Indirect physical quantity measurement method based on multi-modal fusion and cross-domain data driving

Abstract

The invention discloses an indirect physical quantity measurement method based on multi-modal fusion and cross-domain data driving, which comprises the following steps: acquiring finite element simulation data, scaling experiment data and actual system measured data, and carrying out alignment, normalization and label matching on the three types of data to construct a cross-domain multi-modal data set; extracting time domain, frequency domain and space domain physical characteristics from the data; generating a joint feature vector by adopting a feature fusion method based on an attention mechanism; constructing and training a deep neural network model, and introducing domain adaptation loss to improve the generalization ability of the model to a real scene; processing real-time data by using the trained model, outputting a high-precision estimated value of a target physical quantity which cannot be directly measured, and feeding back new data to a data set; outputting a result through a visual interface or a data interface; and finally, online updating of the model is realized through an incremental learning algorithm. The problems that sensors are difficult to arrange in an extreme environment and the measurement precision of a single data source is low are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent measurement and experimental monitoring technology, specifically to an indirect physical quantity measurement method based on multimodal fusion and cross-domain data-driven approach. Background Technology

[0002] In high-end equipment systems such as rocket control systems, aircraft power systems, satellite and payload health management, missile launch systems, ground vehicle mechanical systems, and underwater torpedo condition assessment, there are numerous critical physical quantities that cannot be directly measured by sensors (such as combustion chamber internal temperature, structural stress distribution, and high-speed fluid internal pressure). These physical quantities often exist in extreme environments such as high temperature, high pressure, and strong radiation, making it impossible to deploy sensors or resulting in insufficient accuracy for direct measurement. Traditional indirect measurement methods often rely on a single data source or a simple model, failing to effectively utilize data from multiple sources, leading to poor generalization ability and limited measurement accuracy under complex operating conditions. Summary of the Invention

[0003] The purpose of this invention is to provide an indirect physical quantity measurement method based on multimodal fusion and cross-domain data-driven approach to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an indirect physical quantity measurement method based on multimodal fusion and cross-domain data-driven approach, comprising the following steps:

[0005] S1: Acquire multi-source cross-domain data of the target system, including finite element simulation data, scaled-down experimental data, and actual system measured data;

[0006] S2: Preprocess the three types of data obtained in S1, including data alignment, data normalization and label matching, construct a cross-domain multimodal dataset, and divide it into training set, test set and validation set;

[0007] S3: Extract physically meaningful time-domain, frequency-domain, and spatial-domain features from finite element simulation data, scaled-down experimental data, and actual system measured data, respectively.

[0008] S4: An attention-based feature fusion method is adopted to assign adaptive weights to modal features from different data sources, thereby achieving cross-domain feature fusion and generating a joint feature vector;

[0009] S5: Using the joint feature vector obtained in S4 as input and the target physical quantity as output, construct a deep neural network model. The deep neural network model includes an input layer, multiple hidden layers, and an output layer.

[0010] S6: During the training of deep neural network models, a domain adaptation loss function is introduced, which together with the task loss constitutes the total loss function. Joint optimization enhances the model's generalization ability under different data distributions.

[0011] S7: After the real-time acquired multimodal data is processed sequentially through S2 preprocessing, S3 feature extraction and S4 feature fusion, it is input into the trained deep neural network model, outputting an estimated value of the target physical quantity, and feeding back the real-time acquired data and the corresponding estimated value to the cross-domain multimodal dataset constructed in S2.

[0012] S8: Output the estimated value obtained from S7 through a visual interface or data interface for the state monitoring and control decision of the target system;

[0013] S9: Based on the new data fed back to the dataset from S7, the deep neural network model is fine-tuned online using an incremental learning algorithm to maintain the model's measurement accuracy in dynamic environments.

