The joint distribution of and express and The marginal distribution of

[0039] Step M2: According to the mutual information, the multi-dimensional reconstruction error is introduced to calculate the dimension weights of the dynamic feature vector set and the dimension weights of the time series feature vector set. The formula is as follows:

[0040] ;

[0041] in, Indicates the total number of task dimensions, Represents the dynamic feature vector set The weight of each task dimension, represents the coefficient controlling the reconstruction error weight, express The ideal dynamic feature vector of task dimensions, represents the reconstruction error;

[0042] ;

[0043] in, Represents the time series feature vector set The weight of each task dimension, express The ideal temporal feature vector of each task dimension;

[0044] Step M3: Calculate the total weight according to the dynamic feature vector set dimension weight and the time series feature vector set dimension weight, and perform weighted fusion to obtain the fusion feature. The formula is as follows:

[0045] Compute the total weight of a set of dynamic eigenvectors:

[0046] ;

[0047] in, represents the total weight of the dynamic feature vector set;

[0048] Calculate the total weight of a set of time series feature vectors:

[0049] ;

[0050] in, Represents the total weight of the time series feature vector set;

[0051] Weighted fusion formula:

[0052] ;

[0053] in, Represents fusion features.

[0054] Furthermore, the neurons in the fully connected layer of step B1 receive the fused features and perform weighted summation on the fused features to form a high-dimensional feature representation vector.

[0055] Furthermore, the activation function layer and the Dropout layer of step B1 perform nonlinear transformation on the high-dimensional feature representation vector and randomly discard neurons to obtain a nonlinear feature vector.

[0056] Furthermore, the output layer of step B1 uses a Sigmoid activation function to convert the nonlinear feature vector into an integrated circuit characteristic decision value.

[0057] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0058] The present invention provides an integrated circuit characteristic detection system based on deep learning. The system realizes efficient response and flexible adaptation to changes in circuit characteristics through a dynamic convolution feature extraction module. Unlike the existing static convolution kernel, the GAN-enhanced dynamic convolution feature extraction technology can find the best balance between global and local statistical information, so that the convolution operation can more accurately extract the key features in the circuit. The accuracy and flexibility of circuit detection are improved, especially when facing complex and diverse circuit characteristics, the system's adaptability and detection accuracy are significantly improved.

[0059] Through the time series analysis module, the present invention solves the problem of ignoring the time dependence of circuit characteristics in traditional methods; the introduction of the LSTM model enables the system to effectively capture and analyze the dynamic changes in circuit characteristics, which is particularly suitable for processing complex time series data; when analyzing the circuit operation status, this module can more accurately detect potential anomalies and faults, thereby enhancing the overall reliability of the system and improving the accuracy and timeliness of detection.

[0060] The integrated feature decision module uses the HFCN model and MI-MDRE fusion technology to achieve deep fusion of multiple features and decision optimization; by dynamically adjusting the fusion weights, the present invention can organically combine dynamic features and timing features to generate highly discriminative integrated circuit characteristic decision values; this method effectively solves the problems of information loss and insufficient decision accuracy in traditional feature fusion methods, significantly improves the system's discrimination ability and decision accuracy, and provides strong technical support for the efficiency and reliability of integrated circuit detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a module of an integrated circuit characteristic detection system based on deep learning proposed by the present invention;

[0062] Figure 2 This is a module schematic diagram of the HFCN model proposed in the present invention. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0064] Embodiment 1, according to Figure 1, the present invention provides an integrated circuit characteristic detection system based on deep learning, the system includes a data acquisition module, a dynamic convolution feature extraction module, a time series analysis module, an integrated feature decision module, an anomaly detection and alarm module and a data storage module;

[0065] The data acquisition module collects circuit characteristic data and preprocesses the circuit characteristic data, including removing noise and irrelevant features to obtain preprocessed data;

[0066] The dynamic convolution feature extraction module processes the preprocessed data through the GAN-enhanced dynamic convolution feature extraction technology to generate a dynamic feature vector set;

[0067] The time series analysis module: uses LSTM to perform time series analysis on the dynamic feature vector set, captures the timing characteristics in the integrated circuit, and generates a time series feature vector set;

