Integrated circuit package sealing evaluation method and system under high temperature environment

Abstract

The present invention provides a method and system for evaluating the sealing performance of integrated circuit packaging under a high-temperature environment, and relates to the field of circuit testing technology, including: obtaining temperature, tracer gas concentration and air pressure data of the integrated circuit, extracting temperature nonlinear characteristics, using non-local weighted adaptive anisotropic diffusion filtering and mixed tensor decomposition to obtain multidimensional distribution characteristics of concentration, extracting air pressure frequency domain characteristics, performing feature fusion and spatiotemporal dependency analysis through a quantum deep autoencoder, a multi-head cross-attention network and a heterogeneous spatiotemporal graph neural network, extracting timing features through waveform dynamic time warping and a bidirectional gated recurrent attention network, performing sealing defect diagnosis based on knowledge distillation, predicting failure trends using a causal long short-term memory network, matching failure modes in combination with a dynamic heterogeneous knowledge graph, and using a multi-objective Pareto differential evolution algorithm to evaluate packaging reliability and generate an evaluation report.

Description

Technical Field

[0001] The present invention relates to the technical field of circuit testing, and in particular to a method and system for evaluating the sealing performance of integrated circuit packaging in a high-temperature environment. Background Art

[0002] The sealing performance of integrated circuit packaging directly affects the reliability and service life of the device. The performance degradation of packaging materials and structures will lead to sealing failure and device damage. Therefore, the evaluation of the sealing performance of integrated circuit packaging is crucial to ensure the reliable operation of the device.

[0003] However, traditional methods have shortcomings in detection efficiency, sensitivity, stability and reliability, and are difficult to meet the current needs of integrated circuit reliability assessment. Especially in high-temperature environments, the failure mechanism of devices is more complicated, which puts higher requirements on the evaluation of packaging sealing.

[0004] Therefore, a method is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for evaluating the sealing performance of integrated circuit packaging in a high-temperature environment, which can at least solve some of the problems existing in the prior art.

[0006] According to a first aspect of an embodiment of the present invention, a method for evaluating the sealing performance of an integrated circuit package in a high temperature environment is provided, comprising:

[0007] Acquire temperature data, tracer gas concentration data and air pressure data of the integrated circuit to be tested, decompose the temperature data by using a bidirectional empirical mode decomposition algorithm to obtain a bidirectional temperature intrinsic mode function, extract nonlinear temperature instantaneous eigenvalues ​​by combining fractional-order Hilbert transform, perform noise reduction on the tracer gas concentration data by using a non-local weighted adaptive anisotropic diffusion filter, construct a multidimensional concentration distribution feature matrix by using a mixed tensor decomposition method, obtain air pressure frequency domain eigenvalues ​​by using an adaptive morphological wavelet packet decomposition algorithm for the air pressure data, and input the nonlinear temperature instantaneous eigenvalues, the multidimensional concentration distribution feature matrix and the air pressure frequency domain eigenvalues ​​into a quantum deep autoencoder network to generate a quantum eigenvector;

[0008] Input the quantum feature vector into a multi-head cross attention network to generate an adaptive weight feature vector, construct a heterogeneous spatiotemporal graph neural network, extract the spatiotemporal dependency between feature nodes to obtain a spatiotemporal graph feature vector, perform multi-scale temporal feature alignment on the spatiotemporal graph feature vector through a waveform dynamic time warping algorithm, input the aligned feature sequence into a bidirectional gated recurrent attention network to extract long-term and short-term temporal dependencies to obtain a temporal dynamic feature sequence, and construct a hierarchical probabilistic graph model in combination with a deep variational inference method to obtain a probabilistic feature vector;

[0009] The probability feature vector is input into the sealing defect diagnosis model based on knowledge distillation, and the sealing defect type and the corresponding confidence are output through the collaborative training of the teacher network and the student network. The time series dynamic feature sequence is input into the pre-constructed causal long short-term memory network to predict the failure trend, and a dynamic heterogeneous knowledge graph is constructed to match the failure mode to obtain the failure mode evolution law. Based on the failure trend prediction result and the failure mode evolution law, the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life are calculated through the multi-objective Pareto differential evolution algorithm, and the packaging reliability of the integrated circuit to be tested is rated, the reliability rating result is obtained and a sealing assessment report is output.

[0010] In an optional embodiment,

[0011] The temperature data, tracer gas concentration data and air pressure data of the integrated circuit to be tested are obtained, the temperature data are decomposed by a bidirectional empirical mode decomposition algorithm to obtain a bidirectional temperature intrinsic mode function, the nonlinear temperature instantaneous eigenvalue is extracted in combination with a fractional-order Hilbert transform, the tracer gas concentration data is denoised by a non-local weighted adaptive anisotropic diffusion filter, a multidimensional concentration distribution feature matrix is ​​constructed by a mixed tensor decomposition method, the air pressure data is decomposed by an adaptive morphological wavelet packet decomposition algorithm to obtain the air pressure frequency domain eigenvalue, and the nonlinear temperature instantaneous eigenvalue, the multidimensional concentration distribution feature matrix and the air pressure frequency domain eigenvalue are input into a quantum deep autoencoder network to generate a quantum feature vector, including:

[0012] Arranging temperature sensors at four corner positions and a center position of the integrated circuit to be tested to obtain a temperature data matrix, arranging tracer gas sensors evenly inside the integrated circuit to be tested to obtain a tracer gas concentration data matrix, and arranging air pressure sensors outside the integrated circuit to be tested to obtain an air pressure data matrix;

[0013] Decomposing the temperature data matrix by a bidirectional empirical mode decomposition algorithm, screening the intrinsic mode function that satisfies the conditions that the difference between the number of extreme value points and the number of zero-crossing points does not exceed one, the local mean is zero, at least one zero-crossing point is included between adjacent local extreme value points, and the first and last extreme value points are respectively the local maximum point and the local minimum point, forward decomposing the intrinsic mode function from high to low frequency, and reverse decomposing it from low to high frequency to obtain a bidirectional temperature intrinsic mode function, and extracting nonlinear temperature instantaneous eigenvalues ​​from the bidirectional temperature intrinsic mode function by using fractional-order Hilbert transform and wavelet transform kernel function;

[0014] The tracer gas concentration data matrix is ​​subjected to denoising processing by using a non-local weighted adaptive anisotropic diffusion filter, and the sizes of the similarity window and the search window are adaptively adjusted according to the distribution characteristics of the tracer gas concentration data matrix to obtain a denoised tracer gas concentration data matrix;

[0015] The denoised tracer gas concentration data matrix is ​​constructed as a third-order tensor, wherein the first dimension represents the sensor number, the second dimension represents the time point number, and the third dimension represents the tracer gas concentration value, and the third-order tensor is decomposed by a nuclear tensor decomposition method to obtain a multi-dimensional concentration distribution feature matrix;

[0016] Preprocessing the air pressure data matrix, performing morphological opening and closing operations through disk-shaped structural elements, applying wavelet packet decomposition with adaptive layers to the preprocessed air pressure data, and extracting energy features and singular value features using an adaptive threshold strategy to obtain air pressure frequency domain eigenvalues;

[0017] The nonlinear temperature instantaneous eigenvalue, the multidimensional concentration distribution characteristic matrix and the pressure frequency domain eigenvalue are normalized and converted into a quantum state vector, the quantum state vector is input into a quantum deep autoencoder network, quantum NOT gates, Hadamard gates and rotation gates are set between adjacent layers of the quantum deep autoencoder network, and a quantum eigenvector is obtained through parameterized quantum circuit training, the quantum eigenvector contains comprehensive characteristic information of the temperature, tracer gas concentration and air pressure of the integrated circuit to be tested.

[0018] In an optional embodiment,

[0019] The tracer gas concentration data matrix is ​​subjected to noise reduction processing using a non-local weighted adaptive anisotropic diffusion filter as shown in the following formula:

[0020] ;

[0021] in, represents the rate of change of the tracer gas concentration data matrix over time, C ( x , y , t ) is the tracer gas concentration data matrix, indicating the spatial position ( x , y ) and time t The gas concentration value on x The horizontal coordinate of the sensor location, y Indicates the vertical coordinate of the sensor location, t Indicates time, w ( x , y ) represents the adaptive weight function, ▽C ( x , y ) represents the gas concentration matrix at the spatial position ( x , y ), T represents transpose, ▽ C ( x , y , t ) represents the tracer gas concentration data matrix at time t The spatial gradient of .

[0022] In an optional embodiment,

[0023] Extracting nonlinear temperature instantaneous eigenvalues ​​from the bidirectional temperature intrinsic mode function using fractional-order Hilbert transform and wavelet transform kernel function includes:

[0024] Adding a bidirectional temperature intrinsic mode function to a preset fractional-order Hilbert transform module, performing fractional differentiation on the bidirectional temperature intrinsic mode function to obtain a fractional-order derivative signal, performing Hilbert transform on the fractional-order derivative signal to obtain a first complex-valued analytical signal, and calculating a first nonlinear temperature instantaneous eigenvalue sequence based on the modulus and phase of the first complex-valued analytical signal;

[0025] Inputting the bidirectional temperature intrinsic mode function into a wavelet transform kernel function module, constructing a wavelet transform kernel function by using Morley wavelet as a wavelet basis function, performing continuous wavelet transform on the bidirectional temperature intrinsic mode function by using the wavelet transform kernel function to obtain a wavelet coefficient matrix on a time-frequency scale plane, and extracting wavelet ridges from the wavelet coefficient matrix to obtain a second nonlinear temperature instantaneous eigenvalue sequence;

[0026] Performing maximum and minimum value normalization processing on the first nonlinear temperature instantaneous eigenvalue sequence and the second nonlinear temperature instantaneous eigenvalue sequence to obtain a first normalized eigenvalue sequence and a second normalized eigenvalue sequence;

[0027] Calculate the signal-to-noise ratio of the first normalized eigenvalue sequence and the second normalized eigenvalue sequence, determine a first weight coefficient corresponding to the first normalized eigenvalue sequence and a second weight coefficient corresponding to the second normalized eigenvalue sequence based on the signal-to-noise ratio, multiply the first normalized eigenvalue sequence by the first weight coefficient to obtain a first weighted sequence, multiply the second normalized eigenvalue sequence by the second weight coefficient to obtain a second weighted sequence, and add the first weighted sequence to the second weighted sequence to obtain a fused nonlinear temperature instantaneous eigenvalue sequence.

