Chip anti-radiation performance parameter selection system based on multi-modal data

Abstract

The application discloses a kind of chip anti-radiation performance parameter selection system based on multi-modal data, including parameter acquisition module, chip anti-radiation multi-modal performance parameter correlation discovery module, chip anti-radiation performance parameter selection module and visual display module, the selection system of the present application can produce the entire anti-radiation performance parameter of relevant chip searched, and it is associated with analysis, provide the chip list of better anti-radiation performance for researcher to select, by collecting the multi-modal data of chip anti-radiation, and automatically find the correlation therein, realizes the automation and intelligent of anti-radiation performance parameter selection;Anti-radiation performance parameter selection based on deep learning technology can improve the speed and accuracy of chip anti-radiation performance parameter selection work, by intelligent analysis and processing multi-modal data, reduce the research and development personnel query comparison analysis work, reduce chip research and development cost and time, improve work efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of electronic information technology, and relates to a chip anti-radiation performance parameter selection system based on multi-modal data. BACKGROUND

[0002] In the field of modern technology, electronic chips have been widely used. However, in some special environments, such as nuclear power plants, aerospace, medical treatment, etc., these chips may be affected by high-energy radiation, resulting in equipment damage or performance degradation. Therefore, designing and selecting chips with anti-radiation performance has become an important task.

[0003] Currently, the selection of chip anti-radiation performance parameters is usually based on a small amount of data and empirical rules, and lacks systematicness and reliability. Therefore, a chip anti-radiation performance parameter selection system based on multi-modal data is needed to improve the reliability of chip anti-radiation performance selection and design.

[0004] Currently, the defects of the prior art mainly manifest in the following aspects:

[0005] 1. Lack of systematicness: The chip anti-radiation performance parameter selection of the prior art is usually based on a small amount of data and empirical rules, and cannot comprehensively consider various influencing factors and data sources of chip anti-radiation performance.

[0006] 2. Poor reliability: The chip anti-radiation performance parameter selection of the prior art is usually based on a small amount of data and empirical rules, and cannot accurately and reliably determine the anti-radiation performance parameters of the chip.

[0007] 3. Low efficiency: The chip anti-radiation performance parameter selection of the prior art usually requires researchers to manually screen and analyze, which is low in efficiency. SUMMARY

[0008] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a chip anti-radiation performance parameter selection system based on multi-modal data, to improve the speed and accuracy of chip anti-radiation performance parameter selection, to reduce chip research and development cost and time, and to improve work efficiency.

[0009] In order to achieve the above purpose, the following technical solutions are adopted in the present application:

[0010] The chip anti-radiation performance parameter selection system based on multi-modal data comprises:

[0011] A parameter acquisition module: for acquiring anti-radiation multi-modal data of different types of anti-radiation chips, and storing the data into a multi-modal database according to different data structures, the anti-radiation multi-modal data including but not limited to chip model, packaging process, radiation dose, working temperature and failure rate;

[0012] a chip anti-radiation multi-modal performance parameter correlation relationship discovery module: for integrating and correlating the anti-radiation multi-modal data of anti-radiation chips obtained from different sources to construct the anti-radiation multi-modal data correlation relationship of the chips, the anti-radiation multi-modal data types of the anti-radiation chips including chip anti-radiation performance description text, anti-radiation performance experiment result curve data or chip circuit structure diagram;

[0013] a chip anti-radiation performance parameter selection module: for performing entity matching between the chip demand conditions input by the researchers and the multi-modal database, transmitting the matched multi-modal performance parameters to the chip anti-radiation multi-modal performance parameter correlation relationship discovery module for processing, and automatically selecting suitable chip anti-radiation performance parameters and corresponding chip models according to the multi-modal data collected by the chip parameter acquisition module and the analysis processing results of the chip anti-radiation multi-modal performance parameter correlation relationship discovery module for the researchers to select;

[0014] a visual display module: for presenting the chip models and their anti-radiation performance parameter results obtained by the chip anti-radiation performance parameter selection module to the researchers, and the researchers can filter and sort the results according to their needs.

[0015] Further, in the chip anti-radiation performance parameter selection module, the researchers describe the demand for chips, and the system extracts key information according to the input text to obtain chip demand condition information chip anti-radiation performance indicators LET or total ionizing dose effect.

