An agent-driven sensor structure design optimization method and apparatus
By using an agent-driven sensor structure design optimization method, sensor structure parameters are automatically generated using large language models and deep learning models. This solves the problem of cumbersome and time-consuming traditional design methods and achieves efficient and low-cost sensor design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI INST OF MICROSYSTEM & INFORMATION TECH CHINESE ACAD OF SCI
- Filing Date
- 2026-01-20
- Publication Date
- 2026-06-09
AI Technical Summary
Traditional sensor structure design methods are cumbersome and time-consuming, relying on theoretical calculations and simulations, making it difficult to fully explore design options, resulting in long R&D cycles and high costs.
A sensor structure design optimization method driven by intelligent agents is adopted. It utilizes a large language model and a deep learning model, and automatically generates sensor structure parameters through a multi-level attention mechanism and a Transformer architecture. The design results are then optimized by combining the L-BFGS-B optimization algorithm.
It significantly improves sensor design efficiency, reduces time and labor costs, and enables the automation and human-computer interaction of sensor design, allowing users to quickly optimize designs without needing to master complex professional knowledge.
Smart Images
Figure CN122174600A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and intelligent sensor chip design, and in particular to a method and apparatus for optimizing sensor structure design driven by an intelligent agent. Background Technology
[0002] Gas sensors hold significant importance in the fields of new energy and health and safety. For example, in the production, storage, and transportation of clean energy sources such as hydrogen, they can accurately detect leaks, provide timely warnings to avoid dangers such as explosions, ensure the safe and stable operation of energy systems, and facilitate the large-scale application of new energy. In terms of health and safety, the concentrations of pollutants in the environment, such as sulfur dioxide and hydrogen sulfide, provide data support for air quality assessment and pollution control. Exhaled breath components (such as hydrogen sulfide), as novel gaseous signaling molecules, are of great significance in disease diagnosis. Changes in their levels can assist in the diagnosis of airway inflammatory diseases such as asthma and COPD, reflecting the disease control status and the risk of acute exacerbations. They can also provide clues for the diagnosis of intestinal diseases, kidney function assessment, and oral diseases such as halitosis and periodontitis.
[0003] The development of high-performance sensors is inseparable from sensor design optimization. Currently, in the field of gas sensor research and development, traditional structural design methods have long faced numerous challenges.
[0004] 1) The complex structure of sensors, relying on theoretical calculations and simulations for parameter scanning, is tedious and time-consuming, severely restricting R&D efficiency and innovation capabilities. Traditional design approaches often begin with a deep understanding of the sensor's working principles, followed by predicting sensor performance through theoretical models or simulations. Taking MEMS gas sensors as an example, the sensor structure typically includes a support layer, heating layer, isolation layer, and sensing layer from bottom to top, each with different structural dimensions (length, width, thickness, etc.), resulting in more than ten different dimensions and material parameters. Theoretical calculations require engineers to have a strong theoretical foundation and the ability to accurately set simulation parameters to cover as much of the design space as possible. Furthermore, theoretical calculation equations are often simplifications of complex real-world scenarios, making them prone to design errors. Even with more realistic simulations (such as multiphysics coupled finite element simulations) that can iterate through various parameters multiple times, it remains a tedious and time-consuming process, typically requiring tens of hours or even days of computation. And if the calculation and simulation results do not meet user requirements, recalculation is necessary.
[0005] 2) The selection and combination of parameters are often based on experience or limited datasets, meaning the design process struggles to fully explore all potential design options. More importantly, this reliance on theoretical calculations and simulations often makes it difficult to directly translate results into practical design experience, as each design iteration is an independent event based on specific assumptions and conditions, lacking a systematic mechanism for accumulating experience. Therefore, even after multiple design cycles, engineers struggle to develop a reusable and scalable set of design guidelines or rules, forcing each new product development to start from scratch, repeatedly performing extensive theoretical calculations and simulation verifications, significantly extending the R&D cycle and increasing costs. Summary of the Invention
[0006] The technical problem to be solved by the present invention is to provide an agent-driven sensor structure design optimization method that can improve sensor design efficiency and reduce time and labor costs.
[0007] The technical solution adopted by this invention to solve its technical problem is: to provide a method for optimizing the design of a sensor structure driven by an intelligent agent, comprising the following steps:
[0008] Obtain the design requirements for the sensor;
[0009] The design requirements of the sensor are input into the intelligent agent, which outputs optimized structural parameters of the sensor; wherein, the intelligent agent includes:
[0010] The large language model part is used to extract the sensor's performance targets and structural size constraints from the design requirements;
[0011] The deep learning model part is used to output optimized structural parameters of the sensor based on the sensor's performance objectives and structural size constraints.