[0068] The integrated feature decision module: constructs an HFCN model, inputs a dynamic feature vector set and a time series feature vector set into the HFCN model, performs feature fusion, dimensionality reduction and nonlinear transformation through MI-MDRE fusion technology, and generates an integrated circuit characteristic decision value;

[0069] The anomaly detection and alarm module: monitors the operating status of the integrated circuit in real time according to the integrated circuit characteristic decision value, and when an anomaly is detected, issues an alarm to the user through a sound alarm, a visual alarm and a text message notification method;

[0070] The data storage module stores circuit characteristic data, dynamic feature vector sets, time series feature vector sets and integrated circuit characteristic decision values, and feeds back to the dynamic convolution feature extraction module, the time series analysis module and the anomaly detection and alarm module.

[0071] Embodiment 2, based on the above embodiment, collects circuit characteristic data, including the following aspects:

[0072] Voltage characteristics: including voltage values ​​of integrated circuit nodes, transient voltage changes, and stable voltage levels;

[0073] Current characteristics: including the current size of circuit components, transient current changes, leakage current, and power supply current stability;

[0074] Frequency response: including operating frequency, frequency stability, and frequency deviation of oscillators and filter components;

[0075] Temperature characteristics: Monitor the temperature distribution and temperature change trend of the integrated circuit through thermistor or thermocouple sensors;

[0076] Timing characteristics: including the signal's rise time, fall time, delay time and pulse width;

[0077] Noise characteristics: measure power supply noise, signal noise and external interference noise in the circuit.

[0078] In the third embodiment, the dynamic convolution feature extraction module processes the preprocessed data by using the GAN-enhanced dynamic convolution feature extraction technology, and specifically includes the following steps:

[0079] Step S1: Feature map generation: perform preliminary convolution operation on the preprocessed data, extract the basic feature map, calculate the global statistical information of the basic feature map to capture macro features; and calculate the local statistical information to capture subtle regional features;

[0080] Global statistics capture macroscopic features including: global mean, global variance, global maximum and minimum, global contrast, and global energy;

[0081] Local statistics capture subtle regional features including: local mean, local variance, local maximum and minimum, local gradient, and local contrast.

[0082] Step S2: Dynamic convolution kernel generation: Generate adaptive convolution kernels through GAN using global and local statistical information;

[0083] Step S3: dynamic convolution operation: use the adaptive convolution kernel to perform a convolution operation on the basic feature map to obtain a convolution feature map;

[0084] The convolution feature map includes: spatial distribution features, frequency domain features, texture features, edge features, temporal features and signal correlation features;

[0085] Spatial distribution characteristics include:

[0086] Signal strength distribution: indicates the spatial distribution of signal strength in different areas, reflecting the difference in signal strength in various parts of the circuit;

[0087] Circuit layout features: represent the geometric structure and layout of the circuit, including the relative positions and connection patterns between components.

[0088] Frequency domain features include:

[0089] Spectrum distribution: represents the frequency components and distribution of the signal in the circuit, capturing the changes of different frequency components in space;

[0090] Harmonic features: Extract the harmonic components in the circuit signal.

[0091] Texture features include:

[0092] Surface texture information: The microscopic structural features of the circuit surface are captured through convolution operations to obtain the surface roughness and directionality;

[0093] Signal texture features: Analyze the local change pattern of the signal and capture the fluctuation, noise and interference characteristics of the signal.

[0094] Edge features include:

[0095] Contour and edge information: Identify edge and contour features in circuit images, reflecting the boundaries of components and signal paths;

[0096] Edge strength and direction: measure the strength and direction of edges, and analyze the distribution and direction of edges in the circuit.

[0097] Timing characteristics include:

[0098] Timing change information: captures the changing characteristics of the signal over time, reflecting the dynamic behavior and response time of the circuit;

[0099] Timing correlation features: Analyze the timing correlation between different signals and reveal the time dependence between signals.

[0100] Signal correlation characteristics:

[0101] Correlation between signals: Analyze the spatial correlation of signals to capture the synchronization and mutual influence of signals;

[0102] Cross-signal features: Extract cross-signal characteristics to reflect signal coupling and interference in the circuit.

[0103] Step S4: Feature map integration: Integrate the convolution feature maps, including merging and smoothing the convolution feature maps, to generate a dynamic feature vector set.

