AI chip performance evaluation method for astronavigation in field of astronavigation communication

Abstract

The invention discloses an astronavigation AI chip performance evaluation method oriented to the field of astronavigation communication. The method comprises the steps that a neural network CABNet is designed and trained; the neural network CABNet comprises three modules which all use combination operation, and the first module comprises convolution and batch normalization; the second module comprises two residual channel modules; the third module comprises a convolution linear unit and a rectification linear unit; and evaluating the AI chip through the neural network CABNet to obtain an AI chip performance evaluation index. According to the method, the problem that a performance index evaluation method of the AI chip for space navigation in a typical communication scene is lacked is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of integrated circuit evaluation and relates to a method for evaluating the performance of an aerospace AI chip for the aerospace communication field. Background Art

[0002] With the vigorous development of artificial intelligence technology, more and more scenarios are using AI algorithms in the aerospace field. With the rapid development of deep learning network models, the differences between models for different scenarios are becoming increasingly larger. From convolutional neural networks to Transformers, the structure of network models is gradually becoming more complex and diversified. Most of these algorithms require high-quality data. The capabilities of hardware are often the bottleneck for such large data volumes and high computing tasks. The increasing attention to deep learning has greatly stimulated the design of smart chips, and major companies have proposed corresponding hardware design solutions. With more and more AI chips in aerospace, how to evaluate the AI ​​performance of these chips based on aerospace applications has become a research hotspot.

[0003] At present, one of the fields where aerospace uses AI chips the most is the communication field, such as channel estimation, encoding and decoding and other tasks. There is no channel estimation algorithm in the existing evaluation benchmark programs, and none of them are oriented to aerospace applications. The complex and diverse structure of network models leads to increasing differences in operators between models. Therefore, the existing general evaluation programs cannot be used to evaluate the performance of AI chips in typical aerospace application scenarios. In addition, there is no unified standard and process for the performance evaluation of AI chips at present, so there is an urgent need to solidify a set of AI chip evaluation methods suitable for the field of aerospace communications. Summary of the invention

[0004] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art, propose an AI performance evaluation method for the field of aerospace communications, and solve the problem of lack of performance indicator evaluation methods in typical communication scenarios of AI chips for aerospace.

[0005] The technical solution of the present invention is: a method for evaluating the performance of an aerospace AI chip for the aerospace communication field, comprising:

[0006] Design and train a neural network CABNet; the neural network CABNet includes three modules, all of which use combined operations, the first module includes convolution and batch normalization; the second module includes two residual channel modules; the third module includes convolution and rectified linear units;

[0007] The AI ​​chip is evaluated through the neural network CABNet to obtain the AI ​​chip performance evaluation indicators.

[0008] Each residual channel module integrates a channel attention mechanism to focus on the features most relevant to channel estimation.

[0009] Each residual channel module contains a combination of three convolutions, batch normalization, and rectified linear units, as well as a global average pooling layer and a combination of two layers of convolutions and rectified linear units.

[0010] The 5G NTN-TDL channel model is used to generate channel data, and the neural network CABNet is trained using the channel data.

[0011] When training the neural network CABNet, stochastic gradient descent is used as the optimizer, and the initial learning rate is set to 0.1. Every 1 / 3 of the total iterations, the learning rate is reduced to 1 / 10 of its original value.

[0012] When training the neural network CABNet, in order to prevent overfitting, L2 regularization is added, and the Momentum strategy with a momentum of 0.9 is used to accelerate convergence.

[0013] The evaluation indicators include operator support, AI application performance and AI processing capability.

[0014] The AI ​​chip is evaluated by the neural network CABNet to obtain the AI ​​chip performance evaluation indicators, including:

[0015] The neural network CABNet is trained by simulating channel data sets to obtain the NMSE index of the floating-point model, the floating-point model is quantified to obtain the real-point model, the real-point model is deployed on the hardware board, and the NMSE index under the measured model is obtained by testing on the hardware board. The measured model is compared with the NMSE index of the floating-point model to obtain the evaluation indicators of the AI ​​chip, including application performance and operator support; finally, the computing power of the AI ​​chip when performing tasks under the real-point model of the neural network CABNet is calculated to obtain the AI ​​processing capability index.

[0016] The advantages of the present invention compared with the prior art are:

[0017] 1) The present invention designs a CABNet neural network model for aerospace communication scenarios. When the AI ​​chip is run under this model, the AI ​​performance index results close to the actual aerospace application conditions are obtained, which is more helpful for users to select and use.

