A small sample distillation model fine-tuning method for nuclear and biochemical emergency rescue field

Abstract

The embodiment of the disclosure provides a small sample distillation model rapid fine-tuning method for the field of nuclear biological and chemical emergency rescue, comprising: constructing a noisy cross-modal data pair based on the asynchronous sampling time sequence of a nuclear biological and chemical multi-modal sensor; freezing the parameters of a pre-trained student model, and extracting the intermediate feature representation of the cross-modal interaction module thereof as a basic representation; based on the alignment relationship between the intermediate feature representation and the original asynchronous time sequence data, a time sequence elastic alignment regularization term is constructed and introduced into the model fine-tuning loss function; through iterative optimization, the student model is self-adapted to align multi-modal time sequence under a small amount of labeled samples, and the multi-source decision conflict caused by pseudo-causal correlation is relieved. Through the scheme of the embodiment of the disclosure, the time sequence elastic alignment regularization term can be applied to the cross-modal interaction module of the student model to solve the multi-source decision conflict problem caused by the interference of pseudo-causal correlation between modes.

Description

Technical Field

[0001] This application relates to the field of nuclear, biological, and chemical protection technology, specifically to a method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency rescue. Background Technology

[0002] In the field of nuclear, biological, and chemical (NBC) emergency response, due to the extreme complexity of the on-site environment and the diverse physical characteristics of sensors, a multimodal sensing system is often constructed using gamma spectrometers, electrochemical gas sensors, and thermal imaging modules to comprehensively assess information such as radiation leaks, toxic gas diffusion, and heat source distribution. However, limited by differences in hardware triggering mechanisms and data transmission delays, these sensors generally suffer from asynchronous sampling problems in actual deployments (typical timestamp deviations can reach ±120 ms), making it difficult to strictly align cross-modal data in terms of time sequence. Existing small-sample fine-tuning methods based on knowledge distillation typically assume that features between modes are naturally coupled in the time dimension, ignoring the pseudo-causal correlation interference introduced by asynchronous sampling—that is, the model may incorrectly misjudge the delayed response of one mode as a causal triggering relationship of another mode, thus causing decision conflicts during multi-source information fusion. For example, it may incorrectly attribute subsequent gas concentration changes to previous radiation pulses, leading to misjudgment of the root cause of NBC events. To address this issue, there is an urgent need for a method that can apply physically-aware temporal elastic alignment regularization constraints to the cross-modal interaction modules of the student model during fine-tuning, so as to explicitly correct feature misalignment caused by time deviation and eliminate the negative impact of spurious causal associations on multimodal fusion decision-making. Summary of the Invention

[0003] In view of this, the present disclosure provides a method for rapid fine-tuning of a small-sample distillation model in the field of nuclear, biological and chemical emergency rescue, which at least partially solves the problems existing in the prior art.

[0004] A rapid fine-tuning method for a small-sample distillation model in the field of nuclear, biological, and chemical emergency response includes: Noisy cross-modal data pairs were constructed based on asynchronous sampling time series of nuclear biochemical multimodal sensors; The parameters of the pre-trained student model are frozen, and the intermediate feature representations of its cross-modal interaction modules are extracted as the basic representations. Based on the alignment relationship between the intermediate feature representation and the original asynchronous time series data, a time series elastic alignment regularization term is constructed and introduced into the model fine-tuning loss function; Iterative optimization enables the student model to adaptively align with multimodal time series with a small number of labeled samples, mitigating multi-source decision-making conflicts caused by spurious causal relationships.

[0005] According to one embodiment, the construction of noisy cross-modal data pairs based on asynchronous sampling time series from a nuclear biochemical multimodal sensor includes: The radiation intensity signal R[i], gas concentration C[j], and thermal field intensity T[k] in the environment were collected by a gamma spectrometer, an electrochemical gas sensor, and a thermal imaging module, respectively. Construct cross-modal data frame pairs (x[i], x[j], x[k]) based on timestamp deviation Δt, where x[i], x[j], and x[k] represent time series segments of the corresponding modalities, and Δt ranges from -120 ms to +120 ms. Different modal data are non-strictly synchronized, and noise σ[n] is introduced to perturb the cross-modal data pairs. The value range of σ[n] is [1, 0]. If the current sampling time t is greater than the maximum allowed time interval Tmax, then discard the data set; otherwise, retain it for subsequent processing.

[0006] According to one embodiment, the extraction of intermediate feature representations of cross-modal interaction modules includes: Perform forward propagation on the student model on the training data and extract the feature output F[m] from the m-th layer; Convert the feature output F[m] into a normalized vector, in the form F_norm[m] = (F[m] - μ) / σ, where μ is the channel mean and σ is the channel standard deviation; Generate independent cross-modal attention weights A[l] for each feature channel; We obtain the basic representation F_base = Σ[l](A[l] · F_norm[m]) by weighted summation of the feature vector F_norm[m].

[0007] According to one embodiment, constructing a time-series flexible alignment regularization term based on the alignment relationship between the intermediate feature representation and the original time-series data includes: The dynamic alignment coefficient ρ(t) of cross-modal features is calculated by sliding window matching; Define the elastic alignment function E = α·|Δt| + β·cos(θ), where α and β are the time offset and semantic consistency weighting parameters, respectively, and θ represents the cross-modal feature similarity angle; For each batch of feature pairs, calculate the elasticity loss term L_align = (1 / N)Σ[i=1 to N] E[i]; The regularization term is added to the final fine-tuning objective function only if L_align is greater than the threshold ε (ε ranges from [5, 0]).

[0008] According to one embodiment, the iterative optimization of adaptive aligned multimodal timing includes: Set the learning rate η and the maximum number of iterations Iter_max, where the value of η is in the range of [001, 01]. For each time t, calculate the current timing alignment state s[t] = aligned_count[t] / total_pairs[t]; If s[t] < 7, it indicates that the temporal alignment is insufficient. In this case, the feature pairs should be resampled and the alignment constraints should be strengthened. The early stopping mechanism is used to monitor and verify the rate of loss reduction. Training ends when there is no significant improvement for three consecutive iterations.