[0028] In an optional embodiment,

[0029] The quantum feature vector is input into a multi-head cross attention network to generate an adaptive weight feature vector, a heterogeneous spatiotemporal graph neural network is constructed, the spatiotemporal dependency between feature nodes is extracted to obtain a spatiotemporal graph feature vector, the multi-scale temporal feature alignment of the spatiotemporal graph feature vector is performed through a waveform dynamic time warping algorithm, the aligned feature sequence is input into a bidirectional gated recurrent attention network to extract long-term and short-term temporal dependencies to obtain a temporal dynamic feature sequence, and a hierarchical probabilistic graph model is constructed in combination with a deep variational inference method to obtain a probabilistic feature vector including:

[0030] Input the quantum feature vector into a pre-set multi-head cross attention network, perform linear transformation on the quantum feature vector to obtain a query matrix, a key matrix and a value matrix, perform dot product operation on the query matrix and the key matrix to obtain an attention score matrix, use a softmax function on the attention score matrix to obtain an attention weight matrix, multiply the attention weight matrix and the value matrix to obtain a weighted feature matrix, and concatenate and linearly transform the weighted feature matrix to obtain an adaptive weight feature vector matrix;

[0031] The adaptive weight feature vector matrix is ​​used to construct a space-time graph according to the time series relationship in the time dimension, the feature vector of each time point in the adaptive weight feature vector matrix is ​​used as a node of the space-time graph, a directed edge connection is established between the nodes at adjacent time points in the space-time graph, and the feature vector of each node in the space-time graph and the feature vectors of the adjacent nodes of the current node are aggregated by a graph convolutional network to obtain a space-time graph feature vector matrix;

[0032] Perform wavelet decomposition on the feature vector matrix of the space-time graph in the time dimension to obtain multiple feature subsequences of different time scales, calculate the corresponding dynamic time warping distance matrix for each feature subsequence, solve the optimal matching path based on the dynamic time warping distance matrix in combination with a dynamic programming algorithm, map the feature subsequences onto a unified time axis according to the optimal matching path to obtain time-aligned feature subsequences, perform wavelet reconstruction on the time-aligned feature subsequences to obtain time-aligned feature vector matrix of the space-time graph;

[0033] Input the time-aligned spatiotemporal graph feature vector matrix into a bidirectional gated recurrent attention network, the forward gated recurrent unit and the backward gated recurrent unit of the bidirectional gated recurrent attention network respectively encode the time-aligned spatiotemporal graph feature vector matrix to obtain a forward hidden state sequence matrix and a backward hidden state sequence matrix, perform attention weighting on the forward hidden state sequence matrix and the backward hidden state sequence matrix to obtain a context vector sequence matrix, concatenate and linearly transform the context vector sequence matrix with the hidden state of the corresponding time step in the time-aligned spatiotemporal graph feature vector matrix to obtain a temporal dynamic feature sequence matrix;

[0034] The time series dynamic feature sequence matrix is ​​input into the deep variational inference model to construct a hierarchical probabilistic graphical model including a first layer, a second layer and a third layer. The first layer is used to extract high-level semantic features, the second layer is used to extract middle-level pattern features, and the third layer is used to extract bottom-level detail features. A variational inference algorithm is used to approximate the posterior distribution of the three-layer features of the hierarchical probabilistic graphical model. The parameters of the deep variational inference model are optimized by maximizing the lower bound of evidence to obtain a probabilistic feature vector matrix containing multi-level semantic information.

[0035] In an optional embodiment,

[0036] The probability feature vector is input into the sealing defect diagnosis model based on knowledge distillation, and the sealing defect type and the corresponding confidence are output through the collaborative training of the teacher network and the student network. The time series dynamic feature sequence is input into the pre-constructed causal long short-term memory network to predict the failure trend, and a dynamic heterogeneous knowledge graph is constructed to match the failure mode to obtain the failure mode evolution law. Based on the failure trend prediction result and the failure mode evolution law, the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life are calculated through the multi-objective Pareto differential evolution algorithm, and the packaging reliability of the integrated circuit to be tested is rated, and the reliability rating result is obtained and the sealing evaluation report is output, including:

[0037] Input the probability feature vector matrix composed of the probability feature vectors into the sealing defect diagnosis model based on knowledge distillation, extract high-level semantic features in the probability feature vector matrix through a deep residual convolutional neural network structure based on the teacher network in the sealing defect diagnosis model, output a first probability distribution of the sealing defect type, extract features through the mobile network structure of the student network based on the student network in the sealing defect diagnosis model, migrate the first probability distribution to the student network through a distillation loss function, optimize the second probability distribution output by the student network in combination with the cross entropy loss function, and obtain the sealing defect type and the corresponding confidence level;

[0038] Input the temporal dynamic feature sequence matrix into the encoder of the causal long short-term memory network, extract the temporal dependency and long-range context information through multi-layer gated recurrent units, obtain the failure trend feature vector, input the failure trend feature vector into the decoder of the causal long short-term memory network, enhance the feature selection ability through Granger causality constraints and sparse regularization terms, and generate a failure trend feature sequence as a failure trend prediction result;

[0039] Based on the failure trend feature sequence, a dynamic heterogeneous knowledge graph including failure mode nodes, failure mechanism nodes, failure symptom nodes and failure impact nodes is constructed, low-dimensional vector representations of nodes and relationships are learned through a graph embedding algorithm, and the semantic information and structural information of nodes are encoded in combination with a graph neural network based on an attention mechanism, node feature vectors are generated, and cluster analysis is performed on the failure mode nodes based on the similarity of the node feature vectors to obtain the evolution law between different failure modes, and association analysis is performed on the neighbor nodes of the failure mode nodes to determine the causal relationship and propagation path between the failure mode and the failure mechanism, failure symptom, and failure impact, so as to obtain the evolution law of the failure mode;

[0040] A multi-objective optimization model is established based on the failure trend prediction results and the failure mode evolution law, and the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life are taken as optimization targets. Material parameters, structural parameters, process parameters and environmental parameters are taken as optimization variables. The multi-objective optimization model is solved by using the Pareto differential evolution algorithm, and the Pareto optimal solution set is obtained through non-dominated sorting and congestion evaluation. A compromise solution is selected from the Pareto optimal solution set as the prediction value of different optimization targets, and the prediction value is graded and evaluated. Rating thresholds are set for the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life, and weight coefficients are assigned. The weighted average score is calculated according to the grade corresponding to the prediction value and the weight coefficient, and a sealing evaluation report including prediction indicators, rating results and weight analysis is generated.

[0041] In an optional embodiment,

[0042] The first probability distribution is transferred to the student network through the distillation loss function as shown in the following formula:

[0043] ;

[0044] in, L represents the loss value, α represents the weight coefficient, N represents the number of samples, M represents the number of categories, i represents the sample index, j Represents the category index, is the first probability distribution output by the teacher network, indicating the teacher network's i The sample in j The predicted probability of the class, is the second probability distribution output by the student network, indicating the student network's i The predicted probability of a sample in the jth class.

[0045] A second aspect of an embodiment of the present invention provides a system for evaluating the sealing performance of integrated circuit packages in a high temperature environment, comprising:

[0046] The first unit is used to obtain temperature data, tracer gas concentration data and air pressure data of the integrated circuit to be tested, decompose the temperature data by using a bidirectional empirical mode decomposition algorithm to obtain a bidirectional temperature intrinsic mode function, extract nonlinear temperature instantaneous eigenvalues ​​in combination with fractional-order Hilbert transform, reduce noise of the tracer gas concentration data by using a non-local weighted adaptive anisotropic diffusion filter, construct a multidimensional concentration distribution feature matrix by using a mixed tensor decomposition method, obtain air pressure frequency domain eigenvalues ​​by using an adaptive morphological wavelet packet decomposition algorithm for the air pressure data, and input the nonlinear temperature instantaneous eigenvalues, the multidimensional concentration distribution feature matrix and the air pressure frequency domain eigenvalues ​​into a quantum deep autoencoder network to generate a quantum eigenvector;

[0047] The second unit is used to input the quantum feature vector into a multi-head cross attention network to generate an adaptive weight feature vector, construct a heterogeneous space-time graph neural network, extract the space-time dependency between feature nodes to obtain a space-time graph feature vector, perform multi-scale temporal feature alignment on the space-time graph feature vector through a waveform dynamic time warping algorithm, input the aligned feature sequence into a bidirectional gated recurrent attention network to extract long-term and short-term temporal dependencies to obtain a temporal dynamic feature sequence, and construct a hierarchical probabilistic graph model in combination with a deep variational inference method to obtain a probabilistic feature vector;

[0048] The third unit is used to input the probability feature vector into a sealing defect diagnosis model based on knowledge distillation, output the sealing defect type and the corresponding confidence through the collaborative training of the teacher network and the student network, input the time series dynamic feature sequence into a pre-constructed causal long short-term memory network to predict the failure trend, construct a dynamic heterogeneous knowledge graph to match the failure mode to obtain the failure mode evolution law, and based on the failure trend prediction result and the failure mode evolution law, calculate the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life through a multi-objective Pareto differential evolution algorithm, rate the packaging reliability of the integrated circuit to be tested, obtain the reliability rating result and output a sealing assessment report.

[0049] According to a third aspect of the embodiments of the present invention,

[0050] An electronic device is provided, comprising:

[0051] processor;

[0052] a memory for storing processor-executable instructions;

[0053] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0054] A fourth aspect of the embodiments of the present invention is:

[0055] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0056] In the present invention, data preprocessing and feature extraction are performed through algorithms such as bidirectional empirical mode decomposition, non-local weighted adaptive anisotropic diffusion filtering, and adaptive morphological wavelet packet decomposition, which effectively eliminates the interference of noise and redundant information, extracts more stable and reliable feature information, adaptively learns and fuses multi-source feature information, extracts key feature patterns and laws in package sealing evaluation, enhances the generalization ability and robustness of the evaluation model, introduces causal long short-term memory networks and dynamic heterogeneous knowledge graphs, realizes long-term prediction of sealing failure trends and dynamic evolution analysis of failure modes, and through a multi-objective Pareto differential evolution algorithm, comprehensively considers multiple influencing factors such as sealing defect diffusion rate, stress distribution characteristics, and failure probability, quantitatively calculates the reliability index and remaining life of the package, and provides an important basis for the reliability design and maintenance of the device. In summary, the present invention realizes high precision, intelligence, and adaptability of package sealing evaluation, provides strong support for the reliability design and testing of integrated circuits, and has good application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of a method for evaluating the sealing performance of integrated circuit packaging in a high temperature environment according to an embodiment of the present invention;

[0058] Figure 2 It is a schematic diagram of the structure of a system for evaluating the sealing performance of integrated circuit packaging in a high temperature environment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution 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, 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.

