A sensor design method that actively adapts to the application environment

Abstract

This invention belongs to the interdisciplinary field of sensor design and artificial intelligence, specifically a sensor design method that proactively adapts to the application environment. This invention employs an active learning approach, actively selecting the most informative samples from unlabeled data for labeling, thereby maximizing model performance with less data acquisition. Compared to traditional methods of multiple sensor fabrications, active learning dynamically assesses the uncertainty or representativeness of data samples during model training, prioritizing data points that have the greatest impact on the model's decision boundaries and the strongest performance improvement potential. This method is not only applicable to a single type of sensor chip, but its database design and algorithm model are compatible with various material systems (including organic / inorganic, flexible / rigid, etc.), and can achieve rapid adaptive modeling and parameter recommendation for different application scenarios through transfer learning, demonstrating good scalability and versatility.

Description

Technical Field

[0001] This invention belongs to the intersection of sensor design and artificial intelligence, or the intersection of intelligent sensor system development and automatic design. It involves using active learning optimization models and machine learning to train sensor parameter design models, specifically a sensor design method that actively adapts to the application environment. Background Technology

[0002] In recent years, with the rapid development of technologies such as artificial intelligence and intelligent manufacturing, sensors, as core components of information perception, have also been advancing along with this trend. As a core component of information perception, the performance of sensors directly affects the overall intelligence level of the system. Currently, traditional sensor design relies on engineering experience, repeated experiments, and limited data feedback. The design process is lengthy, costly, and complex to debug, making it difficult to meet the rapid development needs of today's technology for diverse and customized sensors.

[0003] Existing research, such as data-driven soft sensor development based on deep learning techniques, is mostly data-driven. Its design process still relies heavily on a large number of manually set features, which cannot achieve the effect of training models with less data. Moreover, the models have poor adaptability, and there is currently no unified standard for sensor manufacturing and characterization. Different laboratories use different manufacturing schemes, which leads to bias and inconsistency in the training data.

[0004] Meanwhile, some studies have attempted to introduce reinforcement learning or neural network structures into the sensor response optimization process, but the following problems are common: (1) a large number of labeled samples are required, the sampling overhead is large, and high-quality datasets are very scarce; (2) the design space cannot be effectively explored, and it is easy to get trapped in local optima; (3) it is difficult to achieve automated transfer of the design process, and the design knowledge is difficult to reuse.

[0005] Active learning, as an efficient data acquisition strategy, has demonstrated strong sampling efficiency and model improvement potential in fields such as image recognition and materials design in recent years. However, there is currently a lack of technical pathways to deeply integrate active learning strategies with sensor design, making it difficult to achieve efficient iteration and automatic optimization for performance metrics.

[0006] Therefore, how to combine active learning and machine learning methods to build a design algorithm framework that can automatically optimize sensor structure and parameter configuration has become a key technical problem that urgently needs to be solved in the field of intelligent sensor design. Summary of the Invention

[0007] This invention addresses the technical problems of existing sensor (chip) design methods, such as large sample labeling requirements and uneven distribution of training samples, by providing a sensor design method that actively adapts to the application environment.

[0008] This invention is achieved using the following technical solution: a sensor design method that actively adapts to the application environment, comprising the following steps: Step 1: Collect the design parameters and performance parameters of a certain sensor under different specifications to form an initial data point set; Step 2: Support Vector Machine (SVM) modeling. The SVM is trained using the results of performance evaluations by sensors with different parameter configurations, thereby defining the boundary. Step 3: After preprocessing the initial data point set obtained in Step 1, input it into the Support Vector Machine (SVM). The SVM performs initial screening of the input data based on the predefined boundaries. Data within the boundaries after screening constitutes the database. A single pass through the database is performed, calculating the L2 Euclidean distance between each data point and all other data points. The minimum L2 Euclidean distance for each data point is obtained, and then the average of the minimum L2 Euclidean distances of all data points is taken. It is then determined whether this average is greater than a preset threshold: if it is greater than the threshold, the data point coverage is considered sufficiently broad; if it is less than the threshold, the data point coverage is considered insufficient. Step 4: Active learning loop based on decision trees and artificial neural networks ① Input the data points that exceed the threshold in step three into the decision tree model to generate new, more representative data points; ② Input more representative data points into the artificial neural network, and the artificial neural network outputs the predicted sensor structural parameters and performance parameters; ③ Expand the data points below the threshold in step 3 using data augmentation or genetic algorithms to generate additional training data; sort the additional training data to select target data points that meet the sensor design requirements and add them to the initial data point set to complete the update of the initial data point set; ④ Repeat step three with the updated data point set, complete the filtering, and then go through a new round of traversal. After updating the database, continue to repeat step four, output updated and more representative data points, expand and optimize the boundary of the support vector machine, and output the newly predicted sensor structure parameters and performance parameters; continuously repeat the above process to complete the active learning loop based on decision trees and artificial neural networks.