[0014] Furthermore, in S1, the finite element simulation data is obtained through multiphysics simulation software, the scaled-down experimental data is obtained by arranging sensors on a scaled-down experimental device, and the actual system measured data is obtained by arranging sensors on measurable parts of the actual system. The sensors include thermocouples, fiber optic sensors, acoustic emission sensors, and vibration sensors.

[0015] Furthermore, in S2, data alignment employs a dynamic time warping algorithm or a timestamp-based interpolation method; data normalization uses max-min normalization or Z-score normalization, with the max-min normalization formula being... The Z-score standardization formula is: ,in The mean, The standard deviation is used; label matching assigns target physical quantity labels to each set of data according to experimental or simulation conditions; the data is divided into training, testing and validation sets proportionally.

[0016] Furthermore, in S3, the time-domain features include mean, variance, kurtosis, skewness, and root mean square; the frequency-domain features include spectral peak value, band energy, centroid frequency, and frequency variance; and the spatial-domain features include gradient magnitude, texture features, and shape factor. Optionally, wavelet transform and short-time Fourier transform are used for time-frequency feature extraction. For image data, scale-invariant feature transformation or convolution feature extraction methods are used.

[0017] Furthermore, in S4, the weight calculation of the attention mechanism is implemented through a multilayer perceptron or a recurrent neural network, and the weight calculation formula is as follows:

[0018]

[0019]

[0020] The formula for the fused feature vector is: ,in For the first Feature vectors of each modality and It consists of a learnable weight matrix and a bias vector. It is a learnable weight vector. That is, to be assigned to the first Adaptive attention weights for each data source;

[0021] During training, the attention mechanism automatically reduces the fusion weights of features from unreliable data sources for samples with sensor noise, missing data, or sudden changes in operating conditions.

[0022] Furthermore, in S5, the deep neural network is a convolutional neural network, a long short-term memory network, a Transformer, or a hybrid model; the network structure includes an input layer, multiple hidden layers, and an output layer, with the hidden layer activation function being ReLU, LeakyReLU, or Tanh, and the output layer having a linear activation function; a Dropout layer and a BatchNormalization layer are added to the network.

[0023] Furthermore, in step 6, the domain adaptation loss adopts either the maximum mean difference or the domain classification loss in a domain adversarial neural network; the total loss function is... ,in For the total loss, To predict mission losses, For domain adaptation loss, To balance hyperparameters;

[0024] When the domain adaptation loss adopts the maximum mean difference, it is calculated on the feature representation of the intermediate layer of the deep neural network. Inter-domain alignment is achieved by minimizing the distribution distance between the simulated data batch and the measured data batch in the feature space.

[0025] Furthermore, in S8, the visualization interface displays the real-time curves, historical data, and early warning information of the target physical quantity; the data interface supports Modbus, OPC UA, or MQTT industrial communication protocols.

[0026] Furthermore, in S9, the incremental learning algorithm adopts either elastic weight merging or experience replay algorithm; the model update cycle is set according to actual application requirements.

[0027] When the incremental learning algorithm uses the elastic weight merging algorithm, the penalty term is: The total loss is ,in It is a parameter The diagonal elements of the Fisher information matrix for the old task are used to measure the importance of the parameter to the old task.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] This invention improves the completeness and diversity of the training set by fusing the time, frequency, and spatial features of cross-domain data; it enhances the model's predictive ability under unseen conditions through domain adaptation loss function and attention mechanism; this invention can achieve high-precision indirect measurement in scenarios where sensors cannot be deployed, such as high temperature, high pressure, and strong radiation, and can be directly integrated into existing measurement and control systems, making it suitable for health management and status assessment of complex equipment such as missiles, torpedoes, and spacecraft. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0031] Figure 2 A schematic diagram of the multi-source cross-domain dataset construction process of this invention;

[0032] Figure 3 This is a schematic diagram of feature fusion and neural network of the present invention;

[0033] Figure 4 This is a flowchart illustrating the specific process of the present invention. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0035] Please see Figure 1-4 This invention provides an indirect physical quantity measurement method based on multimodal fusion and cross-domain data-driven approach, comprising the following steps:

[0036] S1: Acquire multi-source cross-domain data of the target system, including finite element simulation data, scaled-down experimental data, and actual system measured data;

[0037] S2: Preprocess the three types of data obtained in S1, including data alignment, data normalization and label matching, construct a cross-domain multimodal dataset, and divide it into training set, test set and validation set;

[0038] S3: Extract physically meaningful time-domain, frequency-domain, and spatial-domain features from finite element simulation data, scaled-down experimental data, and actual system measured data, respectively.