[0016] Further, the chip anti-radiation multi-modal performance parameter correlation relationship discovery module analyzes the chip anti-radiation multi-modal data through deep learning technology to discover the correlation relationship between the anti-radiation performance parameters.

[0017] Further, the specific process for the chip anti-radiation multi-modal performance parameter correlation relationship discovery module to construct the chip anti-radiation multi-modal data correlation relationship is as follows:

[0018] (1.1) Integrate and correlate the anti-radiation multi-modal data of anti-radiation chips obtained from different sources, and randomly extract 1 item from each modal anti-radiation performance parameter of the matched chips to form a multi-modal anti-radiation performance parameter combination;

[0019] (1.2) Different deep learning methods are used for feature representation of different modal data in the combination, i.e. extracting feature vectors;

[0020] For anti-radiation performance experiment result curve data, as time series data, feature representation is performed through the LSTM encoding module as E1=LSTM(M_waveform), and the specific process is as follows:

[0021] First, the time series data input to the LSTM module is normalized to scale the data to the range of 0-1;

[0022] Then, the data is formed into an array with a shape of (n_samples, look_back, n_features), where n_samples is the number of samples, look_back is the parameter indicating how many consecutive time points are used as input, and n_features is the number of features at each time point. Set look_back = 10;

[0023] Finally, the generated array data is processed through an LSTM layer and a fully connected layer to output a feature vector, where the input size of the LSTM layer is the same as the number of features of the data, the size of the hidden layer is 512, and the fully connected layer converts the output of the LSTM layer into a feature vector E1 with a size of (1, 1000);

[0024] For the chip circuit structure diagram, as image data, it is represented as E2 = ResNet-LSTM(M_image) through the ResNet-LSTM encoding module, and the specific process is as follows:

[0025] First, the image is preprocessed and adjusted to an RGB image 3*224*224;

[0026] Then, load the pre-trained ResNet50 model and remove the last fully connected layer and average pooling layer;

[0027] Again, flatten the output of the ResNet50 model into a vector (1, 1, 100352) and input it into a fully connected layer, which converts it into a vector with the required size (1, 1, 512); continue to input the transformed vector into an LSTM layer, which processes it as a sequence of time steps and outputs a final LSTM state;

[0028] Finally, input the LSTM state into a fully connected layer to output a feature vector E2 with a size of (1, 512);

[0029] For the chip anti-radiation performance description text, as text data, it is represented as E3 = Bert(M_text) through the Bert encoding module, and the specific process is as follows:

[0030] First, convert the input text into a digital form, i.e. use the WordPiece tokenizer to tokenize the input text, and map each word to the corresponding number in the BERT model's vocabulary (vocab_size = 21128);

[0031] Then, the position encoding of the tokenization of the text is performed according to the position encoding formula: positional_encoding[i,2*j] = sin(i / 10000^(2*j / hidden_size)) and positional_encoding[i,2*j+1] = cos(i / 10000^(2*j / hidden_size)), where i represents the index of the position encoding, j represents the index of the hidden unit, and hidden_size is the number of hidden units in the BERT model, i.e., hidden_size = 768;

[0032] Again, the above encoded sequence is input into the BERT model for forward propagation, and is converted into a series of vectors through 12 layers of self-attention mechanism, full connection and pooling operation, where each vector represents a token in the input text.

[0033] Finally, the average value of these vectors is output as a feature vector E3 with a size of (1, 768).

[0034] For a given chip anti-radiation performance index LET, it is represented as E4 = Linear(index) through the Linear encoding module, and the specific process is as follows:

[0035] First, the given LET value is taken as input, and the index to be divided is set as num_bins = 10 and the width of the bin is set as bin_width = (let_max - let_min) / num_bins; then, the LET is assigned to a bin, and the LET value is subtracted from the lower limit of the LET range and then divided by the bin width to obtain the index of the bin where the value is located: (let_val - let_min) / bin_width, if the calculated index is out of the range of 0 to (num_bins - 1), it needs to be limited within this range; _ _

[0036] Again, through a loop, the center value of each bin is formed into a discrete vector with a length of 10;

[0037] Finally, feature extraction is performed through a full connection layer linear transformation operation, and a feature vector E4 with a size of (1, 30) is output.