[0009] The main challenges in automated sensor design are the lack of sufficient datasets and high-quality data. This invention addresses this by employing an active learning approach. Specifically, it involves actively selecting the most informative samples from unlabeled data for labeling, thereby maximizing model performance with minimal data acquisition. Compared to traditional methods of creating sensors multiple times in experiments, active learning dynamically assesses the uncertainty and representativeness of data samples during model training, prioritizing the collection of data points that have the greatest impact on the model's decision boundaries and the strongest potential for performance improvement.

[0010] Furthermore, in step two, the support vector machine adopts an SVC model with an RBF kernel function and encapsulates it, treating each output target as an independent SVR sub-model, and automatically assigning tasks during the training and prediction phases to achieve the purpose of multi-dimensional input; the RBF kernel function uses 10-fold cross-validation, and the gamma parameter of the support vector machine is adjusted using Bayesian hyperparameter optimization.

[0011] Furthermore, the formula for the RBF kernel function is as follows: ; x i , x j —Feature vectors of two input samples; γ —The core hyperparameter of the RBF kernel controls the locality of the kernel function.

[0012] Furthermore, in step three, before inputting the data from step one into the support vector machine, the data needs to be preprocessed to transform the input data into trainable feature-label data pairs, and then the input data is normalized.

[0013] Furthermore, in step four, the base regressor of the decision tree model is DecisionTreeRegressor, which is weighted and integrated using AdaBoostRegressor, and encapsulated using MultiOutputRegressor or RegressorChain.

[0014] Furthermore, in each round of active learning in step four, the AdaBoostRegressor of the decision tree model focuses on the data with the largest error in the previous round and assigns it a higher weight, thereby improving the overall fitting accuracy.

[0015] Furthermore, in step four, the input layer of the artificial neural network is defined using Keras's Input layer, the shape of which is determined by the dimension of the input feature vector; the intermediate layer includes a fully connected layer applying the ReLU function and two stacked hidden layers; the output layer uses a linear activation function, the loss function is MSE, and the optimizer is Adam.

[0016] Furthermore, in step four, data augmentation employs two methods: one is SMOTE-based data augmentation, and the other is perturbation-based invariant data augmentation.

[0017] This invention provides a sensor design method that proactively adapts to the application environment. By introducing a data-driven intelligent optimization mechanism, it achieves closed-loop optimization from parameter design and environmental modeling to sensor structure feedback. Compared with existing sensor design methods, this invention has the following significant technical advantages: 1. Improve the accuracy of early-stage design predictions by constructing structure-performance relationships based on multiple parameters.

[0018] This invention employs a coupled modeling approach between sensor structural design parameters (such as thickness, electrode spacing, and material type) and performance response parameters (such as sensitivity and range). Compared to existing design methods that primarily rely on empirical formulas or limited experimental iterations, this invention establishes a unified physical-statistical-data model framework. Through experiments, a structure-performance database for multiple material systems is established, and ensemble regression algorithms (including decision tree regression and support vector regression) are used to screen sensitivity parameters. This allows for the prediction of performance trends before actual chip fabrication, with an average prediction error of less than 7%.

[0019] 2. An active learning mechanism is introduced to achieve efficient parameter space exploration and sample selection.

[0020] Unlike traditional passive sample collection methods, this invention employs an active learning loop. Building upon existing knowledge, it evaluates the parameter space based on prediction uncertainty and sample distribution density, selecting the target data points with the highest information content for subsequent simulations or experimental verification. Experimental comparisons show that the active learning parameter recommendation strategy reduces the number of samples by approximately 41.2% compared to a random selection strategy, and the model performance stabilizes after the third round of sample iterations, significantly improving sample utilization efficiency and model convergence speed.

[0021] 3. Integrate multiple machine learning models to improve the generalization ability of performance prediction and optimized recommendations.

