A Multimodal Dynamic Equilibrium Optimization Fusion Classification Method and System Based on Continuous Incremental Learning

Abstract

This invention belongs to the field of intelligent sensing and pattern recognition technology, and discloses a multimodal dynamic equilibrium optimization fusion classification method and system based on continuous incremental learning. It includes the following steps: Step 1, multimodal database construction and preprocessing; Step 2, feature extraction of single-modal data; Step 3, multimodal classification based on dynamic equilibrium evidence fusion; Step 4, continuous incremental learning of categories based on feature mean matching. This invention solves the problems of existing multimodal fusion recognition methods, such as difficulty in feature alignment, strong model staticity, poor noise robustness, and lack of uncertainty balancing mechanisms, thereby achieving efficient extraction of features within each modality and enhanced noise robustness.

Description

Technical Field

[0001] This invention belongs to the field of intelligent sensing and pattern recognition technology, specifically relating to a multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning. This technology is designed for multimodal target classification and recognition tasks in complex environments. By jointly processing and dynamically learning data collected from different types of sensors, it achieves deep fusion and reliable classification of multimodal information, and can be widely applied in fields such as multi-source remote sensing target recognition, cross-modal intelligent information processing, and equipment fault diagnosis. Background Technology

[0002] In complex and ever-changing environments, observations from a single sensor are often limited by factors such as viewing angle, detection range, and signal-to-noise ratio, making it difficult to comprehensively and stably reflect target characteristics. Multimodal fusion (such as the collaborative analysis of radar RCS signals, visible light images, and electronic warfare signals) can overcome the limitations of single-modal sensors in terms of data coverage, recognition accuracy, and anti-jamming capabilities, providing the system with higher-dimensional and more robust sensing information. However, differences in data structure, temporal resolution, information redundancy, and physical semantics between heterogeneous sensors pose significant alignment difficulties and feature conflict problems for cross-modal fusion.

[0003] Current mainstream multimodal fusion methods mostly rely on static training paradigms and deterministic classification models. Firstly, feature-level fusion methods typically achieve modality combination through weighted concatenation or attention allocation. Model parameters are fixed after training and cannot be dynamically adjusted according to environmental or sample changes, leading to performance degradation in complex contexts. Secondly, traditional deep networks have limited robustness to signal noise and environmental disturbances, especially in low signal-to-noise ratio electronic warfare signals or unevenly lit image scenes, where feature representations are easily contaminated, weakening discriminative ability. Thirdly, existing classification networks mostly use Softmax decision outputs, which only provide deterministic class probabilities and cannot quantify the model's confidence in the prediction results. Furthermore, when the system continuously receives new modalities or new class data, the model often needs retraining, making continuous incremental learning difficult and prone to catastrophic forgetting, compromising the stability of learned knowledge. These shortcomings of existing technologies make it difficult for multimodal recognition systems to maintain a balance and reliability of classification performance in long-term operation and dynamic environments.

[0004] To overcome the above limitations, this invention proposes a multimodal dynamic balance optimization fusion classification technology based on continuous incremental learning. Through innovative network structure design and learning strategies, it achieves deep alignment, adaptive updating, and uncertainty balance modeling of different modal features, thereby ensuring recognition accuracy while enabling the system to operate continuously and stably in complex and changing environments. Summary of the Invention

[0005] This invention addresses the problems of existing multimodal fusion recognition, such as difficulty in feature alignment, strong model staticity, poor noise robustness, and lack of uncertainty balancing mechanism. It provides a multimodal dynamic balance optimization fusion classification method and system based on continuous incremental learning, thereby achieving efficient extraction of features within each modality and enhanced noise robustness.

[0006] To achieve the above objectives, the specific technical solution of the present invention is as follows: In a first aspect, the present invention provides a multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning, which includes the following steps: Step 1: Preprocess and extract features from the multimodal heterogeneous data to obtain semantic feature vectors for each modality; Step 2: Input the feature vectors of each modality into the evidence deep learning classification head to generate non-negative evidence quantity, and calculate the uncertainty of each modality to the classification result accordingly; Step 3: Based on the uncertainty of each modality, dynamically weight and standardize the amount of evidence to suppress the fusion contribution of high-uncertainty modalities; Step 4: Use Dempster's synthesis rules to perform joint reasoning on the standardized multimodal evidence, generate the fused class probability distribution, and evaluate the consistency between modalities through the conflict coefficient; Step 5: When adding new categories, the network output layer is expanded and a memory sample replay mechanism based on feature center similarity is introduced. Combined with the joint optimization strategy of distillation loss and cross-entropy loss, the model can stably maintain old knowledge and continuously learn new knowledge.