[0038] ​​(1.3) The feature vectors of each modal performance parameter are subjected to linear transformation for dimension conversion, so that each performance index is represented in the same feature vector space. The linear transformation L1 has a dimension of (1000, 10), the linear transformation L2 has a dimension of (512, 10), and the linear transformation L3 has a dimension of (768, 10). The converted feature vectors are E'1 = E1*L1, E'2 = E2*L2, and E'3 = E3*L3, all having a size of (1, 10);

[0039] (1.4) The features E'1, E'2, and E'3 in the same vector space are fused by the second dimension splicing method to obtain a common feature representation W = [E'1, E'2, E'3] with a size of (1, 30);

[0040] (1.5) The correlation between the feature representation W and the feature vector E4 in the same feature vector space is calculated using the cosine similarity method, and the result is saved to a list. The cosine similarity function is similarity(x, y) = (x·y) / (||x||·||y||), and the correlation between W and E4 is S = similarity(W, E4). S is added to the list.

[0041] Steps 1.1-1.5 are repeated to obtain a list of correlation degrees of multiple multi-modal data combinations, denoted as {S1, S2,..Sx}.

[0042] Further, the automatic selection process of the chip radiation resistance performance parameter selection module is as follows:

[0043] (2.1) According to the chip requirements input by the research and development personnel, entity matching is performed in the multi-modal database to obtain chip radiation resistance multi-modal performance parameters;

[0044] (2.2) The chip radiation resistance multi-modal data performance parameters are combined and feature fused by using the chip radiation resistance multi-modal performance parameter correlation relationship discovery module, and the feature similarity with the chip radiation resistance performance index LET or total ionizing dose effect, i.e., the data correlation degree S, is calculated.

[0045] (2.3) The correlation degree list {S1, S2,..Sx} is merged, and the chips are sorted according to the correlation degree to obtain a list of chips with higher correlation degrees. The higher the correlation degree, the better the radiation resistance performance of the chip, which is used as a reference object for selection.

[0046] Further, the data acquisition module uses manual uploading, website crawling and data extraction methods to obtain anti-radiation multi-modal data of different types of anti-radiation chips; wherein the manual uploading is the relevant parameters and experimental results of the experimental test chip of the research and development personnel; the website crawling is to use the crawling tool embedded in the parameter acquisition module to collect chip-related data regularly from the specified website; the data extraction is to extract chip-related data from the relevant anti-radiation chip data system.

[0047] Further, the visualization display module displays the relevant data related to the chip with high correlation of chip anti-radiation performance parameters in the form of text, list, line chart or column chart.

[0048] The chip anti-radiation performance parameter selection system based on multi-modal data can generate all anti-radiation performance parameters related to the searched chip, and perform correlation analysis, provide a chip list with better anti-radiation performance for the research and development personnel to select, automatically find the correlation in the multi-modal data of the chip anti-radiation, and realize the automation and intelligentization of the anti-radiation performance parameter selection.

[0049] The present application has the following advantages:

[0050] 1. The anti-radiation performance parameter selection based on deep learning technology can improve the speed and accuracy of chip anti-radiation performance parameter selection;

[0051] 2. Through intelligent analysis and processing of multi-modal data, the research and development personnel's query, comparison and analysis work is reduced, the chip research and development cost and time are reduced, and the work efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The system framework diagram of the present application is shown in the figure;

[0053] Figure 2 The chip anti-radiation multi-modal performance parameter correlation relationship discovery module structure diagram is shown in the figure;

[0054] Figure 3 The ResNet-LSTM encoding module structure diagram is shown in the figure;

[0055] Figure 4 The Linear encoding module structure diagram is shown in the figure;

[0056] Figure 5 The chip anti-radiation performance parameter selection module structure diagram is shown in the figure;

[0057] Figure 6 The multi-modal data combination example diagram is shown in the figure;

[0058] Figure 7 A multi-modal data combination and LET correlation descending order arrangement chart;

[0059] Figure 8 A chip anti-radiation performance parameter selection result display chart. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present application will be described clearly and completely below.

[0061] The system structure of the present application is shown in Figure 1 The chip parameter acquisition module acquires the anti-radiation multi-modal data of the chip, including the circuit structure, the anti-radiation performance experimental result curve data, the chip anti-radiation performance specification text data, etc. The chip anti-radiation multi-modal performance parameter correlation discovery module analyzes the chip anti-radiation multi-modal data through deep learning technology to discover the correlation between the anti-radiation performance parameters. The chip anti-radiation performance parameter selection module automatically selects the suitable chip anti-radiation performance parameters according to the multi-modal data collected by the chip parameter acquisition module and the analysis result of the chip anti-radiation multi-modal performance parameter correlation discovery module. The visual display module displays the multi-modal data of the chip and the selection result in a graphical and intuitive manner.