[0009] According to one embodiment, the method further includes constructing an adversarial training mechanism based on pseudo-causal association interference, including: Extract the latent pseudo-causal features F[p] from all cross-modal feature pairs; The strength of pseudo-causal associations is evaluated using the formula score_pseudo = (1 / k)Σ[l=1 to k] [Cj / Ri], where k is the number of time points. If score_pseudo is greater than λ (λ is the set significance threshold, with a value range of [6, 9]), then an adversarial perturbation ξ[i] is introduced; By using a gradient inversion layer to guide the student model to ignore the influence of spurious correlation features, the robustness of decision-making is improved.

[0010] According to one embodiment, it includes: Construct a triplet decision graph, where nodes represent multi-source decision outcomes d_gamma, d_gas, and d_thermal; Define the decision consistency weight W_d = count_c / count_a, where count_c represents the number of co-occurrences and count_a is the total number of occurrences of any decision combination; If W_d < 8, then the abnormal decision is smoothed and an additional penalty term L_conflict = ω ·MSE(d_exp, d_pred) is added; The final output is the integrated multi-source decision result d_final = w1·d_gamma + w2·d_gas + w3·d_thermal.

[0011] According to one embodiment, the iterative optimization of adaptively aligned multimodal timing further includes: Define the dynamic equilibrium parameter λ(t) = exp(t / τ), where τ is the decay constant; In each round of optimization, the weight of the regularization term is gradually adjusted according to λ(t), emphasizing time alignment in the early stage and focusing on overall prediction accuracy in the later stage. The weighted concatenation of each feature vector is expressed as F_joint = (1-λ(t))·F_base + λ(t)·F_align; A random dropout strategy is used to perform protective pruning on the high-fluctuation feature subspace to ensure model robustness.

[0012] According to one embodiment, the time-flexible alignment regularization term constructed based on intermediate feature representation and original asynchronous time series includes: Use temporal convolutional networks to extract hierarchical features from cross-modal data; Construct a bidirectional long short-term memory network model to estimate the temporal dependencies between features; The alignment error is calculated using the formula E[i] = √[(t_g - t_a)² + (score_g - score_a)²], where t_g is the timestamp of the gamma spectrometer, t_a is the timestamp of the electrochemical sensing data, and score_g and score_a are the corresponding confidence scores. When E[i] is greater than the preset threshold ε, a negative feedback mechanism is forcibly added to correct the timing alignment direction and reduce erroneous associations caused by time offset.

[0013] According to one embodiment, the alignment error calculation method incorporates a physical environment-aware factor, specifically including: Obtain the ambient temperature Temp_env (value range [0℃, 40℃]), relative humidity RH, and electromagnetic noise intensity ENI; The weighting factor ω_env is calculated based on environmental complexity: ω_env = 1 + RH / 100 + ENI / 200. The alignment error formula is modified to E[i] = ω_env · [(t_g - t_a)² + (score_g - score_a)²]; In environments with strong interference, a multi-modal fusion enhancement strategy is automatically activated to reduce the risk of judgment distortion caused by the instability of a single mode.

[0014] This disclosure provides a method for rapid fine-tuning of a small-sample distillation model in the field of nuclear, biological, and chemical emergency rescue. The method includes: constructing noisy cross-modal data pairs based on asynchronous sampling time series from nuclear, biological, and chemical multimodal sensors; freezing the parameters of a pre-trained student model and extracting intermediate feature representations of its cross-modal interaction modules as basic representations; constructing a time-series elastic alignment regularization term based on the alignment relationship between the intermediate feature representations and the original asynchronous time series data and introducing it into the model fine-tuning loss function; and using iterative optimization to enable the student model to adaptively align with the multimodal time series with a small number of labeled samples, thereby mitigating multi-source decision conflicts caused by pseudo-causal associations. The solution of this disclosure can apply a time-series elastic alignment regularization term to the cross-modal interaction modules of the student model to solve the problem of multi-source decision conflicts caused by pseudo-causal association interference between modalities. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the exemplary embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this disclosure and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a rapid fine-tuning method for a small-sample distillation model in the field of nuclear, biological and chemical emergency rescue. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the embodiments of this disclosure will be further described in detail below with reference to the accompanying drawings. The illustrative implementation methods and descriptions of the embodiments of this disclosure are only used to explain the embodiments of this disclosure and are not intended to limit the embodiments of this disclosure.

[0018] This embodiment provides a method for rapid fine-tuning of a small-sample distillation model in the field of nuclear, biological, and chemical emergency response. It aims to address the problems of spurious causal correlation interference and multi-source decision conflicts caused by asynchronous sampling from multimodal sensors, enabling a pre-trained small student model to quickly adapt to complex nuclear, biological, and chemical environments with a limited number of labeled samples. The core of this method lies in applying a novel temporal elastic alignment regularization term to the cross-modal interaction module of the student model, forcing the model to learn feature representations consistent with the physical real temporal sequence, rather than spurious statistical correlations in the data.

[0019] Next, refer to Figure 1 This invention describes a method for rapid fine-tuning of a small-sample distillation model in the field of nuclear, biological, and chemical emergency response. The method includes: Step S1: Construct noisy cross-modal data pairs based on the asynchronous sampling time sequence of the nuclear biochemical multimodal sensor.

[0020] This step aims to simulate the non-ideal state of data collection in real rescue scenarios, providing challenging and representative training samples for subsequent model fine-tuning.

[0021] Specifically, the first step is to deploy an integrated multimodal sensor suite, which includes at least a high-sensitivity gamma spectrometer (for detecting the types and intensities of radionuclides), an electrochemical gas sensor array (for quantitatively detecting the concentrations of various toxic and harmful gases, such as CO, H2S, and NH3), and an uncooled infrared thermal imaging module (for acquiring the temperature distribution of the scene and identifying heat sources and high-temperature areas). In actual emergency rescue operations, due to differences in the physical principles, data readout rates, and communication protocols of the various sensors, the data they collect are not strictly synchronized.

[0022] Assume that the gamma spectrometer acquires the radiation intensity signal R[i] at discrete time t_R[i], the electrochemical gas sensor acquires the gas concentration C[j] at time t_C[j], and the thermal imaging module generates a thermal field intensity image T[k] at time t_T[k]. The system records the timestamp of each data point and identifies the time deviation (Δt) between the three. To construct cross-modal data pairs for training, this step employs a data frame construction method based on a sliding time window.