[0022] In a parameter recommendation system for sensor performance prediction and active learning, this invention integrates multiple models, including decision trees, neural networks, and support vector machines, for fusion prediction, and compares the outputs of different models for error comparison and confidence analysis. Experiments show that in a typical thin-film pressure sensor case, the ensemble model achieves an average R² of 0.96, which is superior to a single model (such as a neural network with an R² of 0.91). This ensemble strategy enhances the model's adaptability to different data scales and avoids overfitting.

[0023] 4. Construct a data-driven feedback loop to support the continuous evolution of sensor design solutions. The closed-loop mechanism of "sensor parameter suggestion → performance modeling → experimental verification → database update" proposed in this invention ensures that each actively learned parameter recommendation continuously approaches the target performance point based on the existing optimal solution, achieving automatic evolution of the design scheme. With the support of an on-chip integrated hardware testing platform, each round of data iteration can shorten the experimental cycle by an average of 3-5 days, effectively reducing the design cycle and resource consumption.

[0024] 5. Supports scalable design across material systems and scenarios.

[0025] The method of this invention is not only applicable to a single type of sensor chip (such as piezoresistive type), but its database design and algorithm model are compatible with multiple material systems (including organic / inorganic, flexible / rigid, etc.). It can also achieve rapid adaptive modeling and parameter recommendation under different usage scenarios (such as high temperature, high humidity, strong interference) through transfer learning, demonstrating good scalability and versatility. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the support vector machine model structure; Figure 2 This is a schematic diagram of an artificial neural network structure; Figure 3 This is a schematic diagram of the decision tree model structure; Figure 4 A schematic diagram illustrating the optimization of machine learning models using data augmentation and particle swarm optimization algorithms; Figure 5 This is a block diagram of the overall structure of the method described in this invention; Figure 6 This is a simulation test diagram of an embodiment of the present invention. Detailed Implementation

[0027] The existing technology has the following shortcomings: The lack of high-quality, standardized datasets has resulted in poor performance of machine learning-based sensor modeling and performance prediction. The uneven distribution of training samples makes it difficult to cover critical operating ranges such as extremely low and ultra-high, which limits the applicability of sensor design in multiple scenarios. Traditional design methods rely heavily on human experience and multiple rounds of experimental verification, resulting in low efficiency, long cycles, and poor scalability, which cannot meet the needs of large-scale and rapid sensor development. Existing machine learning methods have high data requirements and are heavily dependent on training data. In practical applications, they often suffer from problems such as overfitting and poor generalization ability due to insufficient data.

[0028] Therefore, this invention aims to propose an automatic sensor design method based on the fusion of active learning and machine learning. This method can actively select data with high information value for training under conditions of small sample size and weak supervision, effectively improving model accuracy, design efficiency and sensor adaptability, thereby realizing the automated and high-performance design of flexible intelligent sensors under different strain ranges and structural features.

[0029] Technical solution: This invention mainly consists of four parts: dataset construction and updating, machine learning modeling, active learning loop, and optimal selection of data points. 1. A boundary definition based on support vector machines First, a Support Vector Machine (SVM) classifier is trained using data with typical characteristics to define the boundary. Specifically, the SVM is trained using the performance evaluation results of sensors with different parameter configurations. This helps the machine learning model identify sensor design configurations that may have higher performance. The SVM uses an SVC model with an RBF kernel function for classification tasks. RBF is commonly used for non-linear functions. A 10-fold cross-validation is used. The dataset is first divided for each fold, and the model is trained for each fold, summarizing the features, labels, and prediction results of the validation set. The most important parameter in the SVM, gamma, is adjusted using Bayesian optimization. The RBF kernel function mainly controls the influence range of the data points. If its influence range is too large, the model will be too smooth and prone to underfitting; if it is too small, the model will be more prone to overfitting. The formula for RBF is as follows: The final generated model can be used to analyze the relationship between different parameter configurations and performance labels.