[0060] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Figure 1 FIG. 1 is a flow chart of a method for evaluating the sealing performance of integrated circuit packaging in a high temperature environment according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] S1. Obtain temperature data, tracer gas concentration data and air pressure data of the integrated circuit to be tested, decompose the temperature data using a bidirectional empirical mode decomposition algorithm to obtain a bidirectional temperature intrinsic mode function, extract nonlinear temperature instantaneous eigenvalues ​​in combination with fractional-order Hilbert transform, use a non-local weighted adaptive anisotropic diffusion filter to reduce noise on the tracer gas concentration data, use a mixed tensor decomposition method to construct a multidimensional concentration distribution feature matrix, use an adaptive morphological wavelet packet decomposition algorithm on the air pressure data to obtain air pressure frequency domain eigenvalues, and input the nonlinear temperature instantaneous eigenvalues, the multidimensional concentration distribution feature matrix and the air pressure frequency domain eigenvalues ​​into a quantum deep autoencoder network to generate a quantum feature vector;

[0063] The tracer gas concentration data refers to the gas concentration value recorded at different time and space positions, which is usually used to monitor the gas changes in the environment. The bidirectional empirical mode decomposition algorithm is a signal decomposition method that can decompose nonlinear and non-stationary signals into several intrinsic mode functions and reduce the modal aliasing problem through bidirectional analysis. The bidirectional temperature intrinsic mode function is an intrinsic mode function obtained by bidirectional empirical mode decomposition, which is specially used for analyzing temperature signals and extracting the intrinsic characteristics of temperature changes. The fractional-order Hilbert transform is a generalization of the traditional Hilbert transform. By introducing fractional orders, it can be more The instantaneous frequency and phase characteristics of the signal are carefully analyzed. The non-local weighted adaptive anisotropic diffusion filter is an algorithm for signal denoising. It adjusts the diffusion process by introducing non-local information and adaptive gradient distribution, effectively removes noise and maintains signal edges. The hybrid tensor decomposition method is a comprehensive method combining multiple tensor decomposition technologies. By reducing calculation and storage overhead, it can more effectively extract potential structures from high-dimensional data. The adaptive morphological wavelet packet decomposition is a signal decomposition method that combines morphological operations and wavelet packet transforms. By adaptively adjusting the decomposition strategy, it can effectively extract local features in the signal.

[0064] In an optional embodiment,

[0065] The temperature data, tracer gas concentration data and air pressure data of the integrated circuit to be tested are obtained, the temperature data are decomposed by a bidirectional empirical mode decomposition algorithm to obtain a bidirectional temperature intrinsic mode function, the nonlinear temperature instantaneous eigenvalue is extracted in combination with a fractional-order Hilbert transform, the tracer gas concentration data is denoised by a non-local weighted adaptive anisotropic diffusion filter, a multidimensional concentration distribution feature matrix is ​​constructed by a mixed tensor decomposition method, the air pressure data is decomposed by an adaptive morphological wavelet packet decomposition algorithm to obtain the air pressure frequency domain eigenvalue, and the nonlinear temperature instantaneous eigenvalue, the multidimensional concentration distribution feature matrix and the air pressure frequency domain eigenvalue are input into a quantum deep autoencoder network to generate a quantum feature vector, including:

[0066] Arranging temperature sensors at four corner positions and a center position of the integrated circuit to be tested to obtain a temperature data matrix, arranging tracer gas sensors evenly inside the integrated circuit to be tested to obtain a tracer gas concentration data matrix, and arranging air pressure sensors outside the integrated circuit to be tested to obtain an air pressure data matrix;

[0067] Decomposing the temperature data matrix by a bidirectional empirical mode decomposition algorithm, screening the intrinsic mode function that satisfies the conditions that the difference between the number of extreme value points and the number of zero-crossing points does not exceed one, the local mean is zero, at least one zero-crossing point is included between adjacent local extreme value points, and the first and last extreme value points are respectively the local maximum point and the local minimum point, forward decomposing the intrinsic mode function from high to low frequency, and reverse decomposing it from low to high frequency to obtain a bidirectional temperature intrinsic mode function, and extracting nonlinear temperature instantaneous eigenvalues ​​from the bidirectional temperature intrinsic mode function by using fractional-order Hilbert transform and wavelet transform kernel function;

[0068] The tracer gas concentration data matrix is ​​subjected to denoising processing by using a non-local weighted adaptive anisotropic diffusion filter, and the sizes of the similarity window and the search window are adaptively adjusted according to the distribution characteristics of the tracer gas concentration data matrix to obtain a denoised tracer gas concentration data matrix;

[0069] The denoised tracer gas concentration data matrix is ​​constructed as a third-order tensor, wherein the first dimension represents the sensor number, the second dimension represents the time point number, and the third dimension represents the tracer gas concentration value, and the third-order tensor is decomposed by a nuclear tensor decomposition method to obtain a multi-dimensional concentration distribution feature matrix;

[0070] Preprocessing the air pressure data matrix, performing morphological opening and closing operations through disk-shaped structural elements, applying wavelet packet decomposition with adaptive layers to the preprocessed air pressure data, and extracting energy features and singular value features using an adaptive threshold strategy to obtain air pressure frequency domain eigenvalues;

[0071] The nonlinear temperature instantaneous eigenvalue, the multidimensional concentration distribution characteristic matrix and the pressure frequency domain eigenvalue are normalized and converted into a quantum state vector, the quantum state vector is input into a quantum deep autoencoder network, quantum NOT gates, Hadamard gates and rotation gates are set between adjacent layers of the quantum deep autoencoder network, and a quantum eigenvector is obtained through parameterized quantum circuit training, the quantum eigenvector contains comprehensive characteristic information of the temperature, tracer gas concentration and air pressure of the integrated circuit to be tested.

[0072] The zero-crossing point refers to the point where the signal changes from a positive value to a negative value or from a negative value to a positive value, and is usually used for frequency extraction or feature detection in signal analysis. The disk-shaped structure element is a structure element in the shape of a disk used in morphological image processing, and is often used for operations such as image expansion and corrosion, processing noise in images or performing morphological filtering. The quantum state vector is a complex-valued vector that describes the state of a quantum system in quantum mechanics, contains all information about the quantum system, and changes with quantum operations. The quantum NOT gate is one of the basic gates in quantum computing, which is used to flip the state of a quantum bit, which is equivalent to the NOT gate operation in classical computing. The Hadamard gate is a basic quantum gate in quantum computing. By applying a specific transformation to the quantum bit, its state can be transformed into an equal superposition state. The rotation gate is a gate used to rotate a quantum bit in quantum computing, and the control of the quantum bit state is achieved by adjusting the phase and amplitude of the quantum bit.

[0073] Temperature sensors are arranged at the four corners and the center of the IC to be tested to obtain a temperature data matrix. For example, a 5-row N-column matrix can be obtained, where 5 represents 5 sensors and N represents the sampling time point. At the same time, tracer gas sensors are evenly arranged inside the IC to obtain a tracer gas concentration data matrix. For example, a M-row N-column matrix can be obtained, where M represents M sensors. Air pressure sensors are arranged outside the IC to obtain an air pressure data matrix. For example, a 1-row N-column time series data can be obtained.

[0074] Perform bidirectional empirical mode decomposition on the temperature data matrix, and first select the intrinsic mode functions that meet the following conditions: the number of extreme points and the number of zero-crossing points do not differ by more than one; the local mean is zero; there is at least one zero-crossing point between adjacent local extreme points; the first and last extreme points are the local maximum point and the local minimum point, respectively. Assuming that three intrinsic mode functions that meet the conditions are selected, the intrinsic mode functions are first forward decomposed from high to low frequency, and then reversely decomposed from low to high frequency to obtain the bidirectional temperature intrinsic mode function. The fractional Hilbert transform and wavelet transform kernel function are used to extract the nonlinear temperature instantaneous eigenvalues ​​of the bidirectional temperature intrinsic mode function. For example, a matrix with 5 rows and K columns can be extracted, where K represents the number of extracted features.

[0075] The tracer gas concentration data matrix is ​​denoised using a non-local weighted adaptive anisotropic diffusion filter. The filter can adaptively adjust the size of the similarity window and the search window according to the distribution characteristics of the concentration data matrix. Assume that the similarity window size is set to 3x3 and the search window size is set to 5x5. The tracer gas concentration data matrix after denoising is still M rows and N columns. The denoised concentration data matrix is ​​constructed as a third-order tensor, where the first dimension represents the sensor number, the second dimension represents the time point number, and the third dimension represents the tracer gas concentration value. For example, a third-order tensor of M*N*1 can be obtained. The third-order tensor is decomposed by the nuclear tensor decomposition method to obtain a multi-dimensional concentration distribution feature matrix. For example, an M*L matrix can be obtained, where L represents the number of features after decomposition.

[0076] The pressure data matrix is ​​preprocessed, and a disk-shaped structural element is used to perform morphological opening and closing operations. The preprocessed pressure data is decomposed by wavelet packets with an adaptive number of layers, for example, the number of decomposition layers is 3. An adaptive threshold strategy is used to extract energy features and singular value features to obtain the pressure frequency domain eigenvalues. For example, a vector with 1 row and P columns can be obtained, where P represents the number of extracted features.

[0077] The nonlinear temperature instantaneous eigenvalues, multidimensional concentration distribution characteristic matrix and pressure frequency domain eigenvalues ​​are normalized and converted into quantum state vectors. These quantum state vectors are input into the quantum deep autoencoder network. Quantum NOT gates, Hadamard gates and rotation gates are set between adjacent layers of the network. The quantum eigenvector is obtained by parameterized quantum circuit training, which contains the comprehensive characteristic information of the temperature, tracer gas concentration and pressure of the integrated circuit to be tested.