[0023] A time window [-150ms, +150ms] is defined centered on a fixed time anchor point t (e.g., the decision time). Within this window, the gamma spectrometer reading R[i], gas concentration C[j], and thermal imaging frame T[k] closest to t are searched and selected. However, if the timestamp difference |t_x - t_y| between any two data points of R[i], C[j], and T[k] is greater than 120ms, they are considered physically uncorrelated due to excessive time deviation. In this case, instead of simply discarding all data, noise perturbation is introduced to simulate the uncertainty of the sensor itself and environmental interference, constructing more robust training samples.

[0024] The specific construction method is as follows: Select a set of original data points that meet the timestamp deviation condition (|t_x - t_y| ≤ 120ms) to form a basic data frame (R[i], C[j], T[k]). Based on this, an independent noise term σ[n] is artificially added to the data for each modality. The noise σ[n] can be a Gaussian white noise with a mean of 0 and a standard deviation between 0.1 and 1.0. For example, for the count rate R[i] of the gamma spectrometer, the value after adding noise is R'[i] = R[i]. (1+σ_R[n]). The introduction of noise simulates fluctuations in sensor electronic noise or ambient background radiation, forcing the model to learn denoising and robust feature extraction capabilities.

[0025] In addition, to prevent the model from being misled by outdated information, a maximum allowable time interval T_max is set (e.g., T_max = 500ms). If the difference between the current decision time t and the timestamp of any selected sensor data is greater than T_max, the sensor data is considered to be severely outdated and unable to reflect the current situation. The data pair should be discarded and the model should wait for new data to arrive before it is reconstructed.

[0026] Step S2: Freeze the parameters of the pre-trained student model and extract the intermediate feature representations of its cross-modal interaction modules as the basic representation.

[0027] This embodiment employs a pre-trained miniaturized student model, such as a lightweight Transformer model, which has undergone knowledge distillation on a large-scale general multimodal dataset. Before fine-tuning begins, the parameters of the entire student model are frozen, i.e., its gradient update is set to False. This means that in the initial stage of fine-tuning, the model's low-level feature extractors (such as convolutional layers or the first few layers of the Transformer) remain unchanged to preserve their powerful general representation capabilities.

[0028] Then, a batch of noisy cross-modal data pairs constructed in step S1 is input into the student model, and a forward propagation is performed. As the data flows through the model's cross-modal interaction modules (e.g., cross-attention layers or multimodal fusion layers in the model), we extract the intermediate feature output F[m] from the m-th layer (e.g., the 1st cross-attention layer) at the output of that module. Assuming the batch size is B and the feature dimension is D, F[m] is a tensor of shape [B, D].

[0029] To eliminate the influence of amplitude differences between different samples and facilitate subsequent calculations, we normalize F[m]. First, we calculate the mean μ and standard deviation σ for each feature channel along the batch and spatial dimensions (if present). Then, we standardize each feature vector in F[m] to obtain the normalized feature vector F_norm[m] = (F[m] - μ) / σ. This operation ensures that the feature values ​​of different channels are on a relatively consistent scale.

[0030] To further focus on the most important information in cross-modal interaction, this embodiment also dynamically calculates an attention weight A[l] for each feature channel. This weight can be calculated by a lightweight, learnable subnetwork (e.g., a single fully connected layer with a sigmoid activation function) based on F_norm[m] itself, and is used to evaluate the importance of each channel in the current multimodal alignment task.

[0031] Finally, these attention weights are used to perform a weighted summation of the normalized features to obtain the basic representation F_base representing the current sample. The calculation formula is: F_base = Σ[l=1 to D] (A[l]· F_norm[m][l]).

[0032] Step S3: Based on the alignment relationship between the intermediate feature representation and the original asynchronous time series data, construct a time series elastic alignment regularization term and introduce it into the model fine-tuning loss function.

[0033] First, to quantify the relationship between the model's intermediate representation F_base and the original sensor time-series data, we designed a dynamic alignment coefficient ρ(t). The calculation of ρ(t) is based on the idea of ​​sliding window matching. Specifically, we slide a time window of length L over a continuous time series segment. For the original sensor data sequence (e.g., R[i], C[j], T[k]) and the corresponding basic representation sequence F_base(t) within the window, we calculate the minimum alignment path cost between them using the Dynamic Time Warping (DTW) algorithm. The reciprocal of this cost is ρ(t). A higher ρ(t) value indicates a better match between the model's internal representation and the temporal fluctuation pattern of the original sensor signal.

[0034] Based on ρ(t), we further define an elastic alignment function E, which comprehensively considers both time offset and semantic similarity: E = α·|Δt| + β·cos(θ) Here, |Δt| is the absolute value of the actual maximum timestamp deviation of the data pair in step S1, which measures the degree of asynchrony at the physical level. α is a time offset weighting parameter used to control the severity of our penalty for time accuracy. cos(θ) is the cross-modal feature similarity angle, where θ is the angle between different modal feature vectors (e.g., feature subspaces corresponding to gamma and gas separated from F_base) in high-dimensional space, and the cosine value measures their semantic similarity. β is a semantic consistency weighting parameter. The design philosophy of this function is: we hope that when the model faces a large physical time deviation |Δt|, the semantic similarity cos(θ) between its internal representations will decrease accordingly, that is, the model can recognize that the information of these two modalities is temporally misaligned and should not establish a strong causal relationship; conversely, if the physical time deviation is small, the model is encouraged to increase the semantic similarity of its internal representations.

[0035] Subsequently, for all N samples in the current training batch, the average elasticity loss term L_align is calculated: L_align = (1 / N) Σ[i=1 to N] E[i] However, not all cases require the same regularization strength. To avoid over-regularization damaging the model's existing knowledge, we set an adaptive activation threshold ε (ε ranges from [0.5, 1.0]). Only when the calculated L_align is greater than the threshold ε is the current batch of samples considered to have a significant temporal misalignment problem, and only then is the regularization term added to the final fine-tuning objective function. The final loss function L_total can be expressed as: L_total = L_task(student model output, true label) + γ L_align (when L_align>ε) Where L_task is the original loss function for the downstream task (such as hazardous substance classification or concentration regression), and γ is the weighting coefficient that balances the two losses.

[0036] Step S4: Through iterative optimization, the student model adaptively aligns with multimodal time series with a small number of labeled samples, thereby alleviating multi-source decision-making conflicts caused by pseudo-causal associations.