[0030] This part of the program is specifically divided into three parts: data processing, SVM modeling and validation. In the data preprocessing stage, the main tasks include feature processing of the input data, construction of feature vectors, and sample diversity assessment. Feature processing mainly involves extracting key performance indicators from the input experimental data and generating labels. Taking the NV color center magnetic field sensor as an example, the key performance indicators selected are: the position where the sensitivity first reaches 10, the position where the sensitivity first reaches 100, and the maximum value corresponding to the sensor's termination. If no sensitivity value meets the conditions, the endpoint sensitivity is used as a substitute. Constructing feature vectors mainly involves converting the input data into trainable feature-label data pairs and then normalizing the input data. Finally, sample diversity assessment is performed after each round of active learning to evaluate the feature space distance distribution between the currently selected experimental samples and determine whether the sample coverage is sufficiently broad. The specific process is as follows: (1) Load all sample features and call normalization; (2) Iterate through each round of experimental batches; (3) For each round of selected samples, calculate the L2 Euclidean distance between each sample and other samples; (4) Obtain the minimum L2 distance (nearest neighbor) for each sample, and then calculate the average of all samples in that round.

[0031] SVM modeling program To achieve accurate prediction of different performance indicators for different material combinations, a multi-objective regression model was constructed using support vector regression. The overall structure of the model is shown in the figure below. Figure 1 As shown.

[0032] Traditional SVR models only support single-target output. To achieve multi-dimensional label prediction, this invention encapsulates the SVR model. This method treats each output target as an independent SVR sub-model and automatically assigns tasks during training and prediction phases, thus achieving multi-dimensional input. Furthermore, if the input material dimension is one-dimensional, the data will be automatically expanded to maintain a consistent output data structure.

[0033] 2. Active learning loop based on ANN (Artificial Neural Network) and decision tree Active learning iteration is used to diversify training data during the exploration of the parameter space of sensor design.

[0034] This invention selects various random parameter configurations and integrates them with their respective parameter settings to create a dataset using decision trees and artificial neural networks. The dataset is fed into a nonlinear algorithm of ANN and decision trees, and six decision processes are trained with tenfold cross-validation. These decision processes aim to assess the unfamiliarity with target regions within the sensor design space. New data points are formed by combining the modulation curves presented by the recommended sensor design with their corresponding parameter configurations. These data points are integrated into the existing dataset, thereby redefining and updating the boundary of the constructed “navigation model.” This boundary coincides with the boundary core defined by the support vector machine (SVM), both representing the spatial boundary of sensor parameters. However, compared to the initial boundary of the SVM, the “navigation model” boundary is an expansion and optimization of the SVM boundary in the active learning loop, incorporating newly added data points to ensure a more comprehensive coverage of the parameter region within the design space. With repeated iterations of the active learning loop, the boundary of the navigation model gradually expands to accommodate a wider range of data points. Through continuous learning and adaptation, and by integrating new, more representative data points, the machine learning-based navigation model is refined. After multiple iterations of active learning, the model has acquired the ability to predict sensor performance characteristics to a certain extent, and thus the ability to design a certain type of sensor.

[0035] ANN model: such as Figure 2 As shown, the input feature vector of this model comes from the data and label array obtained in the previous step. First, the input feature vector is standardized and normalized, mainly to scale the values ​​of all features to a uniform numerical range, preventing some feature values ​​from dominating the network training process. The input layer of the neural network is defined using Keras's Input layer, and its shape is determined by the dimension of the input feature vector; that is, each sample inputs a vector containing multiple features. This input layer is responsible for receiving the pre-processed feature vector. The input feature vector is then passed to the first fully connected hidden layer, which has multiple neurons. This layer performs the following operations: linear transformation of the input vector; applying the ReLU function; and adding weight regularization L1 and L2 to suppress overfitting. To increase the model capacity, two hidden layers with the same structure are stacked in the network structure: these two hidden layers further extract features and perform non-linear mapping on the data. By stacking multiple hidden layers, the network can learn more complex input-output relationships. To normalize the output after one layer to speed up convergence and improve stability, batch normalization is also used. The final hidden layer output is fed into an output layer that uses a linear activation function suitable for regression tasks. The final output (batch_size, labels_dim) represents the prediction of the target variable for each sample. The loss function is MSE, and the optimizer is Adam.

[0036] Decision Tree Model: This model employs an ensemble learning structure, with the core idea of ​​combining multiple decision trees into a strong predictive model. Its base regressor is DecisionTreeRegressor, which uses AdaBoostRegressor for weighted ensemble processing, thereby improving the model's generalization ability and its ability to fit complex data. To support multi-output regression problems, the model is further encapsulated using MultiOutputRegressor or RegressorChain.