[0039] S4: An attention-based feature fusion method is adopted to assign adaptive weights to modal features from different data sources, thereby achieving cross-domain feature fusion and generating a joint feature vector;

[0040] S5: Using the joint feature vector obtained in S4 as input and the target physical quantity as output, construct a deep neural network model. The deep neural network model includes an input layer, multiple hidden layers, and an output layer.

[0041] S6: During the training of deep neural network models, a domain adaptation loss function is introduced, which together with the task loss constitutes the total loss function. Joint optimization enhances the model's generalization ability under different data distributions.

[0042] S7: After the real-time acquired multimodal data is processed sequentially through S2 preprocessing, S3 feature extraction and S4 feature fusion, it is input into the trained deep neural network model, outputting an estimated value of the target physical quantity, and feeding back the real-time acquired data and the corresponding estimated value to the cross-domain multimodal dataset constructed in S2.

[0043] S8: Output the estimated value obtained from S7 through a visual interface or data interface for the state monitoring and control decision of the target system;

[0044] S9: Based on the new data fed back to the dataset in step S7, the deep neural network model is fine-tuned online using an incremental learning algorithm to maintain the measurement accuracy of the model in dynamic environments.

[0045] like Figure 4 As shown, the specific implementation steps are as follows:

[0046] S1: Obtain finite element simulation data, scaled-down experimental data, and actual system measured data.

[0047] In this step, finite element simulation data is acquired using multiphysics simulation software such as ANSYS, Abaqus, or COMSOL to simulate the physical field distribution of the target equipment under extreme environments such as high temperature, high pressure, and strong radiation, and outputs high-dimensional data including temperature field, stress field, and flow field.

[0048] Scaled-down experimental data are acquired by arranging sensors on a scaled-down experimental setup, which should maintain similar physical properties and boundary conditions to the real system.

[0049] Actual system measured data is obtained by placing high-temperature resistant and radiation-resistant sensors on measurable parts of the actual system. These sensors include, but are not limited to, thermocouples, fiber optic sensors, acoustic emission sensors, and vibration sensors.

[0050] Finite element simulations should undergo mesh independence verification and experimental validation to ensure the reliability of the simulation results. The similarity rate of scaled-down experiments should be determined through dimensional analysis to ensure the comparability of experimental data with the real system.

[0051] S2: Align, normalize, and match the three types of data obtained in S1 to construct a cross-domain multimodal dataset.

[0052] In this step, the acquired simulation data, experimental data, and measured data are preprocessed, including but not limited to data alignment, normalization, and label matching. The processed dataset is divided using a random sampling strategy to construct balanced training, testing, and validation sets.

[0053] Data alignment employs either the Dynamic Time Warping (DTW) algorithm or a timestamp-based interpolation method to align the three types of data on a time scale, ensuring data synchronization.

[0054] Optionally, data normalization employs max-min normalization or Z-score normalization to eliminate the influence of different physical dimensions and orders of magnitude, as shown in the following formula:

[0055] Max-Min Normalization:

[0056]

[0057] Z-score standardization:

[0058]

[0059] in, The mean, Standard deviation;

[0060] Label matching assigns corresponding target physical quantity labels to each set of data based on experimental or simulation conditions, forming a supervised learning dataset;

[0061] Dataset partitioning: The processed data is divided into training set, test set and validation set according to the ratio, usually 7:2:1.