[0037] Before fine-tuning begins, some parameters in the student model related to cross-modal interaction and decision-making (e.g., the cross-attention layer and the final classification / regression head) need to be unfrozen, while the underlying feature extraction layer remains frozen to fully utilize its general feature extraction capabilities and prevent overfitting. A suitable learning rate η is set; considering this is small-sample fine-tuning, η should not be too large. In this embodiment, η = 0.005 is used, and the AdamW optimizer is employed for parameter updates. Simultaneously, the maximum number of iterations, Iter_max, is set to 100 epochs.

[0038] In each iteration, we backpropagate the L_total constructed in step S3 as the optimization objective to update the trainable parameters of the model. The key is that we introduce dynamic monitoring of the temporal alignment state during the optimization process. For each sample at time t on the validation set, we calculate its temporal alignment state s[t] = aligned_count[t] / total_pairs[t]. Here, aligned_count[t] refers to the number of sample pairs judged to be well-aligned temporally based on the model output, with the criterion being ρ(t) > 0.8. If s[t] remains below the preset threshold of 0.7, it indicates that the model has failed to effectively learn temporal alignment. In this case, enhancement strategies can be adopted: resample data pairs containing larger temporal deviations from the training set, or temporarily increase the weight γ of the regularization term L_align to strengthen the alignment constraint.

[0039] Through this iterative optimization, the model's cross-modal interaction modules are progressively calibrated. The learned feature representations are no longer misled by spurious statistical correlations (such as "a gamma pulse is always followed by an increase in gas concentration," even if the actual interval between the two is large), but instead learn to fuse information based on the true physical timeline. For example, the model can correctly correlate nearly simultaneous gamma-ray bursts with abrupt changes in the heat source (potentially pointing to a nuclear explosion), while attributing more distant increases in gas concentration to subsequent wind diffusion (rather than the explosion itself).

[0040] Ultimately, when the decision layer needs to synthesize preliminary decisions (d_gamma, d_gas, d_thermal) from multiple sources such as gamma spectrometer, gas sensor, and thermal imaging, conflicts between these source decisions are significantly reduced because the feature representations at the model's underlying level have undergone temporal alignment correction. For example, when predicting the presence of a toxic chemical leak, the model's d_gas (based on an electrochemical sensor) will not be incorrectly suppressed or amplified by a temporally misaligned temperature spike (d_thermal). A robust and accurate final decision can be obtained through a simple weighted fusion (e.g., d_final = w1·d_gamma + w2·d_gas + w3·d_thermal). Furthermore, to prevent overfitting and ensure convergence stability, an early stopping mechanism is introduced: training is terminated and the best model is saved when the loss on the validation set does not show significant improvement over three consecutive iterations.

[0041] In this invention, for step S1, the asynchronous sampling time sequence of nuclear biochemical multimodal sensors is used to construct noisy cross-modal data pairs. First, the system collects the raw data streams of three types of sensors in real time: the radiation intensity signal R[i] output by the gamma spectrometer is recorded in the form of pulse counting, generating several data points per second, each point with a timestamp t_R[i] accurate to milliseconds; the gas concentration C[j] output by the electrochemical gas sensor array is also accompanied by a timestamp t_C[j], but because the gas sensor response time is slow, its sampling frequency is usually lower than that of the gamma spectrometer; the thermal imaging module outputs the thermal field intensity image T[k] and its timestamp t_T[k] at a fixed frame rate (e.g., 30 frames per second). To construct training samples, the system sets an anchor time t_anchor (e.g., the time when emergency rescue personnel issue a query request), and defines a search window with a length of 240ms (i.e., [-120ms, +120ms]) centered on t_anchor. Within this window, the radiation intensity data point R[i], gas concentration data point C[j], and thermal imaging frame T[k] closest to t_anchor are retrieved respectively, and their respective timestamps t_R[i], t_C[j], t_T[k] and actual time differences Δt_R, Δt_C, Δt_T with t_anchor are recorded. If the absolute difference between any two of the three time differences exceeds 120ms (e.g., |Δt_R - Δt_C|>120ms), then the data set is considered to have significant asynchrony. In this case, instead of simply discarding the data, random noise perturbation σ[n] is introduced to simulate the uncertainty of a real sensor. Specifically, for radiation intensity R[i], a random noise factor σ_R uniformly distributed in the range [0.1, 1.0] is generated, and R[i] is corrected to R'[i] = R[i] × (1 + σ_R); similarly, independent random noise σ_C and σ_T are applied to the pixel values ​​of gas concentration C[j] and thermal imaging frame T[k]. The introduction of noise forces the model to learn to extract stable features from noisy data, thereby enhancing its robustness to sensor noise and environmental interference. In addition, the system sets a maximum allowable time interval T_max = 500ms. If the absolute value of the difference between the timestamp of any sensor data point and t_anchor is greater than T_max, it means that the sensor data is seriously outdated and cannot reflect the real environmental state at the current moment. At this time, the system will discard the data and wait for the next set of data that meets the timeliness requirements to arrive before constructing the sample. The above method generates cross-modal data pairs containing timestamp deviation information and noise perturbation, providing training materials that closely resemble real rescue scenarios for subsequent model fine-tuning.

[0042] In one embodiment, the specific implementation of "extracting intermediate feature representations of cross-modal interaction modules" in step S2 is as follows. First, all parameters of the pre-trained student model are set to a frozen state, that is, no gradient is calculated or parameter updates are performed during forward propagation, so as to retain its prior knowledge learned on a large general dataset. Subsequently, the noisy cross-modal data pairs (including radiation intensity, gas concentration, and thermal imaging frames) constructed in step S1 are input into the student model in batches. When the data flows through the cross-modal interaction module of the model, such as a multi-head cross-attention layer, this embodiment sets a cutoff point at the output of the layer to obtain the output feature tensor F[m] of the m-th layer (e.g., the first cross-attention layer). Assuming the current batch size is B=16 and the feature dimension is D=512, the shape of F[m] is [16, 512]. In order to eliminate the interference caused by the amplitude difference between different samples, F[m] is normalized in terms of channel dimension. Specifically, for each feature channel l (l=1,...,512), its mean μ_l and standard deviation σ_l over all 16 samples in the current batch are calculated. Then, the feature values ​​of each sample are standardized: F_norm[m][b, l]= (F[m][b, l]- μ_l) / σ_l, where b represents the b-th sample in the batch. The standardized feature F_norm[m] has zero mean and unit variance, which facilitates subsequent weighting. Next, in order to highlight the feature channels most important for multimodal alignment, this embodiment introduces a lightweight attention mechanism: F_norm[m] is input into an attention generation module consisting of a single-layer fully connected network and a Sigmoid activation function. This module outputs a weight value A[l] for each feature channel, ranging from 0 to 1, representing the importance of that channel in the current cross-modal interaction task. The formula for calculating the attention weights can be expressed as A = σ(W·F_norm[m]+ b), where W and b are learnable parameters (although the model parameters are frozen at this point, the attention module is dynamically calculated during forward propagation as part of feature extraction and does not participate in gradient updates itself). Finally, the standardized features are multiplied element-wise with the corresponding attention weights and summed to obtain the base representation F_base for each sample: F_base[b] = Σ[l=1 to 512](A[l] · F_norm[m][b, l]). F_base is a 512-dimensional vector that encapsulates the core understanding of the current input data from the cross-modal interaction module. It contains information from multimodal fusion and strengthens the focus on key features through the attention mechanism, providing accurate anchors for subsequent temporal alignment regularization.