[0037] In the prediction of each output dimension, the lowest-level learning unit is the DecisionTreeRegressor, a tree-like model that partitions the feature space and performs non-linear fitting. To avoid overfitting, the model sets a `max_depth` parameter to limit the depth of each tree. Then, the AdaBoostRegressor combines multiple such trees, focusing on the sample with the largest error in the previous round during each training iteration and assigning it higher weights, thereby improving the overall fitting accuracy. This forms a strong model composed of multiple weak learners.

[0038] For multi-objective prediction problems, DTRmodel supports two methods for handling multi-dimensional outputs. The first is the default MultiOutputRegressor, which trains an independent regression model for each output dimension, suitable for situations where the outputs are unrelated. The second method is RegressorChain, which uses the predicted value of the previous objective as input for the prediction of subsequent objectives, suitable for situations where multiple output variables have logical or statistical dependencies.

[0039] During the training phase, the model receives normalized feature and label data and fits it to the selected regressor structure. When label normalization is enabled, the model learns the data distribution at a relative scale, which is beneficial for convergence and performance comparison. The prediction phase uses the same input preprocessing to predict the output from the new data. The predicted result has the same dimension as the label dimension; if normalization is used, it needs to be denormalized to restore it to the original scale. Figure 3 This is a diagram of the decision tree model structure.

[0040] 3. Optimize machine learning models using data augmentation and particle swarm optimization. To further reduce the discrepancy between model predictions and actual sensor performance, it was initially anticipated that over 1000 sensors would need to be designed and evaluated, which was clearly impractical. Therefore, it was decided to apply data augmentation and genetic algorithms to optimize the machine learning-based predictive model, aiming to improve its accuracy and efficiency.

[0041] In optimizing the prediction model, two methods were employed: data augmentation and genetic algorithms. Figure 4 As shown, first, data augmentation techniques are applied to generate additional training data. Then, this augmented data is used to train the decision tree and artificial neural network, resulting in two large pools of decision processes, each containing 200 decision programs. These decision programs are selected by a genetic algorithm that mimics biological evolution.

[0042] At this stage, some decision processes are selected from each decision process pool and combined into multiple candidate models, which are then ranked according to their prediction accuracy.

[0043] Data augmentation employs two methods. One is SMOTE-based data augmentation, which primarily introduces synthetic samples along the material feature dimension to alleviate the problem of insufficient or imbalanced training samples. First, the input feature matrix is ​​classified, with one-hot encoding used for sensor material information. Subset data is extracted according to the dimension of the input sensor and concatenated with corresponding labels to form multi-class independent datasets.

[0044] Within each thickness subset, the function randomly selects minority class samples by enumerating triplet combinations (itertools.combinations) and oversampling using the SMOTE algorithm. Each call to SMOTE generates a set of synthetic samples similar to these triplet combinations. This strategy controls the number of generated samples by limiting the number of combinations and the jump size, keeping it close to the requirements of the numel parameters. The synthetic samples are then reassembled with the original one-hot thickness encoding, and finally, all synthetic samples generated in different dimensions are integrated into a complete dataset. The final result is an enhanced feature and label matrix. This approach, by introducing artificially synthesized samples, effectively increases the diversity of training data and enhances the model's learning ability even when material combinations are sparse.

[0045] Another approach employs perturbation-based invariant data augmentation, introducing slight perturbations into multiple dimensions of the original samples to simulate experimental errors or minute fluctuations in real-world material properties. The model duplicates each original sample numel times, shifting each copy according to a random direction and the maximum allowed perturbation range. The magnitude of the perturbation is determined by the `allowed_variation` parameter, ensuring that the data is augmented in a "near but different" manner without violating the original physical laws. Each newly generated feature retains its original label values, assuming that slight feature changes will not affect its corresponding performance evaluation.

[0046] The algorithm employs DEAP within a hybrid framework of particle swarm optimization and genetic algorithm. During particle initialization, a particle object `part` is first generated, containing initial position, velocity, and an integer mask (`int_mask`). It randomly generates an initial position and a velocity vector `speed` based on the input boundary conditions `pmin` and `pmax`. The positions of integer variables are automatically marked and guaranteed to be integers using `round()` in subsequent updates. The particle update part uses the standard PSO update formula, consisting of the current velocity, inertia term (`w`), individual optimal guidance (`c1`), and global optimal guidance (`c2`). Each particle is iteratively updated according to the PSO logic. The entire optimization function first registers the particle generator and updater, configures the evaluation function, then initializes the particle swarm for initial evaluation, and finally enters the main iteration loop. In the loop, the fitness of each particle is evaluated first, and its historical best and swarm best are updated; updateParticle is called to update position and velocity; if the GA algorithm is used, the GA stage is entered, sampling is performed from the particles to construct the GA subswarm; crossover (cxTwoPoint) and mutation (the mutation function defined earlier) are performed; the fitness of the new individuals is calculated; and the poorly performing individuals in the PSO swarm are updated with the GA results.