[0062] S3: Extract physically meaningful time-domain, frequency-domain, and spatial-domain features from simulation data, experimental data, and measured data, respectively.

[0063] In this step, time-domain features include, but are not limited to: mean, variance, kurtosis, skewness, and root mean square; frequency-domain features include: peak value, band energy, centroid frequency, and frequency variance; and spatial-domain features include: gradient magnitude, texture features, and shape factor.

[0064] For simulation data, physical field distribution data at specific locations and times can also be extracted as spatial features.

[0065] Wavelet transform and short-time Fourier transform (STFT) can be used for time-frequency feature extraction; for image data (such as infrared thermal images), scale-invariant feature transform (SIFT) or convolutional feature extraction methods can be used.

[0066] S4: An attention-based feature fusion method is adopted to assign adaptive weights to different data sources, thereby achieving cross-domain feature fusion and generating a joint feature vector.

[0067] In this step, the weights of each modal feature are calculated using an attention network, which can be a multilayer perceptron or a recurrent neural network. The weight calculation formula is as follows:

[0068]

[0069] in, Let i be the eigenvector of the i-th mode. and For learnable parameters, Let be the attention weights for the i-th modality. The fused feature vector is:

[0070]

[0071] S5: Using the fusion features obtained in S4 as input and the target physical quantity as output, construct a deep neural network model.

[0072] In this step, the deep neural network can be a convolutional neural network (CNN), a long short-term memory network (LSTM), a Transformer, or a hybrid model. The network structure includes an input layer, multiple hidden layers, and an output layer. The activation function for the hidden layers can be ReLU, LeakyReLU, or Tanh, and the output layer uses a linear activation function.

[0073] Dropout and Batch Normalization layers are used to enhance the model's generalization ability and prevent overfitting.

[0074] S6: Introduce domain adaptation loss during the training of deep neural network models to enhance the model's generalization ability under different data distributions.

[0075] In this step, the domain adaptation loss can be either the Maximum Mean Difference (MMD) loss or the domain classification loss from a Domain Adversarial Neural Network (DANN). The total loss function is:

[0076]

[0077] in, This refers to task losses (such as mean squared error). For domain adaptation loss, To balance hyperparameters.

[0078] S7: The real-time collected multimodal data is processed by feature extraction and fusion and then input into the trained deep neural network model. The model outputs a high-precision estimate of the target physical quantity and adds this set of data to the cross-domain multimodal training set established in S2 to increase the model training set.

[0079] S8: Outputs estimation results through a visual interface or data interface for system status monitoring and control decisions.

[0080] In this step, the visualization interface can display real-time curves, historical data, and early warning information of the target physical quantity; the data interface can support industrial communication protocols such as Modbus, OPC UA, or MQTT, and can be integrated with existing control systems.

[0081] S9: Based on newly acquired measured or experimental data, perform online fine-tuning or incremental learning on the deep neural network model to maintain the model's accuracy in dynamic environments.

[0082] In this step, incremental learning algorithms such as Elastic Weights Combining (EWC) or Experience Replay are used to avoid catastrophic forgetting. The model update cycle can be set according to the actual application requirements, such as updating every 24 hours or after each task is completed.

[0083] Example:

[0084] This embodiment utilizes a multimodal fusion and cross-domain data-driven indirect physical quantity measurement method for monitoring and evaluating the airborne power system of a fixed-wing aircraft. The system consists of an airborne AC generator, DC-DC converter, power distribution bus, battery pack, key power supply units, and multi-level mission payloads. Traditional aircraft power system measurement methods rely on direct measurement data from single sensors (such as voltage, current, or temperature sensors), which are easily affected by sensor drift, installation location limitations, environmental interference, and data gaps, leading to inaccurate assessment of key physical quantities (such as battery remaining capacity, AC generator output health, and cabin power supply redundancy). This embodiment employs the method of this invention to estimate and infer key indirect physical quantities using multimodal measurement data without altering the aircraft's existing specifications, interface standards, and electrical system structure. The specific implementation steps are as follows:

[0085] S1: Obtain finite element simulation data, scaled-down experimental data, and actual system measured data.