[0043] In one embodiment, the specific implementation of "constructing a time-series flexible alignment regularization term based on the alignment relationship between intermediate feature representations and original time-series data" in step S3 is as follows. First, to dynamically evaluate the time-series matching degree of cross-modal features, an alignment coefficient ρ(t) based on a sliding window is introduced. Specifically, for a continuous time series segment, a sliding window of length L=10 is set, containing multiple consecutive cross-modal data pairs and their corresponding model base representations F_base(t). For each pair of modal data within the window (e.g., radiation intensity sequence R[i] and gas concentration sequence C[j]), the minimum cumulative distance between the two sequences is calculated using the Dynamic Time Warping (DTW) algorithm. The reciprocal of this distance is used as the alignment coefficient ρ(t) at that moment. The larger the value of ρ(t), the higher the time-series matching degree between the sequences. Next, an elastic alignment function E is constructed, which comprehensively considers two dimensions: time deviation and semantic similarity: E = α·|Δt|+ β·cos(θ). Where |Δt| is the absolute value of the actual maximum timestamp deviation of the current data pair (e.g., the time difference between radiation and gas), α is a preset time offset weighting parameter, which is taken as α=0.5 in this embodiment; θ represents the cross-modal feature similarity angle, which is specifically calculated by extracting the feature sub-vectors F_rad corresponding to the radiation mode and F_gas corresponding to the gas mode from the basic representation F_base, and calculating the cosine value of the angle between these two vectors in high-dimensional space cos(θ) = (F_rad·F_gas) / (|F_rad|·|F_gas|). The closer this value is to 1, the more similar the features of the two modes are semantically. β is a semantic consistency weighting parameter, which is taken as β=0.8 in this embodiment. The design intention of this function is that when the physical time deviation |Δt| is large, even if the features of the two modes are semantically similar (cos(θ) is large), the value of function E will increase due to the time penalty term, thereby guiding the model to reduce the association weight of this pair of modes; conversely, if the physical time deviation is very small, it encourages the model to increase the semantic similarity of its feature representations. Subsequently, for all N=16 samples in the current training batch, the average elastic loss L_align = (1 / 16) Σ[i=1 to 16] E[i] is calculated. In order to adaptively control the strength of regularization, a threshold ε is set, which ranges from [0.5, 1.0]. In this embodiment, ε=0.7 is taken.After each forward propagation, the system checks if L_align is greater than 0.7. If it is, it indicates a significant temporal misalignment in the current batch of samples, requiring regularization. In this case, L_align is added to the final fine-tuning objective function with a weight γ = 0.2, resulting in a total loss L_total = L_task + 0.2·L_align. If L_align ≤ 0.7, the current batch is considered to have good temporal alignment, and no additional regularization is needed; in this case, L_total = L_task. This adaptive mechanism avoids over-regularization on samples that are already temporally aligned, preserving the model's existing effective knowledge.

[0044] In one embodiment, the specific implementation of "adaptive alignment of multimodal temporal sequences through iterative optimization" in step S4 is as follows. During the fine-tuning phase, some parameters involved in cross-modal interaction and decision output in the student model (such as the cross-attention layer and classification head) are first unfrozen, while the underlying feature extraction network remains frozen to balance knowledge retention and task adaptation. A learning rate η is set, ranging from [0.001, 0.01], and in this embodiment, η = 0.005 is used. The AdamW optimizer is employed for parameter updates. Simultaneously, the maximum number of iterations, Iter_max, is set to 100 epochs to prevent infinite training. In each iteration cycle, the system calculates the current temporal alignment state s[t] = aligned_count[t] for the samples at each time step t. / total_pairs[t], where total_pairs[t] is the total number of cross-modal data pairs at time t, and aligned_count[t] is the number of pairs that are determined to be well aligned by the dynamic alignment coefficient ρ(t) (the criterion is ρ(t)>0.8). If s[t]<0.7, it indicates that the current model's temporal alignment of the samples at that time is not sufficient. At this time, the system triggers an enhancement strategy: on the one hand, it resamples a batch of samples with larger time deviations (e.g., |Δt|>80ms) from the training set to increase the alignment difficulty. This forces the model to learn more discriminative temporal features. On the other hand, the regularization term weight γ is temporarily increased, for example, from 0.2 to 0.4, to strengthen the proportion of alignment constraints in the total loss. During optimization, an early stopping mechanism is used to monitor the loss decrease on the validation set: if the validation loss does not improve significantly for three consecutive iterations (decreasing by less than 1%), training is terminated early, and the best-performing model parameters on the validation set are saved. Through these dynamic monitoring and adaptive adjustment strategies, the model can converge efficiently with a small number of labeled samples, ensuring the stability and generalization ability of temporal alignment.

