CN-121659805-B - Lightweight construction method of multi-task AI prediction model in low-altitude economic aviation field

Abstract

The invention relates to the technical field of artificial intelligence model optimization, in particular to a lightweight construction method of a multi-task AI prediction model in the low-altitude economic aviation field, which comprises the steps of obtaining historical execution frequency duty ratios of a plurality of subtasks in the model, and calculating complexity entropy values of the subtasks based on the frequency duty ratios; aiming at parameters to be compressed of the model, calculating importance scores of the parameters to be compressed by combining static weight amplitude and dynamic gradient sensitivity, generating a dynamic compression threshold by combining entropy values and the importance scores, and carrying out structured pruning on the parameters to be compressed with the importance scores lower than the dynamic compression threshold from the model to obtain the lightweight model. The method solves the technical problems of resource waste in a simple task scene and precision damage in a complex task scene caused by the fact that the fixed compression threshold value is adopted and the complexity of the task cannot be adapted to real-time change.

Inventors

• GAO KAI
• XIE KUNMING
• HAN ZHENTAO
• WANG XINGXING
• LI SONGBAI
• Xu Tinghai
• WU HONGJUN
• LI RUI
• DENG XIANFENG
• DENG WENXI
• ZHANG XING
• SUN ZIXUAN
• PENG CHENG
• LUO YUBO
• WANG YONGQIANG
• PENG SEN
• LI YONG
• LIU YI
• HU MIN
• XIE YUN

Assignees

• 成都九洲电子信息系统股份有限公司

Dates

Publication Date

20260508

Application Date

20260206

Claims (6)

1. 1. The light-weight construction method of the multi-task AI prediction model in the low-altitude economic aviation field is characterized by comprising the following steps of: acquiring historical execution frequency duty ratios of a plurality of subtasks in the model, and calculating the complexity entropy value of each subtask based on the frequency duty ratios; aiming at parameters to be compressed of the model, calculating importance scores of the parameters to be compressed by combining static weight amplitude and dynamic gradient sensitivity; Combining the entropy value and the importance score to generate a dynamic compression threshold; Carrying out structured pruning on the parameters to be compressed, the importance scores of which are lower than the dynamic compression threshold value, from the model so as to obtain a lightweight model; The method for calculating the importance scores of the parameters to be compressed comprises the following steps: wherein, the method comprises the steps of, Is a parameter Importance score of (2); Is a parameter Is characterized by the L2 norm square of the parameter to be compressed Is a static weight magnitude of (2); For the loss function L pair parameter Characterizing the parameters to be compressed Alpha and beta are weight coefficients; The dynamic compression threshold is generated by multiplying the entropy value, a preset compression strength coefficient and the median of importance scores of all parameters to be compressed.
2. 2. The method for lightweight construction of a low-altitude economic aviation field multi-task AI prediction model according to claim 1, wherein the preset compression strength coefficient is set to be in a negative correlation with the entropy value.
3. 3. The method for lightweight construction of a low-altitude, economic and aviation domain multi-task AI prediction model as set forth in claim 1, wherein the plurality of subtasks are divided according to a time span of space material demand prediction, including short-term demand prediction, medium-term demand prediction and long-term demand prediction.
4. 4. The method for lightweight construction of a low-altitude, economic and aviation domain multi-task AI prediction model as set forth in claim 1, wherein the complexity entropy value is as follows The method of (1) comprises: wherein, the method comprises the steps of, And the execution frequency of the i-th subtask is the execution frequency duty ratio, and n is the total class number of the subtasks.
5. 5. The method of claim 1, wherein when the model comprises a long-short term memory network layer, the structured pruning comprises removing an entire row or column of a long-short term memory network layer weight matrix.
6. 6. The method for lightweight construction of a low-altitude, economic and aviation domain multi-task AI prediction model as set forth in claim 1, wherein said structured pruning comprises deleting low-importance attention headers corresponding to said parameters to be compressed when said model comprises a transducer layer.

Description

Lightweight construction method of multi-task AI prediction model in low-altitude economic aviation field Technical Field The invention relates to the technical field of artificial intelligence model optimization, in particular to a lightweight construction method of a multi-task AI prediction model in the field of low-altitude economy aviation. Background In a multi-task AI prediction scenario in the low-altitude economic aviation field, for example, real-time decisions involving flight path planning, air traffic management or multi-task AI prediction models, task complexity has the characteristic of high dynamic variation. For example, the prediction task may cover both short-term prediction (simpler mode) that relies on an incident and long-term prediction (significantly more complex) that requires capture of long-term trends. These subtasks for different time spans are quite different in terms of the model's ability to be required. The traditional static compression strategy (i.e. the "one-tool" mode with fixed threshold) cannot adapt to the dynamics, and has fundamental contradiction with the task requirements: 1. Waste of resources and unnecessary delay if a low compression rate is set to cope with the most complex tasks, the model is excessively heavy when a large number of simple tasks are processed, resulting in continuous waste of resources and unnecessary delay. 2. Accuracy loss and safety hazards if a high compression rate is set for extremely light weight, critical information may be lost due to excessive compression when the model processes a complex task, and prediction accuracy may be greatly reduced. To overcome the limitations of static strategies, the prior art proposes a dynamic or adaptive compression method, such as chinese patent application CN117882376 a. This patent discloses an "example adaptive image and video compression in a network parameter subspace using a machine learning system" technique. The core scheme is that a low-dimensional parameter subspace is constructed through methods such as Principal Component Analysis (PCA) and the like in a model training stage. In the inference phase, the system will find an optimal set of coordinates in the subspace for each input specific data instance (per-instance), generating an updated set of model parameters tailored to that specific instance. The system then transmits the compressed data to the decoding side along with the optimal set of "subspace coordinates". However, when the technical scheme is applied to a high-real-time and high-throughput multi-task scene such as the low-altitude economic aviation field, the following disadvantages still exist: 1. the fine granularity of adaptation and the high computational overhead require an optimization calculation for each individual input instance to determine its best coordinates in the parameter subspace. This "case-by-case optimization" overhead (requiring a significant amount of computation) may offset the inferred speed gain from compression, resulting in overall system delay that does not degrade and rises, making it difficult to meet the stringent requirements for low delay (e.g., less than 100 ms). 2. The method is adaptive to the microscopic features of a single data instance, and the sensing of the macroscopic statistical characteristics of the whole task flow is lacked. The method can not actively and prospectively adjust the overall compression strategy of the model from a macroscopic level, and can not identify that the model is in a simple task mode, thereby causing resource waste. 3. The response to the task distribution change is lagged, namely, a core of the response is a fixed subspace constructed based on training data, and when the task distribution in practical application has significant drift, the subspace is possibly not optimal any more, and the self-adaptive effect is reduced. Therefore, a brand new technical scheme is urgently needed in the field, the overall complexity of the task flow can be quantized accurately in real time, and the compression strength of the model can be dynamically and efficiently adjusted based on the overall complexity, so that the calculation and storage resources are saved to the greatest extent while the requirements of high real-time performance and high precision are met. Disclosure of Invention The invention aims to overcome the defects in the prior art and provide a light-weight construction method and a corresponding system for a multi-task AI prediction model in the field of low-altitude economy aviation. Specifically, the invention aims to solve the technical problems that the prior model compression strategy, in particular to the static compression strategy, cannot adapt to real-time change of task complexity due to the adoption of a fixed compression threshold, so that resources are wasted in a simple task scene and the precision is damaged in a complex task scene. Meanwhile, the invention aims to solve the technical proble