[0086] In this embodiment, electrical simulation data, ground experimental data, and flight test data are acquired respectively.

[0087] First, a coupled electro-thermal-chemical finite element model of the battery was established using COMSOL Multiphysics software. The model input parameters included: load current curves for different flight profiles (takeoff, cruise, and landing), ambient temperature (-40℃ to +60℃), and a preset battery aging level (SOH decreasing from 100% to 70%). The model output included: the three-dimensional temperature field distribution inside the battery, the plate current density distribution, the internal ohmic impedance, and the simulated terminal voltage response curve. Mesh independence verification was performed before simulation, and the model parameters were calibrated and verified by comparing the results with cell test data provided by the supplier.

[0088] Secondly, 1:1 scale cell-level and 1:5 scale module-level ground-based experimental rigs were constructed. Sensor arrays were deployed on the rigs for data acquisition, including: Hall current / voltage sensors (measurement accuracy ±0.5%) to acquire electrical signals; an infrared thermal imager (30Hz frame rate, 640x480 resolution) to monitor the two-dimensional temperature field on the shell surface; and embedded fiber Bragg grating temperature sensors (placed at key points inside the simulated shell) to acquire internal point temperatures. Experimental conditions covered standard charge / discharge, pulsed load, high / low temperature cycling, and accelerated aging tests, collecting over 200 sets of experimental data under different health conditions.

[0089] Finally, measurable sensors that are allowed to be placed are installed on the actual aircraft. The data includes: battery terminal voltage and bus current, charge and discharge status recorded by the power management unit (PMU), temperature of the high-temperature resistant thin-film thermocouples mounted on the surface of the battery casing, and thermal images of the casing acquired during scheduled maintenance using a portable infrared thermal imager.

[0090] Before constructing the electrical simulation, the model parameters should be calibrated to ensure that the voltage and current waveforms output by the simulation are consistent with the experimental data. The experimental platform should have the ability to switch between repeatable operating conditions to ensure that the experimental and flight environments are consistent in trend.

[0091] S2: Align, normalize, and match the three types of data obtained in S1 to construct a cross-domain multimodal dataset.

[0092] Since simulation data, experimental data, and measured data differ in terms of time scale, electrical parameter scale, and data sparsity, this step performs unified processing on them.

[0093] First, for charge-discharge cycle data, the Dynamic Time Warping (DTW) algorithm is used to nonlinearly align the corresponding time series of simulation and experimental data based on the voltage curve of the measured data, ensuring that each charge-discharge stage (such as constant current charging, voltage plateau, and discharge cutoff) of all data sources is aligned on the time axis.

[0094] Secondly, the Z-score normalization method is used to process each feature independently. For the j-th feature vector from the k-th data source... Its standardized formula is:

[0095]

[0096] in, These are the original eigenvalues. and These are the mean and standard deviation of the feature on the corresponding data source (training set portion), respectively. This step eliminates the dimensional differences between voltage (V), current (A), temperature (°C), and image pixel values.

[0097] Finally, each aligned data sample (corresponding to a complete charge-discharge cycle or a specific operating condition segment) is assigned its actual or simulated SOH value as a label. The SOH baseline value is determined by preset parameters in the capacity calibration test or simulation model on the test bench. The processed total dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0098] S3: Extract physically meaningful time-domain, frequency-domain, and spatial-domain features from simulation data, experimental data, and measured data, respectively.

[0099] From the aligned data, extract multi-dimensional features that can characterize the battery state.

[0100] In the time domain, features are extracted from current and voltage sequences, including but not limited to: average charge and discharge voltage, variance of voltage curves, voltage drop slope in the middle of discharge, voltage rise rate at the end of charging, and coulombic efficiency of current integral.