[0045] Furthermore, this application also constructs an adversarial training mechanism based on pseudo-causal association interference. Specifically, firstly, all potential pseudo-causal features F[p] are extracted from cross-modal feature pairs. These features typically manifest as frequent temporal co-occurrence of certain modal combinations but lacking physical causal relationships (e.g., a small fluctuation in gas concentration is always followed by a gamma pulse, but it is actually caused by external environmental disturbances rather than nuclear biochemical events). To quantify the strength of this pseudo-association, an evaluation index is designed: score_pseudo = (1 / k) Σ[l=1 to k][Cj / Ri], where k is the number of consecutive time points, and Ri and Cj represent the sampled values ​​of radiation intensity and gas concentration within the same time window, respectively. This ratio reflects the statistical dependence between radiation events and gas responses. A significance threshold λ is set, ranging from [0.6, 0.9], and in this embodiment, λ=0.75 is used. If score_pseudo>λ, then the feature pair is considered to have a strong pseudo-causal association. At this point, the system automatically introduces an adversarial perturbation ξ[i], which is generated based on the gradient direction of the current feature F[p] and aims to enhance the model's sensitivity to spurious associations. Specifically, during the model's forward propagation, a small perturbation is applied to the extracted spurious causal feature F[p]: F_adv[p] = F[p] + ε·sign( _F[p] L_task), where ε is the perturbation magnitude (0.01 in this embodiment). Subsequently, F_adv[p] is fed into an auxiliary pseudo-causal classifier via a Gradient Reversal Layer (GRL), which attempts to predict the presence or absence of pseudo-correlation. During backpropagation, the GRL reverses the gradient sign, causing the student model's backbone network to update in the direction of maximizing the pseudo-causal classifier loss. This guides the student model to ignore features easily exploited by pseudo-correlation and instead focus on truly causal modal interactions. Through this adversarial training, the model can effectively resist the misleading influence of pseudo-correlation features when facing unknown scenarios, significantly improving the robustness of decision-making.

[0046] Furthermore, to mitigate multi-source decision conflicts, this application first constructs a triplet decision graph, where the three nodes represent independent decision results from the gamma spectrometer, gas sensor, and thermal imaging module: d_gamma (radiation event classification result), d_gas (gas type and concentration level), and d_thermal (heat source hazard level). Each decision result is represented as a probability vector. A decision consistency weight W_d = count_c / count_a is defined, where count_c represents the number of times two decision results occur simultaneously in historical training data (co-occurrence count), and count_a represents the total number of occurrences of all possible decision combinations. This weight reflects the historical co-occurrence probability between decision pairs, used to assess the rationality of the current decision. For example, if d_gamma and d_gas rarely occur simultaneously in the training data, but occur simultaneously with high confidence at the current moment, it may indicate a conflict. A consistency threshold of 0.8 is set; if W_d < 0.8, the current decision combination is considered to have an abnormal conflict. At this point, the system smooths out abnormal decisions: decisions with low confidence are downweighted, and an additional penalty term L_conflict = ω · MSE(d_exp, d_pred) is added, where d_exp is the expected decision distribution based on historical patterns, d_pred is the predicted distribution of the current model, and ω is the penalty weight (0.3 in this embodiment). This penalty term guides the model to adjust parameters through backpropagation to reduce the occurrence of similar conflicts in the future. Finally, the comprehensive multi-source decision result d_final is output, using a weighted summation method: d_final = w1·d_gamma + w2·d_gas + w3·d_thermal, where the weights w1, w2, and w3 can be dynamically adjusted according to the sensor signal-to-noise ratio or task requirements (w1=0.4, w2=0.4, w3=0.2 in this embodiment). Through the above mechanism, the model can intelligently arbitrate when multi-source information is inconsistent, outputting a more reliable and stable final decision, significantly improving the comprehensive judgment capability in nuclear, biological, and chemical emergency rescue scenarios.

[0047] Furthermore, this embodiment introduces a dynamic balancing mechanism to adaptively adjust the weight of the temporal alignment regularization term at different training stages. Specifically, a dynamic balancing parameter λ(t) = exp(t / τ) is defined, where t is the current epoch, and τ is a decay constant, ranging from 5 to 10; in this embodiment, τ = 10. This function ensures that λ(t) gradually decays from close to 1 as the training progresses, making λ(t) larger in the early stages of training (when t is small), thus emphasizing the role of the temporal alignment regularization term and forcing the model to prioritize learning the correct temporal relationships across modal data. As training progresses to the later stages (when t increases), λ(t) gradually decreases, and the model shifts its focus to optimizing the overall prediction accuracy. In each round of optimization, the weight of the regularization term is gradually adjusted based on λ(t). Specifically, in the total loss function, the weight coefficient γ(t) = γ_base · λ(t) of the temporal alignment loss L_align is used, where γ_base is the base weight (0.3 in this embodiment). In the feature fusion stage, each feature vector is weighted and concatenated, expressed as F_joint = (1-λ(t))·F_base + λ(t)·F_align, where F_base is the base representation and F_align is the aligned feature representation. This weighted concatenation method ensures that in the early stages of training, the model primarily relies on the aligned features for decision-making, while in the later stages, the original base representations are gradually fused to retain more original information. Furthermore, to ensure model robustness, a random dropout strategy is employed to perform protective pruning on highly volatile feature subspaces: in each training batch, a portion of feature dimensions are randomly dropped with probability p_drop=0.2 to prevent the model from over-relying on certain noise-sensitive feature channels, thereby enhancing generalization ability.

[0048] Furthermore, this embodiment introduces a temporal convolutional network and a bidirectional long short-term memory network to enhance temporal modeling capabilities when constructing the temporal elastic alignment regularization term. Specifically, in step S3, in addition to using dynamic time warping to calculate the alignment coefficient, a temporal convolutional network (TCN) is also used to extract hierarchical features from cross-modal data. The TCN is composed of multiple dilated convolutional layers stacked together, which can capture dependencies at different time scales and output temporal feature representations of each modality. Subsequently, these temporal features are input into a bidirectional long short-term memory network (Bi-LSTM) model. The Bi-LSTM effectively estimates the long-term temporal dependencies between features through hidden state encoding in both forward and backward directions, and outputs the fused features at each time step. Based on the output of Bi-LSTM, the alignment error is calculated using the formula E[i] = √[(t_g - t_a)² +(score_g - score_a)²], where t_g is the timestamp of the gamma spectrometer data, t_a is the timestamp of the electrochemical gas sensor data, and score_g and score_a are the confidence scores of the corresponding modes (obtained by normalizing the classification probability or regression value output by Bi-LSTM). This error formula comprehensively considers time deviation and confidence difference. When two modes are close in time and have consistent confidence, E[i] is smaller, and vice versa. An error threshold ε is set (ε=0.8 in this embodiment). When E[i] is greater than ε, it indicates that there is a significant time shift or confidence mismatch in the current cross-modal pair. The system automatically triggers a negative feedback mechanism: during backpropagation, the alignment gradient direction of the sample is reversed, that is, the model is forced to update parameters in the direction that reduces E[i], thereby correcting the temporal alignment direction and effectively reducing erroneous causal associations caused by time shift.