[0047] To verify the practicality and optimization effect of the method of the present invention in the design process of different types of sensor chips, the present invention selects three typical application scenarios for verification, including flexible pressure sensor design, high-temperature gas sensor modeling and optimization, and multi-environment adaptability illumination sensor chip tuning. Specific implementation methods are as follows: Example 1: Structural Optimization Design of Flexible Piezoresistive Pressure Sensor Application Background: Flexible pressure sensors are required to have high sensitivity and low response hysteresis in wearable electronics and health monitoring devices.

[0048] Implementation process: Polydimethylsiloxane (PDMS) was selected as the substrate material, and a mapping database between structural parameters (such as film thickness, microbump radius, electrode spacing, etc.) and sensitivity was established. Using an active learning mechanism, five target sample parameter points are recommended in each round, with an experimental cycle of 48 hours per round. After four rounds of iteration (a total of 20 experimental points), the average error between the model's prediction performance and the actual measured values ​​was controlled within 6.4%.

[0049] Actual results: The final sensor structure achieved a sensitivity increase of approximately 28.3% and a hysteresis reduction of 17% compared to the initial design, demonstrating that the method of this invention can effectively improve design performance and reduce unnecessary experimental resource consumption.

[0050] Example 2: Adaptive Optimization of Parameters for High-Temperature Metal Oxide Gas Sensing Chip Application background: In scenarios such as industrial emission monitoring and vehicle exhaust monitoring, sensors need to operate in environments above 300°C for extended periods, and conventional design methods are difficult to adapt to rapidly changing high-temperature environments.

[0051] Implementation process: Using SnO2 and ZnO as the base materials, process parameters such as sensitive layer thickness, annealing temperature, and electrode spacing were extracted as design variables. Design a neural network model to predict sensitivity changes at high temperatures and introduce an active sample update strategy; After six rounds of design recommendations and experimental verification, the final stable parameter set showed an ethanol gas response ratio 1.9 times that of the initial scheme at 350°C.

[0052] Actual results: This embodiment demonstrates that the present invention can still effectively guide material selection and structural optimization under high-temperature conditions, ensuring high response while minimizing component drift of less than ±5%, and shortening the optimization cycle by approximately 60% compared to traditional trial-and-error optimization.

[0053] Example 3: Parameter Design and Optimization Application of NV Color Core Magnetism Sensing Chip Application Background: Nitrogen-vacancy (NV) color centers, as an emerging quantum magnetic sensing material, are widely used in fields such as micromagnetic imaging, biomagnetic detection and precision measurement. Their sensing performance is significantly affected by factors such as structural configuration, fabrication process and external coupling structure. Traditional design methods are difficult to predict performance and adapt quickly.

[0054] Implementation process: Using diamond

[111] oriented single crystal as a substrate, the core parameters affecting the magnetic field sensitivity of NV color centers were extracted, including doping concentration, annealing temperature, laser power, microwave field strength and microstructure pattern design, etc. An initial dataset was established, and the target performance indicators were defined as magnetic field sensitivity (nT / √Hz), spectral linewidth (MHz), and photoluminescence intensity change rate. The active learning algorithm in this invention is applied in combination with decision tree and neural network model to model and predict the structure-performance mapping relationship, and experimental verification is carried out after each round of design recommendation; After five rounds of sample updates, an optimized set of design parameters was obtained, enabling the NV color center chip to achieve stable sensitivity improvement under experimental conditions while maintaining a low decoherence width.

[0055] Experimental data description: The original design sample had a sensitivity of 56.3 nT / √Hz under standard test field strength, which was improved to 38.9 nT / √Hz after optimization. The signal-to-noise ratio of photofluorescence is improved by approximately 31%; The spectral linewidth remained stable within 9.2 MHz, which was significantly better than the non-optimized structure group (>15 MHz). The number of overall sample testing rounds was reduced by 40%, and the experimental cycle was shortened by about 12 days.