[0078] In this embodiment, a variety of advanced signal processing and feature extraction technologies are used to effectively extract multi-dimensional information such as temperature, tracer gas concentration and air pressure of the integrated circuit to be tested, and integrate them into the quantum feature vector, thereby improving the accuracy and comprehensiveness of feature extraction. A non-local weighted adaptive anisotropic diffusion filter is used to perform noise reduction processing on the tracer gas concentration data, which effectively suppresses the influence of noise on feature extraction and enhances the robustness of the feature. The extracted feature vector is converted into a quantum state vector, and trained using a quantum deep autoencoder network to generate a quantum feature vector containing comprehensive feature information of the integrated circuit to be tested, which provides a foundation for subsequent quantum computing and quantum machine learning applications. In summary, this embodiment realizes high-quality collection and feature extraction of integrated circuit package sealing related data, lays a solid data foundation for sealing defect diagnosis and reliability assessment, and helps to improve the comprehensiveness and accuracy of package reliability assessment.

[0079] In an optional embodiment,

[0080] The tracer gas concentration data matrix is ​​subjected to noise reduction processing using a non-local weighted adaptive anisotropic diffusion filter as shown in the following formula:

[0081] ;

[0082] in, represents the rate of change of the tracer gas concentration data matrix over time, C ( x , y , t ) is the tracer gas concentration data matrix, indicating the spatial position ( x , y ) and time t The gas concentration value on x The horizontal coordinate of the sensor location, y Indicates the vertical coordinate of the sensor location, t Indicates time, w ( x , y ) represents the adaptive weight function, ▽ C ( x , y ) represents the gas concentration matrix at the spatial position ( x , y ), T represents transpose, ▽ C ( x , y , t ) represents the tracer gas concentration data matrix at time t The spatial gradient of .

[0083] In this embodiment, the non-local weighted adaptive anisotropic diffusion filter can be used to adaptively adjust the filter strength according to the gradient change characteristics of the tracer gas concentration at different positions, and the spatial gradient and the temporal gradient are considered at the same time, so as to realize the joint filtering of the tracer gas concentration data in the time and space dimensions, which can more accurately reflect the dynamic characteristics of gas diffusion. Different diffusion coefficients in different directions can better maintain the directional characteristics of the gas concentration distribution and avoid information loss caused by excessive smoothing. Through the constraint of the gradient term, the mutation characteristics of the tracer gas concentration distribution can be effectively maintained. These characteristics are usually closely related to the sealing defect position and are of great significance to the subsequent defect diagnosis. In summary, this embodiment enables the tracer gas concentration data after noise reduction to more accurately reflect the actual sealing condition of the integrated circuit package, and provide a high-quality data basis for subsequent feature extraction and defect diagnosis.

[0084] In an optional embodiment,

[0085] Extracting nonlinear temperature instantaneous eigenvalues ​​from the bidirectional temperature intrinsic mode function using fractional-order Hilbert transform and wavelet transform kernel function includes:

[0086] Adding a bidirectional temperature intrinsic mode function to a preset fractional-order Hilbert transform module, performing fractional differentiation on the bidirectional temperature intrinsic mode function to obtain a fractional-order derivative signal, performing Hilbert transform on the fractional-order derivative signal to obtain a first complex-valued analytical signal, and calculating a first nonlinear temperature instantaneous eigenvalue sequence based on the modulus and phase of the first complex-valued analytical signal;

[0087] Inputting the bidirectional temperature intrinsic mode function into a wavelet transform kernel function module, constructing a wavelet transform kernel function by using Morley wavelet as a wavelet basis function, performing continuous wavelet transform on the bidirectional temperature intrinsic mode function by using the wavelet transform kernel function to obtain a wavelet coefficient matrix on a time-frequency scale plane, and extracting wavelet ridges from the wavelet coefficient matrix to obtain a second nonlinear temperature instantaneous eigenvalue sequence;

[0088] Performing maximum and minimum value normalization processing on the first nonlinear temperature instantaneous eigenvalue sequence and the second nonlinear temperature instantaneous eigenvalue sequence to obtain a first normalized eigenvalue sequence and a second normalized eigenvalue sequence;

[0089] Calculate the signal-to-noise ratio of the first normalized eigenvalue sequence and the second normalized eigenvalue sequence, determine a first weight coefficient corresponding to the first normalized eigenvalue sequence and a second weight coefficient corresponding to the second normalized eigenvalue sequence based on the signal-to-noise ratio, multiply the first normalized eigenvalue sequence by the first weight coefficient to obtain a first weighted sequence, multiply the second normalized eigenvalue sequence by the second weight coefficient to obtain a second weighted sequence, and add the first weighted sequence to the second weighted sequence to obtain a fused nonlinear temperature instantaneous eigenvalue sequence.

[0090] The complex-valued analytical signal is a complex signal obtained by applying the Hilbert transform to a real-valued signal, wherein the real part is the original signal and the imaginary part is the Hilbert transform of the signal, and is widely used to analyze the instantaneous frequency and envelope of the signal. The wavelet ridge is a feature extracted from the signal by the wavelet transform, and is used to represent the local extreme points and changing trends of the signal, and is particularly suitable for analyzing the multi-scale characteristics of non-stationary signals.

[0091] Obtain bidirectional temperature intrinsic mode functions. For example, a series of intrinsic mode functions (IMFs) are obtained by performing empirical mode decomposition (EMD) on the temperature time series data collected by the temperature sensor, and two representative intrinsic mode functions are selected as bidirectional temperature intrinsic mode functions. Assume that the two intrinsic mode functions are IMF1 and IMF2, and their values ​​are a series of time series data, such as IMF1=[25.1, 25.3, 25.5, 25.2, 25.4...], IMF2=[0.1, 0.2, 0.15, 0.25, 0.18...].

[0092] The obtained bidirectional temperature intrinsic mode functions IMF1 and IMF2 are respectively input into the pre-set fractional Hilbert transform module. The core of this module is to perform fractional differentiation on the input signal. The order of fractional differentiation can be set according to the actual situation, for example, it is set to 0.5, and after performing 0.5-order differentiation on IMF1 and IMF2, two new time series are obtained.

[0093] Perform Hilbert transform on the fractional derivative signals D_IMF1 and D_IMF2 to obtain the first complex-valued analytical signal. The Hilbert transform can convert a real signal into a complex signal, and its imaginary part is the Hilbert transform of the real signal. Two complex-valued analytical signals are obtained, which are recorded as C_IMF1 and C_IMF2. For example, C_IMF1[0.15+0.05i, 0.2+0.1i, -0.3-0.15i, 0.2+0.12i, ...], C_IMF2=[0.1+0.02i, -0.05-0.01i, 0.1+0.03i, -0.07-0.02i, ...].

[0094] The first nonlinear temperature instantaneous eigenvalue sequence is obtained based on the modulus and phase of the first complex-valued analytical signals C_IMF1 and C_IMF2. For example, the modulus and phase of each complex value of C_IMF1 are calculated to obtain two new time series, which represent the instantaneous amplitude and instantaneous frequency of C_IMF1 respectively, and these two time series are combined into a new eigenvalue sequence Feature1. The same operation is performed on C_IMF2 to obtain the eigenvalue sequence Feature2.

[0095] Input the bidirectional temperature intrinsic mode functions IMF1 and IMF2 into the wavelet transform kernel function module. Select Morley wavelet as the wavelet basis function to construct the wavelet transform kernel function.

[0096] Use the constructed wavelet transform kernel function to perform continuous wavelet transform on IMF1 and IMF2 to obtain the wavelet coefficient matrix on the time-frequency scale plane. For example, after performing continuous wavelet transform on IMF1, a matrix Matrix1 is obtained, in which each row of the matrix represents a scale, each column represents a time point, and the values ​​in the matrix represent the wavelet coefficients of the corresponding scale and time point. Perform the same operation on IMF2 to obtain the matrix Matrix2.

[0097] Extract wavelet ridges from the wavelet coefficient matrices Matrix1 and Matrix2 to obtain the second nonlinear temperature instantaneous eigenvalue sequence. The wavelet ridge corresponds to the scale of the maximum wavelet coefficient at each time point in the matrix. Extract the wavelet ridges of Matrix1 to obtain the eigenvalue sequence Feature3. Extract the wavelet ridges of Matrix2 to obtain the eigenvalue sequence Feature4.

[0098] The first nonlinear temperature instantaneous eigenvalue sequence (Feature1 and Feature2) and the second nonlinear temperature instantaneous eigenvalue sequence (Feature3 and Feature4) are normalized by the maximum and minimum values ​​to obtain the first normalized eigenvalue sequence (N_Feature1 and N_Feature2) and the second normalized eigenvalue sequence (N_Feature3 and N_Feature4). Assuming that the maximum value in Feature1 is 10 and the minimum value is 1, each value in N_Feature1 will be converted to a value between 0 and 1.

[0099] Calculate the signal-to-noise ratio of the first normalized feature value sequence (N_Feature1 and N_Feature2) and the second normalized feature value sequence (N_Feature3 and N_Feature4). Assume that the signal-to-noise ratio of N_Feature1 is 10 and the signal-to-noise ratio of N_Feature3 is 5.

[0100] Based on the signal-to-noise ratio, the first weight coefficient corresponding to the first normalized eigenvalue sequence and the second weight coefficient corresponding to the second normalized eigenvalue sequence are determined. The higher the signal-to-noise ratio, the larger the corresponding weight coefficient. For example, according to the signal-to-noise ratio of N_Feature1 and N_Feature3, it can be determined that the weight coefficient of N_Feature1 is 0.67 and the weight coefficient of N_Feature3 is 0.33. The same operation is performed on N_Feature2 and N_Feature4.

[0101] The first normalized feature value sequence is multiplied by the first weight coefficient to obtain a first weighted sequence. For example, N_Feature1 is multiplied by 0.67 to obtain Weighted_Feature1.

[0102] The second normalized feature value sequence is multiplied by the second weight coefficient to obtain a second weighted sequence. For example, N_Feature3 is multiplied by 0.33 to obtain Weighted_Feature3.

[0103] The first weighted sequence is added to the second weighted sequence to obtain the fused nonlinear temperature instantaneous eigenvalue sequence. For example, Weighted_Feature1 is added to Weighted_Feature3 to obtain the final fused eigenvalue sequence Fused_Feature1. The same operation is performed on Weighted_Feature2 and Weighted_Feature4 to obtain Fused_Feature2.

[0104] In this embodiment, by combining fractional-order Hilbert transform and wavelet transform, the nonlinear characteristics of the temperature signal can be extracted from different angles, thereby improving the comprehensiveness and accuracy of feature extraction. Through signal-to-noise ratio weighted fusion, the influence of noise on feature extraction can be effectively reduced, and the noise resistance of the feature can be enhanced. The extracted nonlinear temperature instantaneous eigenvalues ​​can more accurately reflect the trend and law of temperature change, thereby improving the reliability of temperature monitoring. In summary, this embodiment not only ensures the accuracy and completeness of temperature feature extraction, but also improves the robustness of feature expression, providing a reliable feature basis for subsequent sealing defect diagnosis, and can effectively capture temperature abnormality characteristics related to sealing defects.