[0007] As a further technical solution of the present invention: the multimodal heterogeneous data includes radar RCS signals, visible light images, and electronic warfare signals; each modality is feature extracted using a dedicated deep network. The visible light image extracts visual semantic features through a ConvNeXtMvt network that includes a ConvNeXt backbone, a Transformer global modeling module, and a Cross-Attention fusion layer. The RCS signal is extracted through a ResNet network and mapped to an RCS feature vector through global average pooling and a fully connected layer; The electronic warfare signals are extracted through a deep residual shrinkage network, which employs an adaptive soft threshold function to achieve channel-level noise suppression.

[0008] As a further technical solution of the present invention: the non-negative evidence quantity output by the evidence deep learning classification head is used to construct the Dirichlet distribution parameters. Then calculate the expected probability of the category. and modal uncertainty ,in This represents the total number of categories.

[0009] As a further technical solution of the present invention: the dynamic weighted standardization process is as follows: ; in, For the uncertainty of mode m, the weight term Used to adaptively reduce the evidence contribution of high-uncertainty modes.

[0010] As a further technical solution of the present invention: in the Dempster synthesis rule, the three-modal evidence is converted into a basic probability allocation function, and its joint synthesis formula is: ; Among them, the conflict coefficient Used to measure the degree of inconsistency between three modes of information. ;like If it is determined to be a complete conflict, a system alarm or backup fusion mechanism will be triggered; when At that time, the synthesis result By normalization factor Eliminate the impact of conflict and achieve joint fusion of evidence from three sources.

[0011] As a further technical solution of the present invention: the memory sample replay mechanism includes: before task switching, selecting the representative sample in the feature space that is closest to its category center for each old category to form a memory set. This, along with the current new category samples, forms a joint training set. .

[0012] As a further technical solution of the present invention: assuming the old model is in the sample The output probability of is The output of the new model is The distillation loss is: ; minimize The new model maintains relative stability of the decision boundary of the old category while learning new categories, thus significantly reducing the phenomenon of "catastrophic forgetting".

[0013] As a further technical solution of the present invention: the total loss function of the joint optimization strategy is: ; For cross-entropy loss, It is a dynamic balancing coefficient that is adaptively adjusted according to the ratio of new and old category samples to balance plasticity and stability.

[0014] As a further technical solution of the present invention: the network output layer is expanded when a new category is added: the weight parameters of the old category remain unchanged, and only the classification header parameters corresponding to the new category are randomly initialized to ensure the continuity of historical knowledge and the scalability of the structure.

[0015] Secondly, this invention provides a multimodal dynamic equilibrium optimization fusion classification system based on continuous incremental learning, comprising: The multimodal preprocessing module is used to perform structured preprocessing on radar RCS signals, visible light images, and electronic warfare signals. The multimodal feature extraction module is used to extract semantic feature vectors for the three modalities respectively; The dynamic evidence fusion module is used to perform EDL classification, uncertainty calculation, evidence weighting standardization, and Dempster joint reasoning. The continuous incremental learning module is used to expand the network output layer when new categories are added, build a memory sample set based on feature center matching, and jointly optimize model parameters through distillation loss and cross-entropy loss to suppress catastrophic forgetting. The adaptive control module is used to dynamically adjust the balance coefficient according to the task stage. And trigger an alarm or backup mechanism when the fusion conflict coefficient reaches a threshold.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, centered on evidence-based deep learning, innovatively maps the quantifiable evidence quantity and uncertainty distribution generated by evidence-based deep learning into basic probability assignments. Furthermore, it employs an improved Dempster synthesis rule at the decision layer to achieve multimodal information fusion. This mechanism overcomes the limitations of traditional feature concatenation or weighted averaging methods in adaptively allocating weights between modalities. It maintains robust classification performance even when multi-source sensor signals contain noise, missing information, or conflicts, achieving reliable integration of heterogeneous modal information.