[0056] Figure 6 shows the sensitivity simulation test results under different experimental conditions: (a) and (b) show the effects of laser power and microwave power on the NV color center magnetic sensor; (c) and (d) show the effects of doping concentration and bias magnetic field on the sensor, reflecting the optimization direction in the parameter recommendation iteration process.

[0057] Actual results: This embodiment verifies the applicability of the present invention in quantum device design. Under the premise of high parameter space dimension and narrow process window, it can efficiently screen key variables and optimize design combinations through data-driven methods, significantly improve the sensitivity and stability of NV color center devices, and reduce experimental burden and testing costs.

Claims

1. A sensor design method that actively adapts to the application environment, characterized in that, Includes the following steps: Step 1: Collect the design parameters and performance parameters of a certain sensor under different specifications to form an initial data point set; Step 2: Support Vector Machine Modeling. The support vector machine is trained using the results of performance evaluations by sensors with different parameter configurations, thereby defining the boundary. Step 3: After preprocessing the initial data point set obtained in Step 1, input it into the Support Vector Machine (SVM). The SVM performs initial screening of the input data based on the predefined boundaries. Data within the boundaries after screening constitutes the database. A single pass through the database is performed, calculating the L2 Euclidean distance between each data point and all other data points. The minimum L2 Euclidean distance for each data point is obtained, and then the average of the minimum L2 Euclidean distances of all data points is taken. It is then determined whether this average is greater than a preset threshold: if it is greater than the threshold, the data point coverage is considered sufficiently broad; if it is less than the threshold, the data point coverage is considered insufficient. Step 4: Active learning loop based on decision trees and artificial neural networks ① Input the data points that exceed the threshold in step three into the decision tree model to generate new, more representative data points; ② Input more representative data points into the artificial neural network, and the artificial neural network outputs the predicted sensor structural parameters and performance parameters; ③ Expand the data points below the threshold in step 3 using data augmentation or genetic algorithms to generate additional training data; The additional training data is sorted to select target data points that meet the sensor design requirements and added to the initial data point set to complete the update of the initial data point set; ④ Repeat step three with the updated data point set, complete the filtering, and then go through a new round of traversal. After updating the database, continue to repeat step four, output updated and more representative data points, expand and optimize the boundary of the support vector machine, and output the newly predicted sensor structure parameters and performance parameters; continuously repeat the above process to complete the active learning loop based on decision trees and artificial neural networks.

2. The sensor design method for actively adapting to the application environment as described in claim 1, characterized in that, In step two, the support vector machine adopts an SVC model with an RBF kernel function and encapsulates it, treating each output target as an independent SVR sub-model, and automatically assigning tasks during the training and prediction phases to achieve the purpose of multi-dimensional input; the RBF kernel function uses 10-fold cross-validation, and the gamma parameter of the support vector machine is adjusted using Bayesian hyperparameter optimization.

3. The sensor design method for actively adapting to the application environment as described in claim 2, characterized in that, The formula for the RBF kernel function is as follows: ； x i , x j —Feature vectors of two input samples; γ —The core hyperparameter of the RBF kernel controls the locality of the kernel function.

4. The sensor design method for actively adapting to the application environment as described in claim 1, characterized in that, In step three, before inputting the data from step one into the support vector machine, the data needs to be preprocessed to transform the input data into trainable feature-label data pairs, and then the input data is normalized.

5. The sensor design method for actively adapting to the application environment as described in claim 1, characterized in that, In step four, the base regressor of the decision tree model is DecisionTreeRegressor, which is weighted and integrated using AdaBoostRegressor, and encapsulated using MultiOutputRegressor or RegressorChain.

6. The sensor design method for actively adapting to the application environment as described in claim 5, characterized in that, In each round of active learning in step four, the AdaBoostRegressor of the decision tree model focuses on the data with the largest error in the previous round and assigns it a higher weight, thereby improving the overall fitting accuracy.

7. The sensor design method for actively adapting to the application environment as described in claim 1, characterized in that, In step four, the input layer of the artificial neural network is defined using Keras's Input layer, the shape of which is determined by the dimension of the input feature vector; the intermediate layer includes a fully connected layer applying the ReLU function and two stacked hidden layers; the output layer uses a linear activation function, the loss function is MSE, and the optimizer is Adam.

8. The sensor design method for actively adapting to the application environment as described in claim 1, characterized in that, In step four, data augmentation is performed in two ways: one is SMOTE-based data augmentation, and the other is perturbation-based invariant data augmentation.

Images