[0104] Embodiment 4: This embodiment is based on the above embodiment, and step S2 specifically includes the following steps:

[0105] Step S21: constructing a GAN model, initializing the generator and the discriminator, including initializing the generator weights and the discriminator weights, the generator receiving random noise, global statistical information and local statistical information as input, generating an initial dynamic convolution kernel, and the discriminator evaluating the output quality of the initial dynamic convolution kernel of the generator;

[0106] Step S22: Define the total loss function of the generator, optimize the output quality of the initial dynamic convolution kernel, the generator calculates the loss value according to the total loss function, adjusts the weight of the generator, and generates an optimized dynamic convolution kernel. The formula is as follows:

[0107] α ， β ， γ= soft max(MLP([ L g lobal , L local , L adv ])) ;

[0108] in, represents a multi-layer perceptron neural network, Represents the activation function for multi-classification problems, represents the global statistical information loss, represents the local statistical information loss, Represents resistance to loss, represents the global statistical information loss weight parameter, represents the local statistical information loss weight parameter, represents the adversarial loss weight parameter;

[0109] ;

[0110] in, represents the total loss function;

[0111] Step S23: The discriminator evaluates the optimized dynamic convolution kernel again, outputs the evaluation result, calculates the discriminator loss value based on the evaluation result, and adjusts the discriminator weight according to the discriminator loss value to obtain the discriminator feedback. The generator further adjusts the generator weight according to the discriminator feedback, further adjusts the optimized dynamic convolution kernel, and outputs the adaptive convolution kernel.

[0112] Embodiment 5, based on the above embodiment, the time series analysis module uses LSTM to perform time series analysis on a dynamic feature vector set, specifically comprising the following steps:

[0113] Step C1: sequence construction and standardization: organize the dynamic feature vector set into a time series data in chronological order, and perform standardization processing on the time series data, including normalization, denoising and data smoothing, to obtain standardized time series data;

[0114] Step C2: Time series feature learning: Input the standardized time series data into LSTM, capture the time series features existing in the circuit, extract the key patterns and change trends in the standardized time series data, and generate a set of time series feature vectors.

[0115] Embodiment 6, according to Figure 2 The integrated feature decision module inputs the dynamic feature vector set and the time series feature vector set into the HFCN model, and performs feature fusion, dimensionality reduction and nonlinear transformation through the MI-MDRE fusion technology, which specifically includes the following steps:

[0116] Step B1: Constructing the HFCN model, which includes an input layer, a feature fusion layer, a fully connected layer, an activation function layer, a Dropout layer, and an output layer; each layer of the HFCN model is composed of neurons, and the HFCN model comprehensively analyzes the dynamic feature vector set and the time series feature vector set to generate the integrated circuit characteristic decision value;

[0117] Step B2: Divide the dynamic feature vector set and the time series feature vector set into a training set and a validation set, initialize the HFCN model weights, train the HFCN model through the training set, update the HFCN model weights through the back propagation algorithm, and evaluate the HFCN model performance through the validation set.

[0118] Embodiment 7: This embodiment is based on the above embodiment. The feature fusion layer of step B1 fuses the dynamic feature vector set and the temporal feature vector set through MI-MDRE fusion technology to form a fusion feature.

[0119] Embodiment 8, based on the above embodiment, the process of fusing the dynamic feature vector set and the temporal feature vector set by using the MI-MDRE fusion technology specifically includes the following steps:

[0120] Step M1: Calculate the mutual information between the dynamic feature vector set and the time series feature vector set. The mutual information formula is as follows:

[0121] ;

[0122] in, represents the task dimension, Represents the first The dynamic characteristics of the task dimension, Represents a set of time series feature vectors; express and The mutual information of Represents a collection symbol. express A specific feature of express A specific feature of express and The joint distribution of and express and The marginal distribution of

[0123] ;

[0124] in, Represents the first The temporal characteristics of the task dimension, represents a set of dynamic feature vectors, express and The mutual information of express A specific feature of express A specific feature of express and The joint distribution of and express and The marginal distribution of

[0125] Step M2: According to the mutual information, the multi-dimensional reconstruction error is introduced to calculate the dimension weights of the dynamic feature vector set and the dimension weights of the time series feature vector set. The formula is as follows:

[0126] ;

[0127] in, Indicates the total number of task dimensions, Represents the dynamic feature vector set The weight of each task dimension, represents the coefficient controlling the reconstruction error weight, express The ideal dynamic feature vector of task dimensions, represents the reconstruction error;

[0128] ;

[0129] in, Represents the time series feature vector set The weight of each task dimension, express The ideal temporal feature vector of each task dimension;

[0130] Step M3: Calculate the total weight according to the dynamic feature vector set dimension weight and the time series feature vector set dimension weight, and perform weighted fusion to obtain the fusion feature. The formula is as follows:

[0131] Compute the total weight of a set of dynamic eigenvectors:

[0132] ;

[0133] in, represents the total weight of the dynamic feature vector set;

[0134] Calculate the total weight of a set of time series feature vectors:

[0135] ;

[0136] in, Represents the total weight of the time series feature vector set;

[0137] Weighted fusion formula:

[0138] ;

[0139] in, Represents fusion features.

[0140] Embodiment 9: This embodiment is based on the above embodiment, and the neurons in the fully connected layer of step B1 receive fused features and perform weighted summation on the fused features to form a high-dimensional feature representation vector.

[0141] Embodiment 10: This embodiment is based on the above embodiment. The activation function layer and the Dropout layer in step B1 perform nonlinear transformation on the high-dimensional feature representation vector and randomly discard neurons to obtain a nonlinear feature vector.

[0142] Embodiment 11: This embodiment is based on the above embodiment, and the output layer of step B1 uses a Sigmoid activation function to convert the nonlinear feature vector into an integrated circuit characteristic decision value.

[0143] Embodiment 12, based on the above embodiment, the abnormality detection and alarm module monitors the operation status of the integrated circuit in real time according to the integrated circuit characteristic decision value, and when an abnormality is detected, issues an alarm to the user through a sound alarm, a visual alarm and a text message notification method;

[0144] In this embodiment, a high-frequency oscillator integrated circuit is used to generate a stable clock signal, and key performance parameters include operating frequency, temperature stability, power supply voltage and signal noise level;

[0145] Get the integrated circuit characteristic decision value:

[0146] Frequency stability decision value: 0.85 (normal range is 0.9-1.1, below the lower limit);

[0147] Power supply voltage stability decision value: 1.0 (normal range);

[0148] Temperature stability decision value: 0.95 (normal range);

[0149] Noise level decision value: 1.0 (normal range);

[0150] The stability decision value of the operating frequency is lower than the lower limit of the normal range. The frequency of the oscillator is unstable and there is a potential risk of failure. The alarm module is activated:

[0151] Sound alarm: emits a continuous beep to alert the equipment manager;

[0152] Visual alarm: Flashing red on the monitoring screen to warn that there is an abnormality in the device;

[0153] SMS notification: Send SMS notification to relevant technical personnel, the content of the SMS is: "Warning: The oscillator frequency is unstable and below the normal range, and needs to be checked immediately."

Claims

1. An integrated circuit characteristic detection system based on deep learning, characterized in that: It includes data acquisition module, dynamic convolution feature extraction module, time series analysis module, integrated feature decision module, anomaly detection and alarm module and data storage module; The data acquisition module collects circuit characteristic data and preprocesses the circuit characteristic data to obtain preprocessed data; The dynamic convolution feature extraction module processes the preprocessed data through the GAN-enhanced dynamic convolution feature extraction technology to generate a dynamic feature vector set; The time series analysis module: uses LSTM to perform time series analysis on the dynamic feature vector set, captures the timing characteristics in the integrated circuit, and generates a time series feature vector set; The integrated feature decision module: establishes an HFCN model, inputs a dynamic feature vector set and a time series feature vector set into the HFCN model, performs feature fusion, dimensionality reduction and nonlinear transformation through MI-MDRE fusion technology, and generates an integrated circuit characteristic decision value; The integrated feature decision module inputs the dynamic feature vector set and the time series feature vector set into the HFCN model for feature fusion, dimensionality reduction and nonlinear transformation, which specifically includes the following steps: Step B1: Constructing the HFCN model, which includes an input layer, a feature fusion layer, a fully connected layer, an activation function layer, a Dropout layer, and an output layer; each layer of the HFCN model is composed of neurons, and the HFCN model comprehensively analyzes the dynamic feature vector set and the time series feature vector set to generate the integrated circuit characteristic decision value; Step B2: Divide the dynamic feature vector set and the time series feature vector set into a training set and a validation set, initialize the HFCN model weight, train the HFCN model through the training set, update the HFCN model weight through the back propagation algorithm, and evaluate the HFCN model performance through the validation set; The feature fusion layer of step B1 fuses the dynamic feature vector set and the temporal feature vector set through MI-MDRE fusion technology to form a fusion feature; The process of fusing the dynamic feature vector set and the temporal feature vector set by using the MI-MDRE fusion technology specifically includes the following steps: Step M1: Calculate the mutual information between the dynamic feature vector set and the time series feature vector set; Step M2: According to the mutual information, a multi-dimensional reconstruction error is introduced to calculate the dimension weights of the dynamic feature vector set and the dimension weights of the temporal feature vector set; Step M3: Calculate the total weight according to the dynamic feature vector set dimension weight and the temporal feature vector set dimension weight, and perform weighted fusion to obtain the fused feature.