[0049] To further enhance the model's adaptability in complex environments, this embodiment introduces a physical environment perception factor in the alignment error calculation. Specifically, during the data acquisition and model inference phases, the physical parameters of the on-site environment are acquired simultaneously: ambient temperature (Temp_env), ranging from [0℃, 40℃], which is read in real time by a temperature sensor deployed on-site; relative humidity (RH), ranging from 0% to 100%, acquired by a humidity sensor; and electromagnetic noise intensity (ENI), normalized to [0, 200], measured by a spectrum analyzer or a dedicated electromagnetic interference sensor to determine the background noise level. The weighting factor ω_env = 1 + RH / 100 + ENI / 200 is calculated based on the environmental complexity. This factor reflects the impact of environmental severity on the reliability of sensor data: higher humidity and stronger electromagnetic noise result in a larger ω_env, meaning stricter alignment constraints are needed to overcome environmental interference. The alignment error formula in claim 9 is modified to E[i] = ω_env ·√[(t_g - t_a)² + (score_g - score_a)²], which multiplies the environmental weight factor by the original error value. This results in a greater alignment loss for the same time-series deviation and confidence difference under harsh environments, thus forcing the model to pay more attention to alignment accuracy. Furthermore, in environments with strong interference (e.g., ω_env > 2.0), the system automatically activates a multimodal fusion enhancement strategy. Specifically, this involves dynamically adjusting the fusion weights of each modality feature, reducing the weights of single-modal features that are more susceptible to environmental influences (e.g., electrochemical gas sensors are prone to drift under high temperature and humidity), while increasing the weights of relatively stable modes (e.g., thermal imaging can still work stably under electromagnetic interference). This reduces the risk of overall judgment distortion caused by single-modal instability and ensures the robustness of the decision.

[0050] Furthermore, this embodiment introduces a scenario context-aware mechanism when constructing multimodal cross-temporal alignment regularization terms to dynamically adjust alignment constraints based on different types of nuclear, biological, and chemical (NBC) events. First, a pre-stored emergency event scenario identifier S = {S1, S2, ..., Sn} is loaded using a prior knowledge base. Each Si represents a different NBC scenario type, such as S1 for radioactive material leakage (primarily gamma radiation), S2 for toxic gas diffusion (primarily electrochemical gas sensor response), and S3 for fire and explosion (primarily thermal imaging anomalies). This knowledge base is built based on historical rescue cases and expert experience, storing prior information on the modal importance under each scenario. During model fine-tuning and inference, the system initially determines the current scenario type using real-time sensing data (e.g., if radiation intensity increases significantly but gas concentration does not change significantly, it is determined to be scenario S1). Based on the determined scenario Si, a contextualized weight W_s = [w_gas, w_thermal, w_gamma] is constructed. This weight vector reflects the relative importance of each modality in the current scenario. For example, in scenario S1, the weights can be set to w_gamma=0.7, w_gas=0.2, w_thermal=0.1; in scenario S2, they are adjusted to w_gamma=0.2, w_gas=0.6, w_thermal=0.2. Subsequently, a multi-task objective function is constructed: L_total = λ·L_main + (1-λ)·L_align(Si), where L_main is the main task loss (e.g., classification loss), and L_align(Si) is a temporal alignment regularization term based on scenario Si, calculated by multiplying the original L_align by the contextualized weights W_s for modal weighting. The parameter λ is a balancing factor, whose value is dynamically adjusted based on the real-time detected scenario: when the scenario features are clear (e.g., extremely strong radiation signals), the value of λ is appropriately reduced to emphasize the main task; when the scenario is ambiguous or contains ambiguity, the value of λ is increased to strengthen the alignment constraint. Through this scene-aware dynamic adjustment mechanism, the model can flexibly balance the relationship between feature alignment and task prediction under different nuclear biochemical events. While ensuring alignment accuracy, it avoids weakening key modal information in specific scenarios due to excessive regularization, and ultimately achieves the best balance between model flexibility and accuracy.

[0051] The present invention provides a method for rapid fine-tuning of a small-sample distillation model in the field of nuclear, biological, and chemical emergency rescue. The method includes: First, constructing noisy cross-modal data pairs based on asynchronously sampled time-series data from nuclear, biological, and chemical multimodal sensors (gamma spectrometer, electrochemical gas sensor, and thermal imaging equipment); then, freezing the parameters of a pre-trained student model and extracting intermediate feature representations of its cross-modal interaction modules as basic representations; next, utilizing the alignment relationship between this basic representation and the original asynchronous time-series data, constructing a time-series elastic alignment regularization term, and introducing this regularization term into the model's fine-tuning loss function, thereby guiding the model to adaptively align the time-series relationships between different modalities with limited labeled samples; finally, improving the model's cross-modal consistency through iterative optimization to alleviate multi-source decision-making conflicts caused by pseudo-causal associations between multimodal data.

[0052] Specifically, this invention addresses the problem of spurious causal interference caused by inaccurate alignment of asynchronous multimodal data by constructing a temporal elastic alignment regularization term. In nuclear, biological, and chemical emergency rescue, data collected by different sensors exhibit significant timestamp deviations (±120 ms), which leads to difficulties in aligning information across modalities and may result in misjudgments or decision conflicts due to time misalignment. Traditional alignment methods are often unable to adapt to such large time deviations, causing the model to learn incorrect modal associations. The temporal elastic alignment regularization term designed in this method measures the alignment degree of the student model between different modal data and constrains and optimizes it during training. This enables the model to more accurately identify and fuse the true relationships between modalities even with low sample sizes, effectively suppressing interference from spurious causal associations, improving the model's robustness and accuracy in complex environments, and thus achieving efficient, accurate, and stable multi-source decision-making.

[0053] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this invention, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency response, characterized in that, include: Noisy cross-modal data pairs were constructed based on asynchronous sampling time series of nuclear biochemical multimodal sensors; The parameters of the pre-trained student model are frozen, and the intermediate feature representations of its cross-modal interaction modules are extracted as the basic representations. Based on the alignment relationship between the intermediate feature representation and the original asynchronous time series data, a time series elastic alignment regularization term is constructed and introduced into the model fine-tuning loss function; Iterative optimization enables the student model to adaptively align with multimodal time series with a small number of labeled samples, mitigating multi-source decision-making conflicts caused by spurious causal relationships.