[0101] In the frequency domain characteristics, a Fast Fourier Transform (FFT) is performed on the voltage fluctuation signal during the charging and discharging process to extract its energy distribution in the 0-100Hz range and calculate the centroid frequency. and frequency variance :

[0102]

[0103]

[0104] in, For frequency, This represents the power spectral density at the corresponding frequency.

[0105] In the spatial domain features, the following features are extracted for infrared thermal images:

[0106] a) Statistical characteristics: average temperature, maximum temperature, and temperature standard deviation of the entire battery casing surface;

[0107] b) Texture features: Contrast and correlation are calculated based on the Gray-Level Co-occurrence Matrix (GLCM);

[0108] c) Shape characteristics: Area proportion and contour complexity of high-temperature regions (pixel sets with temperatures more than 10% above the average temperature).

[0109] Optionally, in power electronic units such as converters, local texture changes of the heat sink thermal image can be further extracted to indirectly assess the aging degree of the device.

[0110] S4: An attention-based feature fusion method is adopted to assign adaptive weights to different data sources, thereby achieving cross-domain feature fusion and generating a joint feature vector.

[0111] This approach effectively fuses heterogeneous features from three data sources: simulation, experiment, and field measurement. Let the feature vectors extracted from the three data sources be... (simulation), (experiment), (Actual measurement). First, they are spliced ​​together as follows: .

[0112] The weights for each modality are calculated using an attention network:

[0113]

[0114]

[0115] in, The feature vector representing the i-th data source. and It consists of a learnable weight matrix and a bias vector. It is a learnable weight vector. This refers to the adaptive attention weights assigned to the i-th data source.

[0116] The joint feature vector z after fusion is calculated as follows:

[0117]

[0118] This mechanism enables the model to dynamically adjust its dependence on different data source features based on the characteristics of the input samples, ensuring that the invention adaptively selects reliable observation data at different stages of flight and achieves stable estimation of indirect physical quantities of the battery.

[0119] S5: Using the fusion features obtained in S4 as input and the target physical quantity as output, construct a deep neural network model.

[0120] Using the fused feature vector z as input, and the battery SOH estimate... To produce the output, a deep regression model is constructed. This embodiment uses the following hybrid network structure:

[0121] Input layer: Receives a fused feature vector z with dimension d.

[0122] Feature Refinement Layer: Two fully connected (Dense) layers with 256 and 128 neurons respectively, using the ReLU activation function, and both followed by a Batch Normalization layer and a Dropout layer (dropout rate 0.3).

[0123] Temporal-spatial feature extraction branch: After reshaping the features, they are input into a CNN module containing two convolutional layers (kernel size 3, channel numbers 32 and 64) to extract spatial correlations between features. Simultaneously, the features are input into a single-layer LSTM network (64 hidden units) to extract sequence dependencies. The output features of the CNN and LSTM are flattened and concatenated.

[0124] Regression Output Layer: The concatenated features are passed through a fully connected layer with 32 neurons and ReLU activation, and finally connected to a linearly activated single-neuron output layer to obtain the SOH estimate. .

[0125] S6: Introduce domain adaptation loss during the training of deep neural network models to enhance the model's generalization ability under different data distributions.

[0126] Total loss function Loss from predicting tasks Domain Adaptation Loss Weighted composition:

[0127]

[0128] in, This is a trade-off hyperparameter, which is set to 0.5 in this embodiment.

[0129] Task loss: Mean squared error (MSE) is used.

[0130]

[0131] Domain adaptation loss: Maximum mean difference (MMD) is used to reduce the difference in feature distribution between the simulated domain (S) and the measured domain (F). It is calculated at the output of the first layer of the feature refinement layer.

[0132]

[0133] in, and These represent batch features from the simulation and experimental domains, respectively. It is mapped to the regenerating kernel Hilbert space H through the Gaussian kernel function.

[0134] The model uses the Adam optimizer, is trained on the training set, and its performance is monitored on the validation set to prevent overfitting.