2. The integrated circuit characteristic detection system based on deep learning according to claim 1, characterized in that: The dynamic convolution feature extraction module processes the preprocessed data through the GAN-enhanced dynamic convolution feature extraction technology, specifically including the following steps: Step S1: Feature map generation: perform preliminary convolution operation on the preprocessed data, extract the basic feature map, calculate the global statistical information of the basic feature map to capture macro features; and calculate the local statistical information to capture subtle regional features; Step S2: Dynamic convolution kernel generation: Generate adaptive convolution kernels through GAN using global and local statistical information; Step S3: dynamic convolution operation: use the adaptive convolution kernel to perform a convolution operation on the basic feature map to obtain a convolution feature map; Step S4: Feature map integration: Integrate the convolution feature maps, including merging and smoothing the convolution feature maps, to generate a dynamic feature vector set.

3. The integrated circuit characteristic detection system based on deep learning according to claim 2, characterized in that: The step S2 specifically includes the following steps: S21: Build a GAN model, initialize the generator and discriminator, including initializing the generator weights and the discriminator weights. The generator receives random noise as input and generates an initial dynamic convolution kernel. The discriminator evaluates the output quality of the generator's initial dynamic convolution kernel. S22: Define the total loss function of the generator, optimize the output quality of the initial dynamic convolution kernel, the generator calculates the loss value according to the total loss function, adjusts the weight of the generator, and generates an optimized dynamic convolution kernel; S23: The discriminator evaluates the optimized dynamic convolution kernel again, outputs the evaluation result, calculates the discriminator loss value based on the evaluation result, and adjusts the discriminator weight according to the discriminator loss value to obtain the discriminator feedback. The generator further adjusts the generator weight according to the discriminator feedback, further adjusts the optimized dynamic convolution kernel, and outputs the adaptive convolution kernel.

4. The integrated circuit characteristic detection system based on deep learning according to claim 1, characterized in that: The time series analysis module uses LSTM to perform time series analysis on a dynamic feature vector set, specifically including the following steps: Step C1: sequence construction and standardization: organize the dynamic feature vector set into a time series data in chronological order, and perform standardization processing on the time series data, including normalization, denoising and data smoothing, to obtain standardized time series data; Step C2: Time series feature learning: Input the standardized time series data into LSTM, capture the time series features existing in the circuit, extract the key patterns and change trends in the standardized time series data, and generate a set of time series feature vectors.

5. The integrated circuit characteristic detection system based on deep learning according to claim 1, characterized in that: The neurons in the fully connected layer of step B1 receive the fused features and perform weighted summation on the fused features to form a high-dimensional feature representation vector.

6. The integrated circuit characteristic detection system based on deep learning according to claim 5, characterized in that: The activation function layer and the Dropout layer of step B1 perform nonlinear transformation on the high-dimensional feature representation vector and randomly discard neurons to obtain a nonlinear feature vector.

7. The integrated circuit characteristic detection system based on deep learning according to claim 6, characterized in that: The output layer of step B1 uses a Sigmoid activation function to convert the nonlinear feature vector into an integrated circuit characteristic decision value.

Images