[0017] 2. This invention introduces a continuous incremental learning mechanism, enabling the model to dynamically expand and stably maintain knowledge when facing new modalities or new categories of samples. By jointly optimizing cross-entropy loss and distillation loss, this invention maintains the stability of the decision boundaries of old categories while continuously learning new knowledge, effectively suppressing the "catastrophic forgetting" problem that is common in traditional deep networks in incremental scenarios. This mechanism allows the multimodal classification model to continuously evolve under gradually changing data distribution conditions, achieving synergistic optimization of knowledge transfer and memory balance, and providing reliable support for long-term intelligent recognition in complex electromagnetic and visual hybrid scenarios.

[0018] 3. The multimodal dynamic equilibrium optimization fusion classification technology based on continuous incremental learning proposed in this invention solves the problems of inconsistent multimodal heterogeneous data, difficulty in adaptive mode weights, and insufficient model sustainability of traditional methods through collaborative innovation in multiple levels such as modal feature extraction, uncertainty quantification, dynamic equilibrium fusion and continuous learning. It provides a robust, accurate and continuously optimizable technical solution for multimodal target classification in complex environments.

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning, as presented in this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be described in more detail below with reference to the accompanying drawings.

[0022] In the accompanying drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, of the embodiments of the present invention.

[0023] The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0024] The following is in conjunction with the appendix Figure 1 The embodiments of the present invention will be described in detail below.