[0062] The implementation steps of the present application are as follows:

[0063] I. Parameter acquisition module

[0064] (1) The data acquisition adopts manual uploading, website crawling and data extraction to obtain the relevant parameters of different types of chips, including the information of chip materials, structure, process, etc. Among them, the manual uploading is mainly the relevant parameters and experimental results of the chip tested by the research and development personnel; the website crawling is mainly to use the embedded crawling tool of the parameter acquisition module to collect chip-related data from the specified website regularly; the data extraction is mainly to extract chip-related data from the data system of the manufacturers or units cooperating with the present application unit.

[0065] (2) Since the same modal data obtained by different ways has differences, it needs to be unified according to certain rules, such as date format, numerical value format, so as to standardize the data format by converting the format; the null value, repeated value, abnormal value, etc. in the data are cleaned to ensure the integrity and accuracy of the data.

[0066] (3) According to the different structures of the data, it is stored in the multi-modal database.

[0067] II. Chip anti-radiation multi-modal performance parameter correlation discovery module

[0068] (1) Randomly select one item from each modal anti-radiation performance parameter of the matched chip (such as anti-radiation performance experimental result curve data, circuit structure diagram, chip anti-radiation performance specification text data) to form a multi-modal anti-radiation performance parameter combination. Let the extracted data be M.

[0069] (2) Different deep learning methods are used for feature representation of different modal data in the combination, that is, feature vectors are extracted.

[0070] For anti-radiation performance experimental result curve data, as time series data, it is represented by an LSTM encoding module as E1=LSTM(M_waveform). In this process, the input of the LSTM module is time series data. First, normalize the time series data to scale the data to the range of 0-1 to make it easier for the LSTM model to learn. Then, form an array with a shape of (n_samples, look_back, n_features) (n_samples is the number of samples, look_back parameter indicates how many consecutive time points are used as input, set look_back=10, n_features is the number of features at each time point). Finally, the generated array data is processed by an LSTM layer and a fully connected layer (linear layer) to output a feature vector, where the input size of the LSTM layer is the same as the number of features of the data, and the size of the hidden layer is 512. The fully connected layer converts the output of the LSTM layer into a feature vector with a size of (1, 1000) E1.

[0071] For circuit structure diagram, as image data, it is represented by a ResNet-LSTM encoding module (such as Figure 4 shown) as E2=ResNet-LSTM(M_image). In this process, first, the image is preprocessed and adjusted to an RGB image (3*224*224); then, load the pre-trained ResNet50 model and remove the last fully connected layer and average pooling layer to extract features from the image; again, flatten the output of the ResNet50 model into a vector (1, 1, 100352) and input it into a fully connected layer, which converts it into a vector with the desired size (1, 1, 512); continue to input the transformed vector into an LSTM layer, which processes it as a sequence of time steps and outputs a final LSTM state; finally, input the LSTM state into a fully connected layer to output a feature vector E2 with a size of (1, 512).

[0072] For the chip anti-radiation performance description text data, as text data is represented as E3=Bert(M_text) by the Bert encoding module. In this process, first, the input text is converted into digital form, that is, the input text is segmented by using the WordPiece tokenizer, and each word is mapped to the corresponding number in the vocabulary table (vocab_size=21128) of the BERT model; then, the segmented text is positionally encoded according to the position encoding formula: positional_encoding[i,2*j]=sin(i / 10000^(2*j / hidden_size)) and positional_encoding[i,2*j+1]=cos(i / 10000^(2*j / hidden_size)), where i represents the index of the position encoding, j represents the index of the hidden unit, and hidden_size is the number of hidden units in the BERT model, that is, hidden_size=768; again, the above encoding sequence is input into the BERT model for forward propagation, and is converted into a series of vectors through 12 layers of self-attention mechanism, full connection and pooling operation, where each vector represents a token in the input text; finally, the average value of these vectors is output as a feature vector E3 with a size of (1, 768).