[0018] 2) The present invention applies the CABNet neural network model. CABNet contains operators in the deep learning model commonly used in the field of aerospace communications, so the operator support of the AI ​​chip can be analyzed through the evaluation results.

[0019] 3) The present invention applies the CABNet neural network model. By evaluating under this neural network, the AI ​​application performance indicators of the AI ​​chip in typical communication scenarios can be obtained.

[0020] 4) The present invention applies the CABNet neural network model. By using the neural network for evaluation, the processing capability indicators of AI chips in aerospace communication scenarios can be obtained, such as actual throughput, computing power, and energy efficiency indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The figure is a flow chart of the method of the present invention.

[0022] Figure 2 Evaluate neural network models for aerospace AI chips based on CABNet. DETAILED DESCRIPTION

[0023] The present invention designs an AI performance evaluation method for the field of aerospace communications. Aiming at channel estimation tasks, the performance evaluation method for AI chips is mainly designed from two aspects.

[0024] First, design a neural network model that can include typical aerospace application scenarios and typical aerospace application operators.

[0025] Second, design AI chip performance evaluation indicators, including operator support, AI application performance and AI processing capabilities.

[0026] The present invention is an AI performance evaluation method for the field of aerospace communications. The specific method flow is shown in the accompanying drawings. Figure 1 .

[0027] First, a neural network named CABNet (Channel Attention Based Network, CABNet) was designed. The floating-point model was trained by simulating channel data sets, and the NMSE index under the channel estimation task was obtained. Then the floating-point model was quantized to obtain a fixed-point model, which was deployed on the hardware board. The NMSE index under the channel estimation task was obtained through the test results on the hardware board. The results obtained by the floating-point model and the fixed-point model were compared to obtain the application performance of the AI ​​chip, operator support and other evaluation results. Finally, the AI ​​processing capability index was obtained by calculating the computing power when executing tasks under the CABNet model.

[0028] The neural network model in this invention is named CABNet, which is a neural network model based on the channel attention mechanism. Figure 2As shown in the figure, the channel information features are selectively enhanced by combining the channel attention mechanism. Using the channel attention mechanism to perform channel estimation is a method often used in the research of channel estimation intelligent algorithms in the field of aerospace communications. The input of the model is the channel estimated by the classical least squares method, and the noisy channel after the real part and the imaginary part are separated. The noisy channel is input into the neural network model CABNet, and the model outputs a channel estimation result that is closer to the true value.

[0029] The CABNet network model consists of three modules, all of which use combined operations. The first module includes convolution (Conv) and batch normalization (BN), and the third module includes convolution (Conv) and rectified linear unit (ReLU). The first module uses convolution and batch normalization to extract feature information, greatly improve the model training speed, and enhance the stability of the model under various data conditions. The third module uses convolution and linear rectifier units to introduce nonlinear capabilities, allowing the network to learn complex patterns and features, ensuring the expressiveness and resolution of the input features and providing high-quality output. Among them, 64×(3×3×2) means that the kernel is 3×3, the input channel is 2, and the output channel is 64.

[0030] The second module is the residual group, which includes the core components of two CABNet networks - residual channel modules (RB). RB integrates the channel attention mechanism to focus on the features most relevant to channel estimation. Each RB contains a combination of 3 Conv, BN and ReLU layers, a global average pooling layer (Global Pooling) and a combination of two layers of Conv and ReLU. RBs are stacked to form residual groups, which enhances the feature learning ability of the model. The residual learning framework can alleviate the vanishing gradient problem and enable deeper networks to be effectively trained. The global average pooling operation can compress the spatial dimension of the input feature map, emphasize important features and reduce irrelevant features.

[0031] Network training uses the 5G NTN-TDL channel model to generate channel data. First, the least square method is used to obtain the preliminary estimated channel information by receiving data Y and sending data X. Will As the network input, CABNet obtains the estimated data closer to the true channel value H The training process includes: using stochastic gradient descent (SGD) as the optimizer, the initial learning rate is set to 0.1, and the learning rate is reduced to 1 / 10 of its original value every 1 / 3 of the total number of iterations. To prevent overfitting, L2 regularization is added, and the Momentum strategy with a momentum of 0.9 is used to accelerate convergence.

[0032] The types of operators included in the CABNet model are listed in Table 1.

[0033] Table 1 Operator types included in the CABNet model

[0034]

[0035] The present invention is an AI performance evaluation method for the field of aerospace communications. The evaluation process is shown in Figure 1 .

[0036] First, the CABNet model is trained, tested, and verified using a data set. The data set used can be a simulation data set or measured channel data in a space environment. The data set used in this patent is a simulation data set generated based on the 5GNTN-TDL channel model. After completing the training of the CABNet model, the NMSE indicator of the floating-point model is obtained.