[0025] Example 1 refer to Figure 1 As shown, this example implementation provides a multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning, which may specifically include the following steps: Step 1: Construction and preprocessing of the multimodal database; Step 101: Visible light data is acquired by a multi-angle imaging system and processed through pixel-level alignment, scale normalization and illumination equalization to ensure the spatial consistency of the target subject under different viewing angles and magnifications; Step 102: RCS signal data is generated using electromagnetic simulation. Reasonable incident angle, polarization mode and frequency range are set, and the comparability of signals in the angle and amplitude dimensions is ensured through angular domain resampling, amplitude normalization and smoothing window processing. Step 103: Electronic warfare signal data is generated by simulation using the TorchSig library, covering a variety of typical modulation methods such as BPSK, QPSK, and 16QAM. Different communication environments are simulated by setting parameters such as signal-to-noise ratio, carrier frequency, and attenuation rate to suppress out-of-band interference and maintain the separability between different modulation samples. Step 2: Feature extraction from single-modal data; Step 201: Visible light images are feature extracted using a ConvNeXtMvt network based on a ConvNeXt+Transformer+Cross-Attention structure. The process involves the ConvNeXt backbone extracting local convolutional features, the Transformer module implementing global dependency modeling, and the Cross-Attention layer fusing multi-scale semantic information. Let the input visible light image sample be... ,in , These represent the height and width of the image, respectively, with 3 indicating the number of RGB channels. The input image, after standardization and data augmentation, is fed into the ConvNeXtMvt network for feature extraction. The extraction process can be represented as ,in, , , These are the outputs of the intermediate layers of the network. The output visual semantic features are then normalized using L2. ; Step 202: Extract RCS features using ResNet. Let the input be the preprocessed RCS signal matrix. ,in Indicates the number of sampling frequency points. This represents the number of angular domain samples in the simulated radar; the ResNet network uses a skip connection structure. The network output is mapped to a fully connected layer after global average pooling. For the input feature map, The convolution and activation transformations are represented by residual connections, which ensure gradient stability and the ability to represent deep features. The high-level feature maps extracted by the network are denoted as... ,in The network parameter set is then used; subsequently, global average pooling (GAP) and fully connected mapping are applied to the feature maps to obtain the final RCS feature representation. ; Step 203: Extract electronic warfare signal features using a deep residual shrinking network; the input signal is the baseband complex envelope signal after normalization and real-to-virtual separation. ,in Indicates two channels, I and Q. This represents the number of sampling points; the core adaptive soft thresholding function is used for channel-level noise suppression, and its operation is as follows: ,in, The threshold is adaptive and determined by the statistical characteristics of the input channels. , The signal is learned from a deep residual shrinking network; the network output is then processed by a multi-layer shrinking residual module and global average pooling to obtain the characteristic representation of the electronic warfare signal. ; Step 3: Multimodal classification based on dynamic equilibrium evidence fusion; Step 301: For each mode (rcs represents radar RCS signal, vis represents visible light image, ew represents electronic warfare signal), its feature vector After passing through the corresponding EDL classification head, a non-negative evidence vector is generated. The corresponding Dirichlet distribution parameters and expected class probabilities are respectively The uncertainty for this sample is defined as follows: ,in For the number of categories, here Characterizing the mode pair category support, and This describes the model's confidence level in the current input; Step 302: To avoid the impact of differences in evidence strength scales between different modalities on the fusion effect, uncertainty-weighted standardization is performed on the evidence from each modality. Weighting terms This is used to reduce the contribution of highly uncertain modes in the final fusion, thereby improving the robustness and reliability of the fusion results; Step 303: Implement joint reasoning of trimodal evidence using Dempster's rules of composition. Let the recognition framework be... , For the set of all possible classification results, the basic probability distribution function (BPA) for each modality is: ,satisfy ,in Let be a subset of propositions, representing some or all of the target hypotheses. When multiple modal evidence sources are independent, the joint synthesis of three modalities is defined as... Among them, the conflict coefficient Used to measure the degree of inconsistency between three modes of information. ;like This indicates that the three modal evidences are completely conflicting and cannot be effectively synthesized; when At that time, the synthesis result By normalization factor Eliminate the impact of conflict and achieve joint fusion of evidence from three sources; Step 4: Continuous incremental learning of categories based on feature mean matching; Step 401: Let the first... At this stage, the model has learned the set of categories. In the Phase, new category set introduced Then the set of incremental task categories is The corresponding network output layer parameters are from Expand to When initializing network parameters, the weights of the old classes remain unchanged, and only the parameters of the newly added classes are randomly initialized, thereby ensuring the scalability of the model structure and the continuity of historical knowledge. Step 402: To prevent the forgetting of old task knowledge during training on a new task, a distillation term is introduced into the loss function. This is achieved by constraining the consistency between the current model output and the previous model output to preserve old knowledge. Let the old model be in the sample... The output probability of is The output of the new model is Then the distillation loss is defined as ;minimize The new model maintains relative stability of the decision boundary of the old category while learning new categories, thus significantly reducing the phenomenon of "catastrophic forgetting". Step 403: Before each task switch, select a representative subset of samples from the old task samples. and compare it with the current task sample Together they form a new training set Sample selection is based on the class center similarity criterion. ,in For sample feature vectors, As the category center, For the first The number of samples selected for each class ensures that the replayed samples represent the old class information in terms of feature distribution; Step 404: During training, the model's total loss function is defined as follows: ,in This is a balancing coefficient used to adjust the learning weights of new and old tasks; For cross-entropy loss, the expression is: ,in The training dataset for the current stage contains samples of the new class. Compared with old category memory samples , As an indicator function, when the sample The real label is the first The value is 1 if the class is correct, and 0 otherwise; it is used to select the prediction probability corresponding to the correct class during the summation process. The adjustment can be made dynamically according to the task stage: when the model faces a large number of new categories, the size can be increased. Weights are used to improve plasticity; when there are many old task samples, reinforcement is applied. Constraints are used to maintain stability.

[0026] Example 2 This invention provides a multimodal dynamic balance optimization fusion classification technology based on continuous incremental learning. Through innovative network structure design and learning strategies, it achieves deep alignment, adaptive updating, and uncertainty balance modeling of different modal features, thereby ensuring recognition accuracy while enabling the system to operate continuously and stably in complex and changing environments.

[0027] Specifically, the following steps are included: Step 1: Preprocess and extract features from the multimodal heterogeneous data to obtain semantic feature vectors for each modality; Step 2: Input the feature vectors of each modality into the evidence deep learning classification head to generate non-negative evidence quantity, and calculate the uncertainty of each modality to the classification result accordingly; Step 3: Based on the uncertainty of each modality, dynamically weight and standardize the amount of evidence to suppress the fusion contribution of high-uncertainty modalities; Step 4: Use Dempster's synthesis rules to perform joint reasoning on the standardized multimodal evidence, generate the fused class probability distribution, and evaluate the consistency between modalities through the conflict coefficient; Step 5: When adding new categories, the network output layer is expanded and a memory sample replay mechanism based on feature center similarity is introduced. Combined with the joint optimization strategy of distillation loss and cross-entropy loss, the model can stably maintain old knowledge and continuously learn new knowledge.