2. The method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency rescue as described in claim 1, characterized in that, The construction of noisy cross-modal data pairs based on asynchronous sampling time series of nuclear biochemical multimodal sensors includes: The radiation intensity signal R[i], gas concentration C[j], and thermal field intensity T[k] in the environment were collected by a gamma spectrometer, an electrochemical gas sensor, and a thermal imaging module, respectively. Construct cross-modal data frame pairs (x[i], x[j], x[k]) based on timestamp deviation Δt, where x[i], x[j], and x[k] represent time series segments of the corresponding modalities, and Δt ranges from -120 ms to +120 ms. Different modal data are non-strictly synchronized, and noise σ[n] is introduced to perturb the cross-modal data pairs. The value range of σ[n] is [0.1, 1.0]. If the current sampling time t is greater than the maximum allowed time interval Tmax, then discard the data set; otherwise, retain it for subsequent processing.

3. The method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency rescue as described in claim 1, characterized in that, The extraction of intermediate feature representations for cross-modal interaction modules includes: Perform forward propagation on the student model on the training data and extract the feature output F[m] from the m-th layer; Convert the feature output F[m] into a normalized vector, in the form F_norm[m] = (F[m] - μ) / σ, where μ is the channel mean and σ is the channel standard deviation; Generate independent cross-modal attention weights A[l] for each feature channel; We obtain the basic representation F_base = Σ[l](A[l] · F_norm[m]) by weighted summation of the feature vector F_norm[m].

4. The method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency rescue as described in claim 1, characterized in that, The construction of the time-series flexible alignment regularization term based on the alignment relationship between the intermediate feature representation and the original time-series data includes: The dynamic alignment coefficient ρ(t) of cross-modal features is calculated by sliding window matching; Define the elastic alignment function E = α·|Δt| + β·cos(θ), where α and β are the time offset and semantic consistency weighting parameters, respectively, and θ represents the cross-modal feature similarity angle; For each batch of feature pairs, calculate the elasticity loss term L_align = (1 / N)Σ[i=1 to N] E[i]; The regularization term is added to the final fine-tuning objective function only if L_align is greater than the threshold ε (ε ranges from [0.5, 1.0]).

5. The method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency rescue as described in claim 1, characterized in that, The iterative optimization of adaptive aligned multimodal timing includes: Set the learning rate η and the maximum number of iterations Iter_max, where the value of η is in the range of [0.001, 0.01]. For each time t, calculate the current timing alignment state s[t] = aligned_count[t] / total_pairs[t]; If s[t] < 0.7, it indicates that the temporal alignment is insufficient. In this case, the feature pairs should be resampled and the alignment constraints should be strengthened. The early stopping mechanism is used to monitor and verify the rate of loss reduction. Training ends when there is no significant improvement for three consecutive iterations.

6. The method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency rescue as described in claim 1, characterized in that, It also includes adversarial training mechanisms built on pseudo-causal association interference, including: Extract the latent pseudo-causal features F[p] from all cross-modal feature pairs; The strength of pseudo-causal associations is evaluated using the formula score_pseudo = (1 / k)Σ[l=1 to k] [Cj / Ri], where k is the number of time points. If score_pseudo is greater than λ (λ is the set significance threshold, with a value range of [0.6, 0.9]), then an adversarial perturbation ξ[i] is introduced; By using a gradient inversion layer to guide the student model to ignore the influence of spurious correlation features, the robustness of decision-making is improved.

7. The method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency rescue as described in claim 1, characterized in that, include: Construct a triplet decision graph, where nodes represent multi-source decision outcomes d_gamma, d_gas, and d_thermal; Define the decision consistency weight W_d = count_c / count_a, where count_c represents the number of co-occurrences and count_a is the total number of occurrences of any decision combination; If W_d < 0.8, then the abnormal decision is smoothed and an additional penalty term L_conflict = ω ·MSE(d_exp, d_pred) is added; The final output is the integrated multi-source decision result d_final = w1·d_gamma + w2·d_gas + w3·d_thermal.

8. The method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency rescue as described in claim 1, characterized in that, The iterative optimization of adaptively aligned multimodal timing also includes: Define the dynamic equilibrium parameter λ(t) = exp(t / τ), where τ is the decay constant; In each round of optimization, the weight of the regularization term is gradually adjusted according to λ(t), emphasizing time alignment in the early stage and focusing on overall prediction accuracy in the later stage. The weighted concatenation of each feature vector is expressed as F_joint = (1-λ(t))·F_base + λ(t)·F_align; A random dropout strategy is used to perform protective pruning on the high-fluctuation feature subspace to ensure model robustness.

9. A method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency rescue as described in claim 8, characterized in that, The time-series elastic alignment regularization term constructed based on intermediate feature representation and original asynchronous time series includes: Use temporal convolutional networks to extract hierarchical features from cross-modal data; Construct a bidirectional long short-term memory network model to estimate the temporal dependencies between features; The alignment error is calculated using the formula E[i] = √[(t_g - t_a)² + (score_g - score_a)²], where t_g is the timestamp of the gamma spectrometer, t_a is the timestamp of the electrochemical sensing data, and score_g and score_a are the corresponding confidence scores. When E[i] is greater than the preset threshold ε, a negative feedback mechanism is forcibly added to correct the timing alignment direction and reduce erroneous associations caused by time offset.

10. A method for rapid fine-tuning of a small-sample distillation model for nuclear, biological, and chemical emergency rescue as described in claim 9, characterized in that, The alignment error calculation method introduces a physical environment-aware factor, specifically including: Obtain the ambient temperature Temp_env (value range [0℃, 40℃]), relative humidity RH, and electromagnetic noise intensity ENI; The weighting factor ω_env is calculated based on environmental complexity: ω_env = 1 + RH / 100 + ENI / 200. The alignment error formula is modified to E[i] = ω_env · [(t_g - t_a)² + (score_g - score_a)²]; In environments with strong interference, a multi-modal fusion enhancement strategy is automatically activated to reduce the risk of judgment distortion caused by the instability of a single mode.

Images