[0037] NMSE is a commonly used indicator in channel estimation tasks. The calculation method of NMSE indicator is as follows:

[0038]

[0039] The channel information H is the real channel information, In order to obtain an estimate closer to the actual channel information through CABNet, ||·|| is the binary norm and E(·) is the mathematical expectation.

[0040] After obtaining the floating-point model, deploy the model to the hardware board of the AI ​​chip to be evaluated. In this patent, the Yulong810A board is used for deployment. The CABNet model is generated through the PyTorch deep learning training framework, so the network model trained by CABNet is imported into YL-ACUITY through ONNX conversion, and the network model is quantized to 8 bits. Through the export function, C language code based on ovxlib is generated.

[0041] The test was conducted on the board to obtain the NMSE index results under the measured model, and the results of the measured model were compared with those of the floating-point model. The operator support and application performance of the AI ​​chip in the aerospace channel estimation task were analyzed, and finally the computing power of the chip when performing tasks under the CABNet model was calculated. Table 2 shows the NMSE index evaluation results of the AI ​​chip application performance under different compression rates of different network models. Table 3 shows the processing power index results of the AI ​​chip when performing tasks under the CABNet model.

[0042] Table 2 NMSE index evaluation results

[0043]

[0044] Table 3 Comparison of evaluation hardware indicators

[0045]

[0046] As can be seen from Table 2, the NMSE index of the CABNet model is higher than that of other network models, and the estimated channel information is more accurate. It also shows that the AI ​​chip used supports all operators in the CABNet model and can support more commonly used operators in aerospace channel estimation tasks. Comparing the results of the floating point and measured models, the 8-bit quantization accuracy loss of the AI ​​chip is small, and the computational complexity can be reduced by deploying the quantized model on the chip.

[0047] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solution of the present invention by using the technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for evaluating the performance of an aerospace AI chip for the field of aerospace communications, characterized in that: include: Design and train a neural network CABNet; the neural network CABNet includes three modules, all of which use combined operations, the first module includes convolution and batch normalization; the second module includes two residual channel modules; the third module includes convolution and rectified linear units; The AI ​​chip is evaluated through the neural network CABNet to obtain the AI ​​chip performance evaluation indicators.

2. According to claim 1, a method for evaluating the performance of an aerospace AI chip for the aerospace communication field is characterized in that: Each residual channel module integrates a channel attention mechanism to focus on the features most relevant to channel estimation.

3. According to claim 2, a method for evaluating the performance of an aerospace AI chip for the aerospace communication field is characterized in that: Each residual channel module contains a combination of three convolutions, batch normalization, and rectified linear units, as well as a global average pooling layer and a combination of two layers of convolutions and rectified linear units.

4. According to claim 3, a method for evaluating the performance of an aerospace AI chip for the aerospace communication field is characterized in that: The 5G NTN-TDL channel model is used to generate channel data, and the neural network CABNet is trained using the channel data.

5. According to claim 4, a method for evaluating the performance of an aerospace AI chip for the aerospace communication field is characterized in that: When training the neural network CABNet, stochastic gradient descent is used as the optimizer, and the initial learning rate is set to 0.

1. Every 1 / 3 of the total iterations, the learning rate is reduced to 1 / 10 of its original value.

6. The method for evaluating the performance of an aerospace AI chip for the aerospace communication field according to claim 5, characterized in that: When training the neural network CABNet, in order to prevent overfitting, L2 regularization is added, and the Momentum strategy with a momentum of 0.9 is used to accelerate convergence.

7. The method for evaluating the performance of an aerospace AI chip for the aerospace communication field according to claim 1, characterized in that: The evaluation indicators include operator support, AI application performance and AI processing capability.

8. The method for evaluating the performance of an aerospace AI chip for the aerospace communication field according to claim 4, characterized in that: The AI ​​chip is evaluated by the neural network CABNet to obtain the AI ​​chip performance evaluation indicators, including: The neural network CABNet is trained by simulating channel data sets to obtain the NMSE index of the floating-point model, the floating-point model is quantified to obtain the real-point model, the real-point model is deployed on the hardware board, and the NMSE index under the measured model is obtained by testing on the hardware board. The measured model is compared with the NMSE index of the floating-point model to obtain the evaluation indicators of the AI ​​chip, including application performance and operator support; finally, the computing power of the AI ​​chip when performing tasks under the real-point model of the neural network CABNet is calculated to obtain the AI ​​processing capability index.

Images