[0073] As shown in Figure 4 , for the chip anti-radiation performance index (LET or total ionizing dose effect), it is represented as E4=Linear(index) by the Linear encoding module. Taking the chip anti-radiation performance index LET as an example, in this process, first, the given LET value is taken as input, and the index to be divided is set: num _ bins=10 and the width of the bin: bin_width=(let_max-let_min) / num_bins; then, the LET is assigned to a bin, and the LET value is subtracted from the lower limit of the LET range, and then divided by the bin width, that is, the index of the bin where the value is located: (let_val-let_min) / bin_width, if the calculated index exceeds the range of 0 to (num _ bins-1), it needs to be limited within this range; again, through a loop, the center value of each bin is formed into a discrete vector with a length of 10; finally, the feature is extracted by a fully connected layer linear transformation operation, and a feature vector E4 with a size of (1, 30) is output.

[0074] (3) Since the feature representations of the performance parameters of each modality have different dimensions, linear transformations are used for dimension conversion respectively, so that they can be represented in the same feature vector space with the performance indicators. Let the dimension of linear transformation L1 be (1000, 10), the dimension of linear transformation L2 be (512, 10), and the dimension of linear transformation L3 be (768, 10). The converted feature vectors are E'1 = E1*L1, E'2 = E2*L2, and E'3 = E3*L3, all of which have a size of (1, 10).

[0075] (4) The feature representations E'1, E'2, and E'3 in the same vector space are fused by using the second dimension splicing method to obtain the common feature representation W = [E'1, E'2, E'3] with a size of (1, 30).

[0076] (5) The correlation degree between the feature representation W in the same feature vector space and E4 is calculated by using the cosine similarity method, and the result is saved to a list. Let the cosine similarity function be similarity(x, y) = (x·y) / (||x||·||y||), then the correlation degree between W and E4 is S = similarity(W, E4), and S is added to the list.

[0077] Repeat steps 1-5 to obtain a list of correlation degrees of multiple multi-modal data combinations, denoted as {S1, S2,..Sx}.

[0078] Three, chip anti-radiation performance parameter selection module

[0079] (1) According to the input chip requirements, perform entity matching in the multi-modal database to obtain chip anti-radiation multi-modal performance parameters.

[0080] (2) Use the chip anti-radiation multi-modal performance parameter correlation relationship discovery module to combine and feature fuse the chip anti-radiation multi-modal data performance parameters, and calculate the feature similarity with the chip anti-radiation performance indicators LET or total ionizing dose effect, i.e. data correlation degree S.

[0081] (3) Merge the correlation degree list {S1, S2,..Sx} and sort the chips according to the correlation degree to obtain a list of chips with higher correlation degrees. In this list, chips with higher correlation degrees often have better anti-radiation performance and can be used as a reference for selection.

[0082] Four, visualization display module

[0083] The visualization display module displays the relevant data texts, lists, line charts, column charts, etc. involved in the chip anti-radiation performance parameter correlation degree higher chip list, making it convenient for researchers to compare and evaluate different selection schemes.

[0084] The use process of the application is described below through specific application examples:

[0085] 1. Input chip requirements: "DC / DC converter with anti-single particle capability (LET threshold) greater than 75 MeV.cm2 / mg", after entity matching with the multi-modal database, multiple DC / DC related data are obtained, wherein the chip anti-radiation performance description text data includes: "T030210_T110910_SMSA2812D.pdf", "PPM-96-004.pdf", "B080100a.txt", the circuit structure diagram includes: "12-15V Boost converter circuit.jpg", "1Boost topology structure.jpg", "Product circuit overall block diagram.jpg", "1Buck type DC-DC topology structure.jpg", and the anti-radiation performance experimental result curve data includes: "Boost converter output voltage simulation experimental result.csv", "Buck type DC-DC electrical parameter.xlsx", "DC_DC converter simulation experimental result.csv". Figure 2 .1Boost topology structure.jpg", "Product circuit overall block diagram.jpg", "1Buck type DC-DC topology structure.jpg", and the anti-radiation performance experimental result curve data includes: "Boost converter output voltage simulation experimental result.csv", "Buck type DC-DC electrical parameter.xlsx", "DC_DC converter simulation experimental result.csv". Figure 3 Figure 5 2. As shown in