[0105] S2. Input the quantum feature vector into a multi-head cross attention network to generate an adaptive weight feature vector, construct a heterogeneous spatiotemporal graph neural network, extract the spatiotemporal dependency between feature nodes to obtain a spatiotemporal graph feature vector, perform multi-scale temporal feature alignment on the spatiotemporal graph feature vector through a waveform dynamic time warping algorithm, input the aligned feature sequence into a bidirectional gated recurrent attention network to extract long-term and short-term temporal dependencies to obtain a temporal dynamic feature sequence, and construct a hierarchical probabilistic graph model in combination with a deep variational inference method to obtain a probabilistic feature vector;

[0106] The spatiotemporal dependency refers to the relationship between data or phenomena in the two dimensions of time and space, and is often used to describe dynamic systems with time evolution and spatial distribution characteristics. The waveform dynamic time warping algorithm is an algorithm for comparing signal waveforms at different time scales. It mainly achieves alignment between waveforms by minimizing the nonlinear transformation of time series. It is widely used in speech recognition, signal analysis and other fields. The deep variational inference method is a deep learning method based on variational inference, which aims to approximate the posterior distribution of complex models by optimizing a lower bound, and is often used for inference problems in Bayesian deep learning.

[0107] In an optional embodiment,

[0108] The quantum feature vector is input into a multi-head cross attention network to generate an adaptive weight feature vector, a heterogeneous spatiotemporal graph neural network is constructed, the spatiotemporal dependency between feature nodes is extracted to obtain a spatiotemporal graph feature vector, the multi-scale temporal feature alignment of the spatiotemporal graph feature vector is performed through a waveform dynamic time warping algorithm, the aligned feature sequence is input into a bidirectional gated recurrent attention network to extract long-term and short-term temporal dependencies to obtain a temporal dynamic feature sequence, and a hierarchical probabilistic graph model is constructed in combination with a deep variational inference method to obtain a probabilistic feature vector including:

[0109] Input the quantum feature vector into a pre-set multi-head cross attention network, perform linear transformation on the quantum feature vector to obtain a query matrix, a key matrix and a value matrix, perform dot product operation on the query matrix and the key matrix to obtain an attention score matrix, use a softmax function on the attention score matrix to obtain an attention weight matrix, multiply the attention weight matrix and the value matrix to obtain a weighted feature matrix, and concatenate and linearly transform the weighted feature matrix to obtain an adaptive weight feature vector matrix;

[0110] The adaptive weight feature vector matrix is ​​used to construct a space-time graph according to the time series relationship in the time dimension, the feature vector of each time point in the adaptive weight feature vector matrix is ​​used as a node of the space-time graph, a directed edge connection is established between the nodes at adjacent time points in the space-time graph, and the feature vector of each node in the space-time graph and the feature vectors of the adjacent nodes of the current node are aggregated by a graph convolutional network to obtain a space-time graph feature vector matrix;

[0111] Perform wavelet decomposition on the feature vector matrix of the space-time graph in the time dimension to obtain multiple feature subsequences of different time scales, calculate the corresponding dynamic time warping distance matrix for each feature subsequence, solve the optimal matching path based on the dynamic time warping distance matrix in combination with a dynamic programming algorithm, map the feature subsequences onto a unified time axis according to the optimal matching path to obtain time-aligned feature subsequences, perform wavelet reconstruction on the time-aligned feature subsequences to obtain time-aligned feature vector matrix of the space-time graph;

[0112] Input the time-aligned spatiotemporal graph feature vector matrix into a bidirectional gated recurrent attention network, the forward gated recurrent unit and the backward gated recurrent unit of the bidirectional gated recurrent attention network respectively encode the time-aligned spatiotemporal graph feature vector matrix to obtain a forward hidden state sequence matrix and a backward hidden state sequence matrix, perform attention weighting on the forward hidden state sequence matrix and the backward hidden state sequence matrix to obtain a context vector sequence matrix, concatenate and linearly transform the context vector sequence matrix with the hidden state of the corresponding time step in the time-aligned spatiotemporal graph feature vector matrix to obtain a temporal dynamic feature sequence matrix;

[0113] The time series dynamic feature sequence matrix is ​​input into the deep variational inference model to construct a hierarchical probabilistic graphical model including a first layer, a second layer and a third layer. The first layer is used to extract high-level semantic features, the second layer is used to extract middle-level pattern features, and the third layer is used to extract bottom-level detail features. A variational inference algorithm is used to approximate the posterior distribution of the three-layer features of the hierarchical probabilistic graphical model. The parameters of the deep variational inference model are optimized by maximizing the lower bound of evidence to obtain a probabilistic feature vector matrix containing multi-level semantic information.

[0114] The spatiotemporal graph feature vector matrix is ​​a matrix representing spatiotemporal data, in which each element usually contains the characteristics of spatial position and time point, and is used to analyze the dependencies and patterns in spatiotemporal data. The dynamic time warping distance matrix is ​​a matrix calculated by the dynamic time warping algorithm, which represents the similarity between different time series and is usually used in pattern recognition and time series analysis. The dynamic programming algorithm is an algorithm for solving optimization problems. It decomposes the problem into sub-problems and uses the solutions of the sub-problems to gradually build the solution of the entire problem. It is widely used in path planning, sequence alignment and other problems. The context vector sequence matrix refers to a vector sequence representing the context information of a certain moment or a certain window in natural language processing or time series data analysis, which is used to capture the time dependency in time series data. The evidence lower bound is an important concept in variational inference, which is used to optimize the lower bound of the evidence of the latent variable model and approximate the true posterior distribution by maximizing the lower bound.

[0115] Perform quantum feature extraction on the input time series data. For example, the quantum kernel method is used to map the original time series data to a high-dimensional Hilbert space to obtain a quantum feature vector. Assume that the input time series data is a time series with a length of 100, and the data dimension of each time point is 5. After quantum feature extraction, a 100*10 quantum feature vector matrix is ​​obtained, in which each row represents the quantum feature of a time point.

[0116] The obtained quantum feature vector is input into a pre-set multi-head cross-attention network, and the quantum feature vector is linearly transformed to obtain the query matrix, key matrix and value matrix. For example, for the 10-dimensional quantum feature vector at each time point, it can be multiplied by three different transformation matrices to obtain three matrices with a dimension of 8, which are used as the query matrix, key matrix and value matrix respectively. The query matrix and the key matrix are dot-producted to obtain the attention score matrix. The attention score matrix is ​​processed by the softmax function to obtain the attention weight matrix, and the attention weight matrix is ​​multiplied by the value matrix to obtain the weighted feature matrix. The weighted feature matrices of multiple attention heads are concatenated and linearly transformed to obtain the final adaptive weight feature vector matrix with a dimension of 100*64.

[0117] Construct a heterogeneous spatiotemporal graph based on adaptive weight feature vectors. The feature vector of each time point is used as a node in the graph, and directed edges are established between nodes at adjacent time points to form a spatiotemporal graph. For example, the feature vector of the tth time point is connected with the feature vectors of the t-1th and t+1th time points, and the node features in the spatiotemporal graph are aggregated using a graph convolutional network. For example, the feature vectors of each node and its neighboring nodes are weighted averaged to obtain a new node feature vector. After the graph convolution operation, the spatiotemporal graph feature vector matrix is ​​obtained, and the dimension is still 100*64.

[0118] In order to align the temporal features of different time scales, the feature vector matrix of the space-time graph is aligned at multiple scales, and the feature vector matrix of the space-time graph is decomposed into multiple feature subsequences of different time scales using wavelet decomposition. For example, the feature sequence of 100 time points can be decomposed into two subsequences of 50 time points and 25 time points, and the dynamic time warping (DTW) distance matrix is ​​calculated for each feature subsequence. Based on the dynamic time warping distance matrix, the dynamic programming algorithm is used to find the optimal matching path. According to the optimal matching path, the feature subsequence is mapped to the unified time axis to obtain the time-aligned feature subsequence, and the aligned feature subsequence is reconstructed by wavelet to obtain the time-aligned space-time graph feature vector matrix, and the dimension is still 100*64.

[0119] The time-aligned spatiotemporal graph feature vector matrix is ​​input into the bidirectional gated recurrent attention network. The network contains two gated recurrent units (GRUs), one forward and one backward. The forward GRU starts encoding from the beginning of the sequence, and the backward GRU starts encoding from the end of the sequence. For example, for a 100*64 feature matrix, the forward GRU will obtain a 100*32 hidden state sequence matrix, and the backward GRU will also obtain a 100*32 hidden state sequence matrix. The forward and backward hidden state sequence matrices are weighted by attention to obtain the context vector sequence matrix. The context vector sequence matrix is ​​concatenated and linearly transformed with the hidden state of the corresponding time step to obtain the final temporal dynamic feature sequence matrix with a dimension of 100*128.

[0120] The temporal dynamic feature sequence matrix is ​​input into the deep variational inference model to construct a hierarchical probabilistic graphical model with three layers. The first layer is used to extract high-level semantic features, the second layer is used to extract middle-level pattern features, and the third layer is used to extract low-level detail features. The variational inference algorithm is used to approximate the posterior distribution of the three-layer features, and the model parameters are optimized by maximizing the lower bound of evidence to obtain a probabilistic feature vector matrix containing multi-level semantic information with a dimension of 100*256.

[0121] In this embodiment, by combining quantum features, multi-head cross-attention and spatiotemporal graph modeling, more discriminative and expressive features can be extracted, and the complex spatiotemporal dependencies in the data can be effectively captured. The dynamic time warping algorithm can effectively overcome problems such as inconsistent time series lengths and time offsets, and improve the model's robustness to noise and outliers. The deep variational inference model can learn the underlying probability distribution of the data, provide more interpretable results, and help to deeply understand the laws and mechanisms behind the data. In summary, this embodiment realizes deep mining of integrated circuit packaging reliability-related features, which not only improves the accuracy and interpretability of feature extraction, but also enhances the system's adaptability to complex working conditions, and provides high-quality feature support for subsequent reliability evaluation.