[0135] S7: The real-time collected multimodal data is processed by feature extraction and fusion and then input into the trained deep neural network model. The model outputs a high-precision estimate of the target physical quantity and adds this set of data to the cross-domain multimodal training set established in S2 to increase the model training set.

[0136] The trained model is deployed on airborne edge computing devices or ground station servers. During aircraft operation or ground maintenance, real-time voltage, current, temperature, and thermal imaging data are collected and, after undergoing the same preprocessing, feature extraction, and fusion processes as in S2-S4, are input into the model, which outputs a real-time estimate of the current battery SOH. Simultaneously, these new measured data-estimate pairs (after manual or automatic verification) are added to the training database described in S2 for subsequent model updates, forming a closed loop of "data acquisition - model estimation - data accumulation."

[0137] S8: Outputs estimation results through a visual interface or data interface for system status monitoring and control decisions.

[0138] In this embodiment, the estimation results are output in two ways:

[0139] Visual interface: The ground station software of the maintenance personnel displays the real-time value of SOH, historical decay curve, health rating (such as "excellent, good, medium, poor") and early warning information (such as alarm when SOH is below 80%).

[0140] Data interface: The SOH estimate and health status code are sent to the aircraft health management system (IVHM) or airborne maintenance system through a standardized data interface (MQTT protocol) to guide maintenance decisions.

[0141] S9: Based on newly acquired measured or experimental data, perform online fine-tuning or incremental learning on the deep neural network model to maintain the model's accuracy in dynamic environments.

[0142] To prevent model performance degradation after long-term operation and to adapt to individual differences in aircraft batteries, a periodic update mechanism is established. An incremental learning process is triggered every time a certain amount of new measured data (equivalent to 10 complete charge-discharge cycles) is accumulated. The Elastic Weights Combined (EWC) algorithm is used to update the original model parameters. Based on this, fine-tuning was performed. The key penalty item for EWC is:

[0143]

[0144] in, It is a parameter The diagonal elements of the Fisher information matrix for the old task are used to measure the importance of that parameter to the old task. The total loss becomes This method can effectively constrain the forgetting of old knowledge while learning new data, thus enabling the continuous evolution of the model.

[0145] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An indirect physical quantity measurement method based on multimodal fusion and cross-domain data-driven approach, characterized in that: Includes the following steps: S1: Acquire multi-source cross-domain data of the target system, including finite element simulation data, scaled-down experimental data, and actual system measured data; S2: Preprocess the three types of data obtained in S1, including data alignment, data normalization and label matching, construct a cross-domain multimodal dataset, and divide it into training set, test set and validation set; S3: Extract physically meaningful time-domain, frequency-domain, and spatial-domain features from finite element simulation data, scaled-down experimental data, and actual system measured data, respectively. S4: An attention-based feature fusion method is adopted to assign adaptive weights to modal features from different data sources, thereby achieving cross-domain feature fusion and generating a joint feature vector; S5: Using the joint feature vector obtained in S4 as input and the target physical quantity as output, construct a deep neural network model. The deep neural network model includes an input layer, multiple hidden layers, and an output layer. S6: During the training of deep neural network models, a domain adaptation loss function is introduced, which together with the task loss constitutes the total loss function. Joint optimization enhances the model's generalization ability under different data distributions. S7: After the real-time acquired multimodal data is processed sequentially through S2 preprocessing, S3 feature extraction and S4 feature fusion, it is input into the trained deep neural network model, outputting an estimated value of the target physical quantity, and feeding back the real-time acquired data and the corresponding estimated value to the cross-domain multimodal dataset constructed in S2. S8: Output the estimated value obtained from S7 through a visual interface or data interface for the state monitoring and control decision of the target system; S9: Based on the new data fed back to the dataset from S7, the deep neural network model is fine-tuned online using an incremental learning algorithm to maintain the model's measurement accuracy in dynamic environments.