[0012] The large language model is based on the Deepseek architecture and uses a multi-level attention mechanism to capture information about the sensor's performance goals in the design requirements.
[0013] The deep learning model includes:
[0014] The encoder section takes sensor structural parameters as input and sensor performance as output.
[0015] The decoder section takes sensor performance as input and sensor structural parameters as output.
[0016] Both the encoder and decoder sections adopt the Transformer architecture.
[0017] The encoder portion includes multiple stacked encoders, each encoder comprising:
[0018] The first encoder sublayer consists of a multi-head self-attention layer, a normalization layer, and a residual connection. The multi-head self-attention layer is used to capture the long-range dependencies and coupling relationships between multiple sensor structural parameters.
[0019] The second encoder sublayer consists of a feedforward neural network layer, a normalization layer, and a residual connection. The feedforward neural network layer is used to perform a deep nonlinear transformation on the coupling features output by the first encoder sublayer to fit the nonlinear relationship between structural parameters and performance.
[0020] The decoder section includes multiple stacked decoders and parameter optimization modules; the decoder includes:
[0021] The first decoder sublayer consists of a multi-head attention layer, a normalization layer, and a residual connection. The multi-head attention layer includes a multi-head self-attention layer and a multi-head cross-attention layer. The multi-head self-attention layer is used to capture the correlation between performance parameters and structural size constraints, as well as the priority relationship among multiple performance parameters. The multi-head cross-attention layer is used to align the performance and structural size constraint features of the decoder with the nonlinear relationship between the structural parameters and performance of the encoder, ensuring that the reverse-generated structural parameters conform to the physical laws learned by the encoder.
[0022] The second decoder sublayer consists of a feedforward neural network layer, a normalization layer, and a residual connection. The feedforward neural network layer is used to perform a nonlinear transformation on the positive alignment features output by the second decoder sublayer to fit the inverse nonlinear relationship between performance parameters and structural size parameters.
[0023] The parameter optimization module is used to optimize the output structure size parameters of the decoder.
[0024] The parameter optimization module uses the L-BFGS-B optimization algorithm to optimize the output structure size parameters of the decoder.
[0025] The technical solution adopted by this invention to solve its technical problem is: to provide an agent-driven sensor structure design optimization device, comprising:
[0026] The acquisition module is used to acquire the design requirements of the sensor;
[0027] A design module is used to input the design requirements of the sensor into an intelligent agent to obtain optimized structural parameters of the sensor; wherein, the intelligent agent includes:
[0028] The large language model part is used to extract the performance targets and structural size constraints of the sensor from the design requirements;
[0029] The deep learning model part is used to output optimized structural parameters of the sensor based on the sensor's performance objectives and structural size constraints.
[0030] The technical solution adopted by the present invention to solve its technical problem is: to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned intelligent agent-driven sensor structure design optimization method.
[0031] The technical solution adopted by the present invention to solve its technical problem is: to provide a computer-readable storage medium on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the above-mentioned intelligent agent-driven sensor structure design optimization method are implemented.
[0032] Beneficial effects
[0033] By adopting the above-mentioned technical solution, this invention has the following advantages and positive effects compared with the prior art: The intelligent agent constructed by this invention can directly provide the structural parameters of the sensor according to the user's needs, avoiding tedious parameter scanning and trial and error, significantly improving the efficiency of sensor design, reducing time and manpower costs, realizing the automation of sensor design and human-computer interaction, allowing users to quickly design and optimize sensors without mastering complex professional model knowledge, and integrating fragmented experience into an intelligent agent, which continuously learns and accumulates, updates and iterates itself, and continuously improves the efficiency and performance of sensor R&D. Attached Figure Description
[0034] Figure 1 This is a flowchart of the sensor structure design optimization method driven by the intelligent agent according to the first embodiment of the present invention;
[0035] Figure 2 This is a schematic diagram of the deep learning model part in the first embodiment of the present invention;
[0036] Figure 3 This is a schematic diagram of the training of the deep learning model part in an embodiment of the present invention;
[0037] Figure 4 This is a diagram of the human-computer interaction interface of the large voice model part of the intelligent agent in this embodiment of the invention;
[0038] Figure 5 This is a diagram showing the result of the intelligent agent calling a deep learning model to perform reverse optimization prediction in an embodiment of the present invention. Detailed Implementation
[0039] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.