[0028] Furthermore, the multimodal heterogeneous data includes radar RCS signals, visible light images, and electronic warfare signals; each modality undergoes feature extraction via a dedicated deep network. The visible light image extracts visual semantic features through a ConvNeXtMvt network that includes a ConvNeXt backbone, a Transformer global modeling module, and a Cross-Attention fusion layer. The RCS signal is extracted through a ResNet network and mapped to an RCS feature vector through global average pooling and a fully connected layer; The electronic warfare signals are extracted through a deep residual shrinkage network, which employs an adaptive soft threshold function to achieve channel-level noise suppression.

[0029] Furthermore, the non-negative evidence quantity output by the deep learning evidence classification head is used to construct the Dirichlet distribution parameters. Then calculate the expected probability of the category. and modal uncertainty ,in This represents the total number of categories.

[0030] Furthermore, the dynamic weighted standardization process is as follows: ; in, For the uncertainty of mode m, the weight term Used to adaptively reduce the evidence contribution of high-uncertainty modes.

[0031] Furthermore, in the Dempster synthesis rule, the three-modal evidence is converted into a basic probability assignment function, and its joint synthesis formula is as follows: ; Among them, the conflict coefficient Used to measure the degree of inconsistency between three modes of information. ;like If it is determined to be a complete conflict, a system alarm or backup fusion mechanism will be triggered; when At that time, the synthesis result By normalization factor Eliminate the impact of conflict and achieve joint fusion of evidence from three sources.

[0032] Furthermore, the memory sample replay mechanism includes: before task switching, selecting representative samples in the feature space that are closest to the class center for each old class to form a memory set. This, along with the current new category samples, forms a joint training set. .

[0033] Furthermore, let's assume the old model is in the sample The output probability of is The output of the new model is The distillation loss is: ; minimize The new model maintains relative stability of the decision boundary of the old category while learning new categories, thus significantly reducing the phenomenon of "catastrophic forgetting".

[0034] Furthermore, the total loss function of the joint optimization strategy is: ; For cross-entropy loss, It is a dynamic balancing coefficient that is adaptively adjusted according to the ratio of new and old category samples to balance plasticity and stability.

[0035] Furthermore, the network output layer is extended when a new category is added: the weight parameters of the old categories remain unchanged, and only the classification header parameters corresponding to the new categories are randomly initialized to ensure the continuity of historical knowledge and the scalability of the structure.

[0036] This invention also provides a multimodal dynamic equilibrium optimization fusion classification system based on continuous incremental learning, which includes: The multimodal preprocessing module is used to perform structured preprocessing on radar RCS signals, visible light images, and electronic warfare signals. The multimodal feature extraction module is used to extract semantic feature vectors for the three modalities respectively; The dynamic evidence fusion module is used to perform EDL classification, uncertainty calculation, evidence weighting standardization, and Dempster joint reasoning. The continuous incremental learning module is used to expand the network output layer when new categories are added, build a memory sample set based on feature center matching, and jointly optimize model parameters through distillation loss and cross-entropy loss to suppress catastrophic forgetting. The adaptive control module is used to dynamically adjust the balance coefficient according to the task stage. And trigger an alarm or backup mechanism when the fusion conflict coefficient reaches a threshold.

[0037] This invention solves the problems of inconsistent multimodal heterogeneous data, difficulty in adaptive modal weights, and insufficient model sustainability of traditional methods through collaborative innovation in modal feature extraction, uncertainty quantification, dynamic balance fusion, and continuous learning. It provides a robust, accurate, and continuously optimizable new method for multimodal target classification in complex environments, and is applicable to fields such as multi-source remote sensing target recognition, cross-modal intelligent information processing, and equipment fault diagnosis.

[0038] Thus, the objective of this invention has been achieved.