[0086] 2. As shown in Figure 6 , the obtained data are combined, and different ways are used for feature extraction according to the modes, wherein the text data are subjected to feature extraction and vector representation E1 through a Bert encoding module, the image data are subjected to feature extraction and vector representation E2 through a ResNet-LSTM encoding module as shown in Figure 3 , and the time series data are subjected to feature extraction and vector representation E3 through an LSTM encoding module. The feature representations are fused, the feature vectors are spliced according to the 2nd dimension to obtain a feature vector W = E1 + E2 + E3, then the fused feature vector W and a feature representation E4 (the chip anti-radiation performance index "anti-single particle LET threshold greater than 75 MeV.cm2 / mg" is obtained by a Linear encoding module as shown in Figure 4 , and the cosine similarity is calculated, that is, an association degree S is obtained, and is saved to a list L;

[0087] 3. The above process is repeated until all combinations of association degrees are obtained;

[0088] 4. As shown in Figure 7 , the list L is sorted according to the association degree value from high to low as shown in Figure 8 , and the corresponding content is displayed in a visualization display module, wherein the highest association degree value is the optimal combination, which is the matched DC / DC converter.​

Claims

1. A chip anti-radiation performance parameter selection system based on multi-modal data, characterized in that Comprise: Parameter acquisition module: for collecting different types of anti-radiation chip anti-radiation multi-modal data, and storing in multi-modal database according to different data structure, anti-radiation multi-modal data includes but not limited to chip model, packaging process, radiation dose, working temperature and failure rate; Chip anti-radiation multi-modal performance parameter correlation discovery module: for integrating and correlating the anti-radiation multi-modal data of anti-radiation chips obtained from different sources to construct the chip anti-radiation multi-modal data correlation, the anti-radiation multi-modal data of anti-radiation chips including chip anti-radiation performance specification text, anti-radiation performance experimental result curve data or chip circuit structure diagram; Chip anti-radiation performance parameter selection module: for matching the chip demand conditions input by the research and development personnel with the multi-modal database, transmitting the matched multi-modal performance parameters to the chip anti-radiation multi-modal performance parameter correlation discovery module for processing, and automatically selecting the suitable chip anti-radiation performance parameters and corresponding chip model for the research and development personnel to select according to the multi-modal data collected by the chip parameter acquisition module and the analysis processing results of the chip anti-radiation multi-modal performance parameter correlation discovery module; Visual display module: for presenting the chip model and its anti-radiation performance parameter results obtained by the chip anti-radiation performance parameter selection module to the research and development personnel, and the research and development personnel can filter and sort the results according to their needs; In the chip anti-radiation performance parameter selection module, the research and development personnel describe the demand for the chip, and the system extracts the key information according to the input text to obtain the chip demand condition information chip anti-radiation performance index LET or ionization total dose effect; The specific process of the chip anti-radiation multi-modal performance parameter correlation discovery module to construct the chip anti-radiation multi-modal data correlation is as follows: (1.1) Integrate and correlate the anti-radiation multi-modal data of anti-radiation chips obtained from different sources, and randomly extract 1 item from each modal anti-radiation performance parameter of the matched chip to form a multi-modal anti-radiation performance parameter combination; (1.2) Different deep learning methods are used for feature representation of different modal data in the combination, that is, feature vectors are extracted; For anti-radiation performance experimental result curve data, as time series data, feature representation is performed through LSTM encoding module E1 = LSTM(M_waveform), the specific process is as follows: Firstly, normalize the time series data input to the LSTM module to scale the data to the range of 0-1; Then, form an array with shape (n_samples, look_back, n_features), where n_samples is the number of samples, look_back parameter represents the number of consecutive time points used as input, set look_back = 10, and n_features is the number of features at each time point; Finally, the generated array data is processed through an LSTM layer and a fully connected layer to output a feature vector, where the input size of the LSTM layer is the same as the number of features of the data, the size of the hidden layer is 512, and the fully connected layer converts the output of the LSTM layer into a feature vector E1 with a size of (1, 1000); For the chip circuit structure diagram, as image data, it is represented by the ResNet-LSTM encoding module as E2 = ResNet-LSTM (M_image), and the specific process is as follows: First, the image is adjusted to an RGB image 3 after pre-processing 224 224; Then, load the pre-trained ResNet50 model and remove the last fully connected layer and average pooling layer; Again, flatten the output of the ResNet50 model into a vector (1, 1, 100352) and input it into a fully connected layer, which converts it into a vector (1, 1, 512) with the required size; continue to input the transformed vector into an LSTM layer, which processes it as a sequence of time steps and outputs a final LSTM state; Finally, input the LSTM state into a fully connected layer to output a feature vector E2 with a size of (1, 512); For the chip anti-radiation performance specification text, as text data, it is represented by the Bert encoding module as E3 = Bert(M_text), and the specific process is as follows: First, convert the input text into digital form, i.e., use the WordPiece tokenizer to tokenize the input text and map each word to the corresponding number in the BERT model's vocabulary (vocab_size= 21128); Then, according to the position encoding formula: With The position encoding of the word segmentation of the text is carried out, wherein i represents the index of the position encoding, j represents the index of the hidden unit, hidden_size is the number of hidden units in the BERT model, that is, hidden_size=768; Again, input the above encoding sequence into the BERT model for forward propagation, and through 12 layers of self-attention mechanism, fully connected and pooling operation, it is converted into a series of vectors, where each vector represents a token in the input text; Finally, average these vectors to output a feature vector E3 with a size of (1, 768); For the given chip anti-radiation performance index LET, it is represented by the Linear encoding module as E4 = Linear(index), and the specific process is as follows: First, given a LET value as input, set the index to be divided: and the width of the bin: ; then, assign the LET to a bin, subtract the lower limit of the LET range from the LET value, and divide by the width of the bin to get the index of the bin where the value is located: If the calculated index is out of the range of 0 to ( ), it needs to be limited within this range. Again, through a loop, the center value of each bin is formed into a discrete vector with a length of 10; Finally, perform feature extraction through a fully connected layer linear transformation operation to output a feature vector E4 with a size of (1, 30); (1.3) The eigenvectors of each modal performance parameter are subjected to linear transformation for dimension conversion, so that each performance index is represented in the same eigenvector space. The dimension of linear transformation L1 is (1000, 10), the dimension of linear transformation L2 is (512, 10), and the dimension of linear transformation L3 is (768, 10). The converted eigenvectors are E'1 = E1 L1, E'2 = E2 L2, E'3 = E3 L3, and their sizes are all (1, 10); (1.4) Fuse the feature representations E'1, E'2, E'3 in the same vector space using the 2nd dimension concatenation method to obtain the multi-modal data common feature representation W = [E'1, E'2, E'3] with a size of (1, 30); (1.5) Calculate the correlation degree of the feature representation W in the same feature vector space and the feature vector E4 by using the cosine similarity method, and save the result to the list; let the cosine similarity function be similarity(x, y) = (x·y) / ( x · y ), then the correlation degree of W and E4 is S = similarity(W, E4), and S is added to the list; Repeat steps 1.1-1.5 to obtain a list of correlation degrees for multiple multi-modal data combinations, denoted as {S1, S2,..Sx}.