[0040] The first embodiment of the present invention relates to a method for optimizing the design of a sensor structure driven by an intelligent agent, such as... Figure 1 As shown, it includes the following steps:
[0041] Step 1: Obtain the design requirements for the sensor.
[0042] Step 2: Input the design requirements of the sensor into the intelligent agent and output the structural parameters of the sensor to be optimized.
[0043] The intelligent agent in this embodiment includes a large language model and a deep learning model. The large language model is used to extract the performance objectives and structural size constraints of the sensor from the design requirements; the deep learning model is used to output optimized structural parameters of the sensor based on the performance objectives and structural size constraints.
[0044] The large language model in this embodiment can be based on the Deepseek architecture. Compared to other large language models, the Deepseek architecture is based on the Transformer architecture, accurately capturing key information through a multi-level attention mechanism. It employs sparse training and mixed-precision computation to reduce computational costs and memory usage, supporting deployment on edge devices. Furthermore, Deepseek's open-source ecosystem and low-cost characteristics lower the technical barrier, and its adaptive learning capabilities enable sensor design optimization to move towards intelligence, high precision, and low cost.
[0045] Because sensors have many structural parameters but relatively few performance parameters, forward modeling, which derives performance parameters from structural parameters, offers high accuracy. However, considering that user needs often involve predicting sensor structural parameters based on target sensor performance, this implementation incorporates a reverse intelligent optimization prediction model within the deep learning model section. This model rapidly filters the structural dimension parameters that meet the user's target performance requirements.
[0046] like Figure 2As shown, the deep learning model in this embodiment includes an encoder (i.e., a forward prediction model) and a decoder (i.e., a backward intelligent optimization prediction model). The encoder takes sensor structural parameters as input and sensor performance as output; the decoder takes sensor performance as input and sensor structural parameters as output. Both the encoder and decoder employ the Transformer architecture. Compared to traditional multilayer perceptrons (MLPs) and artificial neural networks (ANNs), the Transformer architecture's self-attention mechanism can process sequential data in parallel, overcoming the temporal limitations of RNNs and significantly improving computational efficiency. Through multi-head attention, it captures long-range dependencies and accurately models the relationship between sensor structural parameters and performance.
[0047] The core objective of the encoder section in this embodiment is to learn the positive correlation between sensor structural parameters and performance parameters, providing physical support for the decoder's reverse prediction, and ensuring that the structural parameters generated in reverse can indeed achieve the target performance in the forward logic. The encoder section in this embodiment includes an input layer, an embedding layer, multiple stacked encoders, and an output layer.
[0048] The input layer receives structural parameters and standardizes them to eliminate interference from differences in parameter dimensions and units. It provides data in a uniform format for subsequent feature extraction, avoiding a decrease in forward mapping accuracy due to inconsistent parameter scales.
[0049] Embedding layers are used to transform discrete / numerical structural parameters into high-dimensional feature vectors (embedding vectors), allowing the model to capture the potential relationships between parameters (rather than just focusing on the numerical value of a single parameter).
[0050] The encoder includes:
[0051] The first encoder sublayer consists of a multi-head self-attention layer, a normalization layer, and a residual connection. The multi-head self-attention layer captures the long-range dependencies and coupling relationships between multiple sensor structural parameters. The model calculates the correlation weights between each structural parameter through the self-attention layer; higher weights indicate a stronger coupling effect on performance, while also capturing the synergistic effects between structural parameters. Multiple independent attention heads can simultaneously focus on coupling relationships in different dimensions, thus achieving comprehensive modeling of the coupling relationships between structural parameters. The normalization layer normalizes the output features of the multi-head self-attention layer (unifying the mean and variance of the data), avoiding gradient vanishing or parameter oscillations during training.
[0052] The second encoder sublayer consists of a feedforward neural network layer, a normalization layer, and a residual connection. The feedforward neural network layer is used to perform a deep nonlinear transformation on the coupling features output by the first encoder sublayer to fit the nonlinear relationship between structural parameters and performance. In this way, the encoder can accurately learn the mapping law between the combination of structural parameters and the numerical values of performance parameters, providing positive knowledge reference for the subsequent decoder.
[0053] The output layer is used to convert the output of the second encoder sublayer into specific sensor performance parameter values.
[0054] The goal of the decoder section in this embodiment is to generate the optimal combination / range of structural parameters that meets the performance requirements and constraints based on the user's target performance parameters and structural size constraints. It relies on the forward patterns learned by the encoder to avoid the reverse prediction deviating from the actual physical logic. The decoder section in this embodiment includes an input layer, an embedding layer, multiple stacked decoders, a parameter optimization module, and an output layer.