[0039] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. 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. A multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning, characterized in that, Includes the following steps: Step 1: Preprocess and extract features from the multimodal heterogeneous data to obtain semantic feature vectors for each modality; Step 2: Input the feature vectors of each modality into the evidence deep learning classification head to generate non-negative evidence quantity, and calculate the uncertainty of each modality to the classification result accordingly; Step 3: Based on the uncertainty of each modality, dynamically weight and standardize the amount of evidence to suppress the fusion contribution of high-uncertainty modalities; Step 4: Use Dempster's synthesis rules to perform joint reasoning on the standardized multimodal evidence, generate the fused class probability distribution, and evaluate the consistency between modalities through the conflict coefficient; Step 5: When adding new categories, the network output layer is expanded and a memory sample replay mechanism based on feature center similarity is introduced. Combined with the joint optimization strategy of distillation loss and cross-entropy loss, the model can stably maintain old knowledge and continuously learn new knowledge.

2. The multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning according to claim 1, characterized in that, The multimodal heterogeneous data includes radar RCS signals, visible light images, and electronic warfare signals; each modality undergoes feature extraction using a dedicated deep network. The visible light image extracts visual semantic features through a ConvNeXtMvt network that includes a ConvNeXt backbone, a Transformer global modeling module, and a Cross-Attention fusion layer. The RCS signal is extracted through a ResNet network and mapped to an RCS feature vector through global average pooling and a fully connected layer; The electronic warfare signals are extracted through a deep residual shrinkage network, which employs an adaptive soft threshold function to achieve channel-level noise suppression.

3. The multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning according to claim 1, characterized in that, The non-negative evidence quantity output by the deep learning evidence classification head is used to construct the Dirichlet distribution parameters. Then calculate the expected probability of the category. and modal uncertainty ,in This represents the total number of categories.

4. The multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning according to claim 1, characterized in that, The dynamic weighted standardization process is as follows: ； in, For the uncertainty of mode m, the weight term Used to adaptively reduce the evidence contribution of high-uncertainty modes.

5. The multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning according to claim 1, characterized in that, In the Dempster synthesis rule, the three-modal evidence is converted into a basic probability assignment function, and its joint synthesis formula is as follows: ； Among them, the conflict coefficient Used to measure the degree of inconsistency between three modes of information. ;like If it is determined to be a complete conflict, a system alarm or backup fusion mechanism will be triggered; when At that time, the synthesis result By normalization factor Eliminate the impact of conflict and achieve joint fusion of evidence from three sources.

6. The multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning according to claim 1, characterized in that, The memory sample replay mechanism includes: before task switching, selecting representative samples in the feature space that are closest to the class center for each old class to form a memory set. This, along with the current new category samples, forms a joint training set. .

7. The multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning according to claim 1, characterized in that, Assume the old model in the sample The output probability of is The output of the new model is The distillation loss is: ； minimize The new model maintains relative stability of the decision boundary of the old category while learning new categories, thus significantly reducing the "catastrophic forgetting" phenomenon.

8. The multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning according to claim 1, characterized in that, The total loss function of the joint optimization strategy is: ； For cross-entropy loss, It is a dynamic balancing coefficient that is adaptively adjusted according to the ratio of new and old category samples to balance plasticity and stability.

9. The multimodal dynamic equilibrium optimization fusion classification method based on continuous incremental learning according to claim 1, characterized in that, The network output layer is expanded when a new category is added: the weight parameters of the old categories remain unchanged, and only the classification header parameters corresponding to the new categories are randomly initialized to ensure the continuity of historical knowledge and the scalability of the structure.

10. A multimodal dynamic equilibrium optimization fusion classification system based on continuous incremental learning, characterized in that, include: The multimodal preprocessing module is used to perform structured preprocessing on radar RCS signals, visible light images, and electronic warfare signals. The multimodal feature extraction module is used to extract semantic feature vectors for the three modalities respectively; The dynamic evidence fusion module is used to perform EDL classification, uncertainty calculation, evidence weighting standardization, and Dempster joint reasoning. The continuous incremental learning module is used to expand the network output layer when new categories are added, build a memory sample set based on feature center matching, and jointly optimize model parameters through distillation loss and cross-entropy loss to suppress catastrophic forgetting. The adaptive control module is used to dynamically adjust the balance coefficient according to the task stage. And trigger an alarm or backup mechanism when the fusion conflict coefficient reaches a threshold.

Images