[0122] S3. Input the probability feature vector into the sealing defect diagnosis model based on knowledge distillation, output the sealing defect type and the corresponding confidence through the collaborative training of the teacher network and the student network, input the time series dynamic feature sequence into the pre-constructed causal long short-term memory network to predict the failure trend, construct a dynamic heterogeneous knowledge graph to match the failure mode and obtain the failure mode evolution law, based on the failure trend prediction result and the failure mode evolution law, calculate the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life through the multi-objective Pareto differential evolution algorithm, rate the packaging reliability of the integrated circuit to be tested, obtain the reliability rating result and output the sealing assessment report.

[0123] The sealing defect diagnosis model is a mathematical model used to detect and identify possible defects or leakage problems in the sealing system. It is usually based on sensor data and feature extraction technology, combined with machine learning algorithms for diagnosis and prediction. The failure trend refers to the trend of system or component performance degradation or failure over time under certain conditions. It is often used to predict the remaining service life or maintenance cycle of equipment. The multi-objective Pareto differential evolution algorithm is an optimization algorithm that aims to solve multi-objective optimization problems. It simulates natural selection and genetic variation processes and uses the Pareto optimal solution set to seek a balance between multiple objectives. It has strong global search capabilities. The sealing defect diffusion rate refers to the rate at which sealing defects propagate or diffuse in the medium. It is usually used to quantify the scope of influence of sealing defects in the system, which helps to assess potential risks and formulate maintenance strategies.

[0124] In an optional embodiment,

[0125] The probability feature vector is input into the sealing defect diagnosis model based on knowledge distillation, and the sealing defect type and the corresponding confidence are output through the collaborative training of the teacher network and the student network. The time series dynamic feature sequence is input into the pre-constructed causal long short-term memory network to predict the failure trend, and a dynamic heterogeneous knowledge graph is constructed to match the failure mode to obtain the failure mode evolution law. Based on the failure trend prediction result and the failure mode evolution law, the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life are calculated through the multi-objective Pareto differential evolution algorithm, and the packaging reliability of the integrated circuit to be tested is rated, and the reliability rating result is obtained and the sealing evaluation report is output, including:

[0126] Input the probability feature vector matrix composed of the probability feature vectors into the sealing defect diagnosis model based on knowledge distillation, extract high-level semantic features in the probability feature vector matrix through a deep residual convolutional neural network structure based on the teacher network in the sealing defect diagnosis model, output a first probability distribution of the sealing defect type, extract features through the mobile network structure of the student network based on the student network in the sealing defect diagnosis model, migrate the first probability distribution to the student network through a distillation loss function, optimize the second probability distribution output by the student network in combination with the cross entropy loss function, and obtain the sealing defect type and the corresponding confidence level;

[0127] Input the temporal dynamic feature sequence matrix into the encoder of the causal long short-term memory network, extract the temporal dependency and long-range context information through multi-layer gated recurrent units, obtain the failure trend feature vector, input the failure trend feature vector into the decoder of the causal long short-term memory network, enhance the feature selection ability through Granger causality constraints and sparse regularization terms, and generate a failure trend feature sequence as a failure trend prediction result;

[0128] Based on the failure trend feature sequence, a dynamic heterogeneous knowledge graph including failure mode nodes, failure mechanism nodes, failure symptom nodes and failure impact nodes is constructed, low-dimensional vector representations of nodes and relationships are learned through a graph embedding algorithm, and the semantic information and structural information of nodes are encoded in combination with a graph neural network based on an attention mechanism, node feature vectors are generated, and cluster analysis is performed on the failure mode nodes based on the similarity of the node feature vectors to obtain the evolution law between different failure modes, and association analysis is performed on the neighbor nodes of the failure mode nodes to determine the causal relationship and propagation path between the failure mode and the failure mechanism, failure symptom, and failure impact, so as to obtain the evolution law of the failure mode;

[0129] A multi-objective optimization model is established based on the failure trend prediction results and the failure mode evolution law, and the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life are taken as optimization targets. Material parameters, structural parameters, process parameters and environmental parameters are taken as optimization variables. The multi-objective optimization model is solved by using the Pareto differential evolution algorithm, and the Pareto optimal solution set is obtained through non-dominated sorting and congestion evaluation. A compromise solution is selected from the Pareto optimal solution set as the prediction value of different optimization targets, and the prediction value is graded and evaluated. Rating thresholds are set for the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life, and weight coefficients are assigned. The weighted average score is calculated according to the grade corresponding to the prediction value and the weight coefficient, and a sealing evaluation report including prediction indicators, rating results and weight analysis is generated.

[0130] The deep residual convolutional neural network is a deep neural network structure that allows signals to skip one or more layers in the network by introducing residual connections, thereby alleviating the gradient vanishing problem and improving the training effect and performance of the network. It is widely used in image processing and feature learning. The Granger causality constraint is a statistical method for analyzing time series data to determine whether one time series can help predict another time series, that is, to check whether there is a causal relationship. This constraint is usually used in fields such as system modeling and economic data analysis.

[0131] Multi-dimensional data collection is performed on integrated circuit packaging, including but not limited to: acoustic microscope images, X-ray images, scanning electron microscope images, and environmental data such as temperature, humidity, and pressure. The collected image data is preprocessed, including image denoising, image enhancement, and feature extraction operations, to extract probabilistic features related to sealing defects, such as the size, shape, and position of holes, and these probabilistic features are combined into probabilistic feature vectors, which in turn form a probabilistic feature vector matrix. At the same time, the dynamic features of the integrated circuit at different time points, such as current, voltage, temperature and other parameters, are collected to form a time series dynamic feature sequence, and a time series dynamic feature sequence matrix is ​​formed.

[0132] The probability feature vector matrix is ​​input into the sealing defect diagnosis model based on knowledge distillation. The model consists of two parts: the teacher network and the student network. The teacher network adopts a deep residual convolutional neural network structure, which can effectively extract high-level semantic features in the probability feature vector matrix. For example, it can identify whether there are cracks, delamination and other defect types, and output the preliminary probability distribution of the sealing defect type, such as the crack probability is 70%, the delamination probability is 20%, and the other defect probability is 10%. The student network adopts a lightweight mobile network structure, and the output probability distribution of the teacher network is transferred to the student network through the distillation loss function. At the same time, the output of the student network is optimized in combination with the cross entropy loss function to obtain more accurate sealing defect types and their corresponding confidence levels. For example, after the student network is optimized, the crack probability is 80% and the confidence level is 95%.

[0133] The time series dynamic feature sequence matrix is ​​input into the pre-built causal long short-term memory network. The encoder part of the network uses multi-layer gated recurrent units to extract dependencies and long-range context information in the time series data, such as the temperature change trend of the chip at different time points. The decoder part uses Granger causality constraints and sparse regularization terms to enhance feature selection capabilities and generate failure trend feature sequences, such as predicting that the chip temperature will continue to rise in the future.

[0134] Based on the failure trend feature sequence, a dynamic heterogeneous knowledge graph is constructed. The knowledge graph contains failure mode nodes (e.g., open circuit, short circuit), failure mechanism nodes (e.g., metal migration, corrosion), failure symptom nodes (e.g., current anomaly, voltage anomaly), and failure impact nodes (e.g., performance degradation, functional failure). The low-dimensional vector representation of nodes and relationships is learned through the graph embedding algorithm, and the semantic information and structural information of the nodes are encoded using the graph neural network based on the attention mechanism to generate node feature vectors. For example, the feature vector of the open circuit node can represent its relationship with other nodes and its own attributes. The failure mode nodes are clustered based on the similarity of the node feature vectors to obtain the evolution law between different failure modes. For example, an open circuit may evolve into a short circuit. At the same time, the neighbor nodes of the failure mode node are associated with each other to determine the causal relationship and propagation path between the failure mode and the failure mechanism, failure symptom, and failure impact. For example, the open circuit is caused by metal migration, which will cause current anomaly and eventually lead to chip functional failure.

[0135] Based on the failure trend prediction results and the evolution law of failure modes, a multi-objective optimization model is established. The sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life are taken as optimization targets, and material parameters (e.g., Young's modulus, Poisson's ratio), structural parameters (e.g., package size, wire bonding strength), process parameters (e.g., curing temperature, curing time) and environmental parameters (e.g., temperature, humidity) are taken as optimization variables. The Pareto differential evolution algorithm is used to solve the multi-objective optimization model to obtain the Pareto optimal solution set, and the compromise solution is selected as the prediction value of different optimization targets. For example, the sealing defect diffusion rate is predicted to be 1 micron per year, the stress distribution characteristic value is 10MPa, the failure probability is 1%, and the remaining life is 10 years. These predicted values ​​are graded and evaluated, for example, the diffusion rate is divided into three levels: low, medium and high. A rating threshold is set for each indicator and a weight coefficient is assigned, for example, the weight coefficient of the remaining life is 0.5 and the weight coefficient of the failure probability is 0.3. The weighted average score is calculated based on the grade and weight coefficient corresponding to the predicted value, and a sealing assessment report containing prediction indicators, rating results and weight analysis is generated.

[0136] For example, assuming that the predicted value of the remaining life of an integrated circuit package is 8 years, the corresponding level is "medium", and the weight coefficient is 0.5; the predicted value of the probability of failure is 0.5%, the corresponding level is "low", and the weight coefficient is 0.3; the levels and weight coefficients of other indicators have also been determined. Then the weighted average score of the package can be calculated as: (0.5*medium score) + (0.3*low score) + ..., and finally the reliability of the package is rated according to the weighted average score, for example, the rating is "good".

[0137] In this embodiment, by combining deep learning and knowledge distillation technology, various types of sealing defects can be more accurately identified and corresponding confidence levels can be given, thereby improving the accuracy of diagnosis. At the same time, the lightweight structure of the student network can improve the operating efficiency of the model and shorten the diagnosis time. Through the causal long short-term memory network and the dynamic heterogeneous knowledge graph, the failure trend of the integrated circuit package can be accurately predicted, and the evolution law between different failure modes and the causal relationship between the failure mechanism, failure signs, and failure effects can be deeply analyzed to provide guidance for the formulation of preventive measures. Through the multi-objective optimization algorithm and the hierarchical evaluation system, the reliability of the integrated circuit package can be comprehensively evaluated, and a sealing evaluation report containing prediction indicators, rating results, and weight analysis can be generated to provide a basis for decision-making. In summary, this embodiment provides comprehensive, accurate, and efficient diagnosis and evaluation support for the sealing problem of the integrated circuit package, and has important engineering application value.