[0055] The input layer receives the target performance and structural size constraints and performs standardization processing. The standardization processing includes normalizing the target performance parameters and converting the structural size constraints into boundary marker vectors. This allows the model to be aware of the legal range of the parameters and avoids generating invalid structural parameters.
[0056] The embedding layer is used to vectorize the target performance, that is, to transform the target performance parameters and structural size constraints into high-dimensional feature vectors, so that the model can capture the relationship between the performance target and the constraints.
[0057] The decoder includes:
[0058] The first decoder sublayer consists of a multi-head attention layer, a normalization layer, and a residual connection. The multi-head attention layer includes a multi-head self-attention layer and a multi-head cross-attention layer. The multi-head self-attention layer captures the correlation between performance parameters and structural size constraints, as well as the priority relationships among multiple performance parameters. This allows the model to understand how to meet performance targets under structural size constraints, providing direction for subsequent association of positive laws. The multi-head cross-attention layer aligns the decoder's performance and structural size constraint features with the nonlinear relationship between the encoder's structural parameters and performance, ensuring that the back-generated structural parameters conform to the physical laws learned by the encoder. This avoids the decoder generating structural parameters that cannot achieve the target performance in forward modeling, ensuring the accuracy of back-prediction (error ±0.1%), and serves as the core bridge connecting forward laws and back-generation. The normalization layer normalizes the features output by the attention layer, preventing parameter oscillations during back-generation and ensuring that the model stably outputs structural parameters that meet the requirements.
[0059] The second decoder sublayer consists of a feedforward neural network layer, a normalization layer, and a residual connection. The feedforward neural network layer is used to perform a nonlinear transformation on the positive alignment features output by the second decoder sublayer to fit the inverse nonlinear relationship between performance parameters and structural size parameters. This allows the decoder to accurately learn the inverse relationship between the target performance and structural parameters, and generate structural parameters that meet the performance requirements.
[0060] The core of the decoder is based on the Transformer architecture, which generates a set of candidate structural parameters that meet the target performance through positive regular alignment. However, there may be three limitations: (1) the output parameters may approach the structural size constraint boundary, making the processing difficult; (2) for multi-target scenarios, it may focus on a single target and ignore the optimal balance of another target; (3) there are small errors.
[0061] To address the aforementioned issues, this embodiment adds a parameter optimization module. This module optimizes the structural dimension parameters output by the decoder. In this embodiment, the parameter optimization module employs the L-BFGS-B optimization algorithm to optimize the decoder's output structural dimension parameters. The input to this module is the initial set of candidate structural parameters generated by the decoder. Its optimization objective is to minimize the error between the performance (calculated value based on forward modeling) corresponding to the structural parameters and the target performance. Using the structural dimension parameters as constraints, during the iteration process, based on the structure-performance forward mapping relationship learned by the encoder, the actual performance corresponding to the candidate parameters is calculated. Through quasi-Newton iterations using L-BFGS-B, the structural parameters are adjusted, and the performance is recalculated until the error is less than a preset value and the parameters are no longer on the constraint boundaries. The output of this parameter optimization module is the optimal range of structural parameter combinations after fine optimization.
[0062] The output layer transforms the finely optimized range of structural parameter combinations into user-usable structural parameter results and performs constraint verification to ensure that the output structural parameters do not conform to structural size constraints. The results from the output layer can be fed back to the large language model part, providing the optimization range for sensor size design through human-computer interaction and question-and-answer format.
[0063] The following example, using the power consumption design optimization of a MEMS semiconductor metal oxide gas sensor, further illustrates this invention. A micro-heating plate is a typical MEMS semiconductor metal oxide gas sensor, comprising a support layer, a heating layer, an isolation layer, and a sensing layer. The electrodes of the sensing layer are coated with metal oxide. The electrodes of the heating layer provide a heating temperature (200°C-400°C) to the semiconductor metal oxide, at which point the adsorbed oxygen on the oxide surface can interact with the measured gas (such as hydrogen sulfide), resulting in electron transfer and a change in the oxide resistance between the sensing electrodes. The magnitude of this change is proportional to the concentration. Therefore, a crucial performance indicator closely related to the structural size of the MEMS gas sensor is heating power consumption. Low-power sensors are also one of the requirements for arrayed, distributed gas detection in real-world scenarios.