2. The multi-modal data based chip radiation hardness performance parameter selection system of claim 1, wherein: The chip anti-radiation multi-modal performance parameter correlation discovery module analyzes the chip anti-radiation multi-modal data through deep learning technology to discover the correlation between the anti-radiation performance parameters.

3. The multi-modal data based chip radiation performance parameter selection system of claim 1, wherein The automatic selection process of the chip anti-radiation performance parameter selection module is: (2.1) According to the chip demand input by the research and development personnel, entity matching is performed in the multi-modal database to obtain chip anti-radiation multi-modal performance parameters; (2.2) The chip anti-radiation multi-modal data performance parameters are combined and feature fused by using the chip anti-radiation multi-modal performance parameter association relationship discovery module, and the feature similarity, i.e. data correlation degree S, with the chip anti-radiation performance index LET or ionizing total dose effect is calculated; (2.3) The correlation degree list {S1, S2,..Sx} is combined, and the chips are sorted according to the correlation degree, to obtain a chip list with higher correlation degree, and the chip with higher correlation degree has better anti-radiation performance, which is taken as the reference object for selection.

4. The multi-modal data based chip radiation performance parameter selection system of claim 1, wherein: The data acquisition module acquires anti-radiation multi-modal data of different types of anti-radiation chips by manual uploading, website crawling and data extraction. Among them, the manual uploading is the relevant parameters and experimental results of the chip tested by the research and development personnel; the website crawling is to regularly collect chip-related data from the specified website by using the crawling tool embedded in the parameter acquisition module; the data extraction is to extract chip-related data from the relevant anti-radiation chip data system.

5. The multi-modal data based chip radiation performance parameter selection system of claim 1, wherein: The visualization display module displays the relevant data related to the chip with higher chip anti-radiation performance parameter correlation degree in the form of text, list, line chart or column chart.

Images