[0138] In an optional embodiment,

[0139] The first probability distribution is transferred to the student network through the distillation loss function as shown in the following formula:

[0140] ;

[0141] in, L represents the loss value, α represents the weight coefficient, N represents the number of samples, M represents the number of categories, i represents the sample index, j Represents the category index, is the first probability distribution output by the teacher network, indicating the teacher network's i The sample in j The predicted probability of the class, is the second probability distribution output by the student network, indicating the student network's i The predicted probability of a sample in the jth class.

[0142] In this embodiment, by learning the probability distribution of the teacher network, the student network can converge faster, and under the same performance, the model scale is smaller, the computational cost is lower, and it is more suitable for deployment on resource-constrained devices. The predicted probability of the teacher network serves as a soft target to guide the student network to learn a finer-grained feature expression. The soft target of the teacher network contains more information than the hard label (true label), which can help the student network learn richer feature representations, thereby improving the generalization ability of the model and reducing the risk of overfitting. In summary, this embodiment realizes the efficient transfer of teacher network knowledge to the student network, which not only ensures the accuracy of sealing defect diagnosis, but also significantly improves the deployment efficiency of the model, providing reliable technical support for practical engineering applications.

[0143] Figure 2 FIG. 1 is a schematic diagram of a system for evaluating the sealing performance of integrated circuit packages under a high temperature environment according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0144] The first unit is used to obtain temperature data, tracer gas concentration data and air pressure data of the integrated circuit to be tested, decompose the temperature data by using a bidirectional empirical mode decomposition algorithm to obtain a bidirectional temperature intrinsic mode function, extract nonlinear temperature instantaneous eigenvalues ​​in combination with fractional-order Hilbert transform, reduce noise of the tracer gas concentration data by using a non-local weighted adaptive anisotropic diffusion filter, construct a multidimensional concentration distribution feature matrix by using a mixed tensor decomposition method, obtain air pressure frequency domain eigenvalues ​​by using an adaptive morphological wavelet packet decomposition algorithm for the air pressure data, and input the nonlinear temperature instantaneous eigenvalues, the multidimensional concentration distribution feature matrix and the air pressure frequency domain eigenvalues ​​into a quantum deep autoencoder network to generate a quantum eigenvector;

[0145] The second unit is used to input the quantum feature vector into a multi-head cross attention network to generate an adaptive weight feature vector, construct a heterogeneous space-time graph neural network, extract the space-time dependency between feature nodes to obtain a space-time graph feature vector, perform multi-scale temporal feature alignment on the space-time graph feature vector through a waveform dynamic time warping algorithm, input the aligned feature sequence into a bidirectional gated recurrent attention network to extract long-term and short-term temporal dependencies to obtain a temporal dynamic feature sequence, and construct a hierarchical probabilistic graph model in combination with a deep variational inference method to obtain a probabilistic feature vector;

[0146] The third unit is used to input the probability feature vector into a sealing defect diagnosis model based on knowledge distillation, output the sealing defect type and the corresponding confidence through the collaborative training of the teacher network and the student network, input the time series dynamic feature sequence into a pre-constructed causal long short-term memory network to predict the failure trend, construct a dynamic heterogeneous knowledge graph to match the failure mode to obtain the failure mode evolution law, and based on the failure trend prediction result and the failure mode evolution law, calculate the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life through a multi-objective Pareto differential evolution algorithm, rate the packaging reliability of the integrated circuit to be tested, obtain the reliability rating result and output a sealing assessment report.

[0147] According to a third aspect of the embodiments of the present invention,

[0148] An electronic device is provided, comprising:

[0149] processor;

[0150] a memory for storing processor-executable instructions;

[0151] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0152] A fourth aspect of the embodiments of the present invention is:

[0153] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0154] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for evaluating the sealing performance of integrated circuit packaging in a high temperature environment, characterized in that: include: Acquire temperature data, tracer gas concentration data and air pressure data of the integrated circuit to be tested, decompose the temperature data by using a bidirectional empirical mode decomposition algorithm to obtain a bidirectional temperature intrinsic mode function, extract nonlinear temperature instantaneous eigenvalues ​​by combining fractional-order Hilbert transform, perform noise reduction on the tracer gas concentration data by using a non-local weighted adaptive anisotropic diffusion filter, construct a multidimensional concentration distribution feature matrix by using a mixed tensor decomposition method, obtain air pressure frequency domain eigenvalues ​​by using an adaptive morphological wavelet packet decomposition algorithm for the air pressure data, and input the nonlinear temperature instantaneous eigenvalues, the multidimensional concentration distribution feature matrix and the air pressure frequency domain eigenvalues ​​into a quantum deep autoencoder network to generate a quantum eigenvector; Input the quantum feature vector into a multi-head cross attention network to generate an adaptive weight feature vector, construct a heterogeneous spatiotemporal graph neural network, extract the spatiotemporal dependency between feature nodes to obtain a spatiotemporal graph feature vector, perform multi-scale temporal feature alignment on the spatiotemporal graph feature vector through a waveform dynamic time warping algorithm, input the aligned feature sequence into a bidirectional gated recurrent attention network to extract long-term and short-term temporal dependencies to obtain a temporal dynamic feature sequence, and construct a hierarchical probabilistic graph model in combination with a deep variational inference method to obtain a probabilistic feature vector; The probability feature vector is input into the sealing defect diagnosis model based on knowledge distillation, and the sealing defect type and the corresponding confidence are output through the collaborative training of the teacher network and the student network. The time series dynamic feature sequence is input into the pre-constructed causal long short-term memory network to predict the failure trend, and a dynamic heterogeneous knowledge graph is constructed to match the failure mode to obtain the failure mode evolution law. Based on the failure trend prediction result and the failure mode evolution law, the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life are calculated through the multi-objective Pareto differential evolution algorithm, and the packaging reliability of the integrated circuit to be tested is rated, the reliability rating result is obtained and a sealing assessment report is output.

2. The method according to claim 1, characterized in that The temperature data, tracer gas concentration data and air pressure data of the integrated circuit to be tested are obtained, the temperature data are decomposed by a bidirectional empirical mode decomposition algorithm to obtain a bidirectional temperature intrinsic mode function, the nonlinear temperature instantaneous eigenvalue is extracted in combination with a fractional-order Hilbert transform, the tracer gas concentration data is denoised by a non-local weighted adaptive anisotropic diffusion filter, a multidimensional concentration distribution feature matrix is ​​constructed by a mixed tensor decomposition method, the air pressure data is decomposed by an adaptive morphological wavelet packet decomposition algorithm to obtain the air pressure frequency domain eigenvalue, and the nonlinear temperature instantaneous eigenvalue, the multidimensional concentration distribution feature matrix and the air pressure frequency domain eigenvalue are input into a quantum deep autoencoder network to generate a quantum feature vector, which includes: Arranging temperature sensors at four corner positions and a center position of the integrated circuit to be tested to obtain a temperature data matrix, arranging tracer gas sensors evenly inside the integrated circuit to be tested to obtain a tracer gas concentration data matrix, and arranging air pressure sensors outside the integrated circuit to be tested to obtain an air pressure data matrix; Decomposing the temperature data matrix by a bidirectional empirical mode decomposition algorithm, screening the intrinsic mode function that satisfies the conditions that the difference between the number of extreme value points and the number of zero-crossing points does not exceed one, the local mean is zero, at least one zero-crossing point is included between adjacent local extreme value points, and the first and last extreme value points are respectively the local maximum point and the local minimum point, forward decomposing the intrinsic mode function from high to low frequency, and reverse decomposing it from low to high frequency to obtain a bidirectional temperature intrinsic mode function, and extracting nonlinear temperature instantaneous eigenvalues ​​from the bidirectional temperature intrinsic mode function by using fractional-order Hilbert transform and wavelet transform kernel function; The tracer gas concentration data matrix is ​​subjected to denoising processing by using a non-local weighted adaptive anisotropic diffusion filter, and the sizes of the similarity window and the search window are adaptively adjusted according to the distribution characteristics of the tracer gas concentration data matrix to obtain a denoised tracer gas concentration data matrix; The denoised tracer gas concentration data matrix is ​​constructed as a third-order tensor, wherein the first dimension represents the sensor number, the second dimension represents the time point number, and the third dimension represents the tracer gas concentration value, and the third-order tensor is decomposed by a nuclear tensor decomposition method to obtain a multi-dimensional concentration distribution feature matrix; Preprocessing the air pressure data matrix, performing morphological opening and closing operations through disk-shaped structural elements, applying wavelet packet decomposition with adaptive layers to the preprocessed air pressure data, and extracting energy features and singular value features using an adaptive threshold strategy to obtain air pressure frequency domain eigenvalues; The nonlinear temperature instantaneous eigenvalue, the multidimensional concentration distribution characteristic matrix and the pressure frequency domain eigenvalue are normalized and converted into a quantum state vector, the quantum state vector is input into a quantum deep autoencoder network, quantum NOT gates, Hadamard gates and rotation gates are set between adjacent layers of the quantum deep autoencoder network, and a quantum eigenvector is obtained through parameterized quantum circuit training, the quantum eigenvector contains comprehensive characteristic information of the temperature, tracer gas concentration and air pressure of the integrated circuit to be tested.

3. The method according to claim 2, characterized in that The tracer gas concentration data matrix is ​​subjected to noise reduction processing using a non-local weighted adaptive anisotropic diffusion filter as shown in the following formula: ； in, represents the rate of change of the tracer gas concentration data matrix over time, C ( x , y , t ) is the tracer gas concentration data matrix, indicating the spatial position ( x , y ) and time t The gas concentration value on x The horizontal coordinate of the sensor location, y Indicates the vertical coordinate of the sensor location, t Indicates time, w ( x , y ) represents the adaptive weight function, ▽ C ( x , y ) represents the gas concentration matrix at the spatial position ( x , y ), T represents transpose, ▽ C ( x , y , t ) represents the tracer gas concentration data matrix at time t The spatial gradient of .