[0064] Since there are more than a dozen different structural dimensions for the various layers of the sensor, this embodiment only selects four of the most critical structural dimensions (side length of the heating plate, length of the support arm, thickness of the support layer, and thickness of the isolation layer), heating temperature, and heating power consumption to demonstrate the effectiveness of this implementation method.
[0065] Figure 3 The training process of the deep learning model is demonstrated. Due to time and cost constraints, it is difficult to fabricate tens of thousands of different sensor size combinations in real-world scenarios. Therefore, a feasible approach is to measure the heating plate side length, support arm length, support layer thickness, insulation layer thickness, heating temperature, and heating power consumption using existing real sensors, calibrate the corresponding finite element simulation model, and then use the finite element simulation model to change the above parameter combinations to obtain 10,000 size design results and their corresponding heating temperatures and power consumptions, which are then used as a dataset. 85% of this dataset is selected as the training set, and 15% as the prediction set. After 100 rounds of training, it can be seen that the loss function LOSS has stabilized, and the accuracy of both the training set and the prediction set can reach approximately 99%.
[0066] Figure 4 This embodiment demonstrates the human-computer interaction interface for the large language model portion of the intelligent agent. Here, the user assumes a target sensor heating power consumption of 10mW and a heating temperature of 300℃. The user then provides reasonable ranges for the heating plate side length, support arm length, support layer thickness, and isolation layer thickness within the specified technological parameters. The intelligent agent is then asked to provide optimized structural parameters that meet the requirements. The interface shows that the large speech model understands and extracts key parameter information, and begins to invoke the deep learning model portion. Figure 5 The results of the intelligent agent calling a dedicated deep learning model for backpropagation optimization prediction are shown. It can be seen that, given a user-specified heating power consumption target of 10mW and a heating temperature of 300℃, the deep learning model provides the optimal design range for each parameter and informs the user of the results, with an error of ±0.1%.
[0067] The time required for the intelligent agent to provide a complete answer from receiving the user's design requirements is approximately 7 seconds. Using the same parameter scan (i.e., the previous 10,000 data combinations) and filtering out the results that meet the user's requirements would take 22.2 hours, thus the sensor design efficiency is improved by 11,000 times.
[0068] It is easy to see that the intelligent agent constructed by this invention can directly provide the structural parameters of the sensor according to the user's needs, avoiding tedious parameter scanning and trial and error, significantly improving the efficiency of sensor design, reducing time and manpower costs, realizing the automation of sensor design and human-computer interaction. Users do not need to master complex professional model knowledge to quickly design and optimize sensors. At the same time, fragmented experience is integrated into an intelligent agent, which updates and iterates itself through continuous learning and accumulation, continuously improving the efficiency and performance of sensor R&D.
[0069] The second embodiment of the present invention relates to an agent-driven sensor structure design optimization device, comprising:
[0070] The acquisition module is used to acquire the design requirements of the sensor;
[0071] A design module is used to input the design requirements of the sensor into an intelligent agent and output optimized structural parameters of the sensor; wherein, the intelligent agent includes:
[0072] The large language model part is used to extract the performance targets and structural size constraints of the sensor from the design requirements;
[0073] The deep learning model part is used to output optimized structural parameters of the sensor based on the sensor's performance objectives and structural size constraints.
[0074] The large language model is based on the Deepseek architecture and uses a multi-level attention mechanism to capture information about the sensor's performance goals in the design requirements.
[0075] The deep learning model includes:
[0076] The encoder section takes sensor structural parameters as input and sensor performance as output.
[0077] The decoder section takes sensor performance as input and sensor structural parameters as output.
[0078] Both the encoder and decoder sections adopt the Transformer architecture.
[0079] The encoder portion includes multiple stacked encoders, each encoder comprising:
[0080] The first encoder sublayer consists of a multi-head self-attention layer, a normalization layer, and a residual connection. The multi-head self-attention layer is used to capture the long-range dependencies and coupling relationships between multiple sensor structural parameters.
[0081] The second encoder sublayer consists of a feedforward neural network layer, a normalization layer, and a residual connection. The feedforward neural network layer is used to perform a deep nonlinear transformation on the coupling features output by the first encoder sublayer to fit the nonlinear relationship between structural parameters and performance.
[0082] The decoder section includes multiple stacked decoders and parameter optimization modules; the decoder includes:
[0083] The first decoder sublayer consists of an attention layer, a normalization layer, and a residual connection. The attention layer includes a multi-head self-attention layer and a multi-head cross-attention layer. The multi-head self-attention layer is used to capture the correlation between performance parameters and structural size constraints, as well as the priority relationship among multiple performance parameters. The multi-head cross-attention layer is used to align the performance and structural size constraint features of the decoder with the nonlinear relationship between the structural parameters and performance of the encoder, ensuring that the reverse-generated structural parameters conform to the physical laws learned by the encoder.