2. The indirect physical quantity measurement method based on multimodal fusion and cross-domain data-driven approach according to claim 1, characterized in that: In S1, finite element simulation data is obtained through multiphysics simulation software, scaled-down experimental data is obtained by arranging sensors on a scaled-down experimental device, and actual system measured data is obtained by arranging sensors on measurable parts of the actual system. The sensors include thermocouples, fiber optic sensors, acoustic emission sensors, and vibration sensors.

3. The indirect physical quantity measurement method based on multimodal fusion and cross-domain data-driven approach according to claim 1, characterized in that: In step S2, data alignment employs a dynamic time warping algorithm or a timestamp-based interpolation method; data normalization uses max-min normalization or Z-score normalization, with the max-min normalization formula being... The Z-score standardization formula is: ,in The mean, The standard deviation is used; label matching assigns target physical quantity labels to each set of data according to experimental or simulation conditions; the data is divided into training, testing and validation sets proportionally.

4. The indirect physical quantity measurement method based on multimodal fusion and cross-domain data driving according to claim 1, characterized in that: In S3, the time-domain features include mean, variance, kurtosis, skewness, and root mean square. Frequency domain features include spectral peak value, band energy, centroid frequency, and frequency variance; spatial domain features include gradient magnitude, texture features, and shape factor; optionally, wavelet transform and short-time Fourier transform are used for time-frequency feature extraction; for image data, scale-invariant feature transformation or convolution feature extraction methods are used.

5. The indirect physical quantity measurement method based on multimodal fusion and cross-domain data-driven approach according to claim 1, characterized in that: In step S4, the weight calculation for the attention mechanism is implemented using a multilayer perceptron or a recurrent neural network, and the weight calculation formula is as follows: ， ， The formula for the fused feature vector is: ,in For the first Feature vectors of each modality and It consists of a learnable weight matrix and a bias vector. It is a learnable weight vector. That is, to be assigned to the first Adaptive attention weights for each data source; During training, the attention mechanism automatically reduces the fusion weights of features from unreliable data sources for samples with sensor noise, missing data, or sudden changes in operating conditions.

6. The indirect physical quantity measurement method based on multimodal fusion and cross-domain data-driven approach according to claim 1, characterized in that: In S5, the deep neural network is a convolutional neural network, a long short-term memory network, a Transformer, or a hybrid model; the network structure includes an input layer, multiple hidden layers, and an output layer, with the hidden layer activation function being ReLU, LeakyReLU, or Tanh, and the output layer having a linear activation function; a Dropout layer and a BatchNormalization layer are added to the network.

7. The indirect physical quantity measurement method based on multimodal fusion and cross-domain data driving according to claim 1, characterized in that: In step 6, the domain adaptation loss uses either the maximum mean difference or the domain classification loss in a domain adversarial neural network; the total loss function is... ,in For the total loss, To predict mission losses, For domain adaptation loss, To balance hyperparameters; When the domain adaptation loss adopts the maximum mean difference, it is calculated on the feature representation of the intermediate layer of the deep neural network. Inter-domain alignment is achieved by minimizing the distribution distance between the simulated data batch and the measured data batch in the feature space.

8. The indirect physical quantity measurement method based on multimodal fusion and cross-domain data-driven approach according to claim 1, characterized in that: In S8, the visualization interface displays real-time curves, historical data, and early warning information of the target physical quantity; the data interface supports Modbus, OPC UA, or MQTT industrial communication protocols.

9. The indirect physical quantity measurement method based on multimodal fusion and cross-domain data driving according to claim 1, characterized in that: In S9, the incremental learning algorithm adopts either elastic weight merging or experience replay algorithm. The model update cycle is set according to the actual application requirements; When the incremental learning algorithm uses the elastic weight merging algorithm, the penalty term is: The total loss is ,in It is a parameter The diagonal elements of the Fisher information matrix for the old task are used to measure the importance of the parameter to the old task.

Images