4. The method according to claim 2, characterized in that: Extracting nonlinear temperature instantaneous eigenvalues ​​from the bidirectional temperature intrinsic mode function using fractional-order Hilbert transform and wavelet transform kernel function includes: Adding a bidirectional temperature intrinsic mode function to a preset fractional-order Hilbert transform module, performing fractional differentiation on the bidirectional temperature intrinsic mode function to obtain a fractional-order derivative signal, performing Hilbert transform on the fractional-order derivative signal to obtain a first complex-valued analytical signal, and calculating a first nonlinear temperature instantaneous eigenvalue sequence based on the modulus and phase of the first complex-valued analytical signal; Inputting the bidirectional temperature intrinsic mode function into a wavelet transform kernel function module, constructing a wavelet transform kernel function by using Morley wavelet as a wavelet basis function, performing continuous wavelet transform on the bidirectional temperature intrinsic mode function by using the wavelet transform kernel function to obtain a wavelet coefficient matrix on a time-frequency scale plane, and extracting wavelet ridges from the wavelet coefficient matrix to obtain a second nonlinear temperature instantaneous eigenvalue sequence; Performing maximum and minimum value normalization processing on the first nonlinear temperature instantaneous eigenvalue sequence and the second nonlinear temperature instantaneous eigenvalue sequence to obtain a first normalized eigenvalue sequence and a second normalized eigenvalue sequence; Calculate the signal-to-noise ratio of the first normalized eigenvalue sequence and the second normalized eigenvalue sequence, determine a first weight coefficient corresponding to the first normalized eigenvalue sequence and a second weight coefficient corresponding to the second normalized eigenvalue sequence based on the signal-to-noise ratio, multiply the first normalized eigenvalue sequence by the first weight coefficient to obtain a first weighted sequence, multiply the second normalized eigenvalue sequence by the second weight coefficient to obtain a second weighted sequence, and add the first weighted sequence to the second weighted sequence to obtain a fused nonlinear temperature instantaneous eigenvalue sequence.

5. The method according to claim 1, characterized in that The quantum feature vector is input into a multi-head cross attention network to generate an adaptive weight feature vector, a heterogeneous spatiotemporal graph neural network is constructed, the spatiotemporal dependency between feature nodes is extracted to obtain a spatiotemporal graph feature vector, the multi-scale temporal feature alignment of the spatiotemporal graph feature vector is performed through a waveform dynamic time warping algorithm, the aligned feature sequence is input into a bidirectional gated recurrent attention network to extract long-term and short-term temporal dependencies to obtain a temporal dynamic feature sequence, and a hierarchical probabilistic graph model is constructed in combination with a deep variational inference method to obtain a probabilistic feature vector including: Input the quantum feature vector into a pre-set multi-head cross attention network, perform linear transformation on the quantum feature vector to obtain a query matrix, a key matrix and a value matrix, perform dot product operation on the query matrix and the key matrix to obtain an attention score matrix, use a softmax function on the attention score matrix to obtain an attention weight matrix, multiply the attention weight matrix and the value matrix to obtain a weighted feature matrix, and concatenate and linearly transform the weighted feature matrix to obtain an adaptive weight feature vector matrix; The adaptive weight feature vector matrix is ​​used to construct a space-time graph according to the time series relationship in the time dimension, the feature vector of each time point in the adaptive weight feature vector matrix is ​​used as a node of the space-time graph, a directed edge connection is established between the nodes at adjacent time points in the space-time graph, and the feature vector of each node in the space-time graph and the feature vectors of the adjacent nodes of the current node are aggregated by a graph convolutional network to obtain a space-time graph feature vector matrix; Perform wavelet decomposition on the feature vector matrix of the space-time graph in the time dimension to obtain multiple feature subsequences of different time scales, calculate the corresponding dynamic time warping distance matrix for each feature subsequence, solve the optimal matching path based on the dynamic time warping distance matrix in combination with a dynamic programming algorithm, map the feature subsequences onto a unified time axis according to the optimal matching path to obtain time-aligned feature subsequences, perform wavelet reconstruction on the time-aligned feature subsequences to obtain time-aligned feature vector matrix of the space-time graph; Input the time-aligned spatiotemporal graph feature vector matrix into a bidirectional gated recurrent attention network, the forward gated recurrent unit and the backward gated recurrent unit of the bidirectional gated recurrent attention network respectively encode the time-aligned spatiotemporal graph feature vector matrix to obtain a forward hidden state sequence matrix and a backward hidden state sequence matrix, perform attention weighting on the forward hidden state sequence matrix and the backward hidden state sequence matrix to obtain a context vector sequence matrix, concatenate and linearly transform the context vector sequence matrix with the hidden state of the corresponding time step in the time-aligned spatiotemporal graph feature vector matrix to obtain a temporal dynamic feature sequence matrix; The time series dynamic feature sequence matrix is ​​input into the deep variational inference model to construct a hierarchical probabilistic graphical model including a first layer, a second layer and a third layer. The first layer is used to extract high-level semantic features, the second layer is used to extract middle-level pattern features, and the third layer is used to extract bottom-level detail features. A variational inference algorithm is used to approximate the posterior distribution of the three-layer features of the hierarchical probabilistic graphical model. The parameters of the deep variational inference model are optimized by maximizing the lower bound of evidence to obtain a probabilistic feature vector matrix containing multi-level semantic information.

6. The method according to claim 1, characterized in that The probability feature vector is input into the sealing defect diagnosis model based on knowledge distillation, and the sealing defect type and the corresponding confidence are output through the collaborative training of the teacher network and the student network. The time series dynamic feature sequence is input into the pre-constructed causal long short-term memory network to predict the failure trend, and a dynamic heterogeneous knowledge graph is constructed to match the failure mode to obtain the failure mode evolution law. Based on the failure trend prediction result and the failure mode evolution law, the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life are calculated through the multi-objective Pareto differential evolution algorithm, and the packaging reliability of the integrated circuit to be tested is rated, and the reliability rating result is obtained and the sealing evaluation report is output, including: Input the probability feature vector matrix composed of the probability feature vectors into the sealing defect diagnosis model based on knowledge distillation, extract high-level semantic features in the probability feature vector matrix through a deep residual convolutional neural network structure based on the teacher network in the sealing defect diagnosis model, output a first probability distribution of the sealing defect type, extract features through the mobile network structure of the student network based on the student network in the sealing defect diagnosis model, migrate the first probability distribution to the student network through a distillation loss function, optimize the second probability distribution output by the student network in combination with the cross entropy loss function, and obtain the sealing defect type and the corresponding confidence level; Input the temporal dynamic feature sequence matrix into the encoder of the causal long short-term memory network, extract the temporal dependency and long-range context information through multi-layer gated recurrent units, obtain the failure trend feature vector, input the failure trend feature vector into the decoder of the causal long short-term memory network, enhance the feature selection ability through Granger causality constraints and sparse regularization terms, and generate a failure trend feature sequence as a failure trend prediction result; Based on the failure trend feature sequence, a dynamic heterogeneous knowledge graph including failure mode nodes, failure mechanism nodes, failure symptom nodes and failure impact nodes is constructed, low-dimensional vector representations of nodes and relationships are learned through a graph embedding algorithm, and the semantic information and structural information of nodes are encoded in combination with a graph neural network based on an attention mechanism, node feature vectors are generated, and cluster analysis is performed on the failure mode nodes based on the similarity of the node feature vectors to obtain the evolution law between different failure modes, and association analysis is performed on the neighbor nodes of the failure mode nodes to determine the causal relationship and propagation path between the failure mode and the failure mechanism, failure symptom, and failure impact, so as to obtain the evolution law of the failure mode; A multi-objective optimization model is established based on the failure trend prediction results and the failure mode evolution law, and the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life are taken as optimization targets. Material parameters, structural parameters, process parameters and environmental parameters are taken as optimization variables. The multi-objective optimization model is solved by using the Pareto differential evolution algorithm, and the Pareto optimal solution set is obtained through non-dominated sorting and congestion evaluation. A compromise solution is selected from the Pareto optimal solution set as the prediction value of different optimization targets, and the prediction value is graded and evaluated. Rating thresholds are set for the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life, and weight coefficients are assigned. The weighted average score is calculated according to the grade corresponding to the prediction value and the weight coefficient, and a sealing evaluation report including prediction indicators, rating results and weight analysis is generated.

7. The method according to claim 6, characterized in that The first probability distribution is transferred to the student network through the distillation loss function as shown in the following formula: ； in, L represents the loss value, α represents the weight coefficient, N represents the number of samples, M represents the number of categories, i represents the sample index, j Represents the category index, is the first probability distribution output by the teacher network, indicating the teacher network's i The sample in j The predicted probability of the class, is the second probability distribution output by the student network, indicating the student network's i The predicted probability of a sample in the jth class.

8. A system for evaluating the sealing performance of integrated circuit packages in a high temperature environment, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain temperature data, tracer gas concentration data and air pressure data of the integrated circuit to be tested, decompose the temperature data by using a bidirectional empirical mode decomposition algorithm to obtain a bidirectional temperature intrinsic mode function, extract nonlinear temperature instantaneous eigenvalues ​​in combination with fractional-order Hilbert transform, reduce noise of the tracer gas concentration data by using a non-local weighted adaptive anisotropic diffusion filter, construct a multidimensional concentration distribution feature matrix by using a mixed tensor decomposition method, obtain air pressure frequency domain eigenvalues ​​by using an adaptive morphological wavelet packet decomposition algorithm for the air pressure data, and input the nonlinear temperature instantaneous eigenvalues, the multidimensional concentration distribution feature matrix and the air pressure frequency domain eigenvalues ​​into a quantum deep autoencoder network to generate a quantum feature vector; The second unit is used to input the quantum feature vector into a multi-head cross attention network to generate an adaptive weight feature vector, construct a heterogeneous space-time graph neural network, extract the space-time dependency between feature nodes to obtain a space-time graph feature vector, perform multi-scale temporal feature alignment on the space-time graph feature vector through a waveform dynamic time warping algorithm, input the aligned feature sequence into a bidirectional gated recurrent attention network to extract long-term and short-term temporal dependencies to obtain a temporal dynamic feature sequence, and construct a hierarchical probabilistic graph model in combination with a deep variational inference method to obtain a probabilistic feature vector; The third unit is used to input the probability feature vector into a sealing defect diagnosis model based on knowledge distillation, output the sealing defect type and the corresponding confidence through the collaborative training of the teacher network and the student network, input the time series dynamic feature sequence into a pre-constructed causal long short-term memory network to predict the failure trend, construct a dynamic heterogeneous knowledge graph to match the failure mode to obtain the failure mode evolution law, and based on the failure trend prediction result and the failure mode evolution law, calculate the sealing defect diffusion rate, stress distribution characteristic value, failure probability and remaining life through a multi-objective Pareto differential evolution algorithm, rate the packaging reliability of the integrated circuit to be tested, obtain the reliability rating result and output a sealing assessment report.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images