[0084] The second decoder sublayer consists of a feedforward neural network layer, a normalization layer, and a residual connection. The feedforward neural network layer is used to perform a nonlinear transformation on the positive alignment features output by the second decoder sublayer to fit the inverse nonlinear relationship between performance parameters and structural size parameters.
[0085] The parameter optimization module is used to optimize the output structure size parameters of the decoder.
[0086] The parameter optimization module uses the L-BFGS-B optimization algorithm to optimize the output structure size parameters of the decoder.
[0087] The third embodiment of the present invention relates to an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the agent-driven sensor structure design optimization method of the first embodiment.
[0088] The fourth embodiment of the present invention relates to a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the agent-driven sensor structure design optimization method of the first embodiment.
[0089] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0090] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction methods implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0093] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for optimizing the design of a sensor structure driven by an intelligent agent, characterized in that, Includes the following steps: Obtain the design requirements for the sensor; The design requirements of the sensor are input into the intelligent agent, which outputs optimized structural parameters of the sensor; wherein, the intelligent agent includes: The large language model part is used to extract the performance targets and structural size constraints of the sensor from the design requirements; The deep learning model part is used to output optimized structural parameters of the sensor based on the sensor's performance objectives and structural size constraints.
2. The sensor structure design optimization method driven by an intelligent agent according to claim 1, characterized in that, The large language model is based on the Deepseek architecture and uses a multi-level attention mechanism to capture information about the sensor's performance goals in the design requirements.
3. The agent-driven sensor structure design optimization method according to claim 1, characterized in that, The deep learning model includes: The encoder section takes sensor structural parameters as input and sensor performance as output. The decoder section takes sensor performance as input and sensor structural parameters as output. Both the encoder and decoder sections adopt the Transformer architecture.
4. The agent-driven sensor structure design optimization method according to claim 3, characterized in that, The encoder portion includes multiple stacked encoders, each encoder comprising: The first encoder sublayer consists of a multi-head self-attention layer, a normalization layer, and a residual connection. The multi-head self-attention layer is used to capture the long-range dependencies and coupling relationships between multiple sensor structural parameters. The second encoder sublayer consists of a feedforward neural network layer, a normalization layer, and a residual connection. The feedforward neural network layer is used to perform a deep nonlinear transformation on the coupling features output by the first encoder sublayer to fit the nonlinear relationship between structural parameters and performance.
5. The agent-driven sensor structure design optimization method according to claim 3, characterized in that, The decoder section includes multiple stacked decoders and parameter optimization modules; the decoder includes: The first decoder sublayer consists of a multi-head attention layer, a normalization layer, and a residual connection. The multi-head attention layer includes a multi-head self-attention layer and a multi-head cross-attention layer. The multi-head self-attention layer is used to capture the correlation between performance parameters and structural size constraints, as well as the priority relationship among multiple performance parameters. The multi-head cross-attention layer is used to align the performance and structural size constraint features of the decoder with the nonlinear relationship between the structural parameters and performance of the encoder, ensuring that the reverse-generated structural parameters conform to the physical laws learned by the encoder. The second decoder sublayer consists of a feedforward neural network layer, a normalization layer, and a residual connection. The feedforward neural network layer is used to perform a nonlinear transformation on the positive alignment features output by the second decoder sublayer to fit the inverse nonlinear relationship between performance parameters and structural size parameters. The parameter optimization module is used to optimize the output structure size parameters of the decoder.
6. The agent-driven sensor structure design optimization method according to claim 5, characterized in that, The parameter optimization module uses the L-BFGS-B optimization algorithm to optimize the output structure size parameters of the decoder.
7. A sensor structure design optimization device driven by an intelligent agent, characterized in that, include: The acquisition module is used to acquire the design requirements of the sensor; A design module is used to input the design requirements of the sensor into an intelligent agent and output optimized structural parameters of the sensor; wherein, the intelligent agent includes: The large language model part is used to extract the performance targets and structural size constraints of the sensor from the design requirements; The deep learning model part is used to output optimized structural parameters of the sensor based on the sensor's performance objectives and structural size constraints.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the agent-driven sensor structure design optimization method as described in any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the agent-driven sensor structure design optimization method as described in any one of claims 1-6.