A method for dynamic evaluation and early warning of refrigerated cargo transport quality

Abstract

The application provides a kind of perishable goods frozen transport quality dynamic evaluation and early warning method, belong to frozen product transportation technical field, the application is based on three-layer discrete wavelet decomposition and adaptive threshold soft cut-off noise reduction based on Dab wavelet base function is carried out to the original time sequence signal from multiple sensors, four-way heterogeneous time sequence signals after denoising are sent into four independent asynchronous computing pipelines connected with lock-free ring buffer, the forgetting gate coefficient of local asymmetric gate control cycle unit in pipeline and Arrhenius temperature integral item are coupled explicitly to force to retain long-range historical characteristics of quality deterioration, the characteristic tensor of four pipelines is fused by mixed range attention fusion layer to be weighted and fused with Markov state transition probability, combined with hierarchical early warning threshold, trigger early warning instruction, solve the technical problem that long-range feature real-time fusion reasoning of quality deterioration cannot be realized under the condition of asynchronous arrival of multi-source heterogeneous sensor signals.

Description

Technical Field

[0001] This invention belongs to the field of frozen goods transportation technology, and specifically relates to a method for dynamic assessment and early warning of the quality of frozen transport of perishable goods. Background Technology

[0002] In the frozen transport of perishable goods, various sensors, including those for temperature, humidity, vibration, and gas concentration, are widely deployed at cold chain nodes. Existing quality assessment methods often employ a synchronous batch processing framework, waiting for all sensor data to arrive completely within the same clock cycle before uniformly performing signal denoising, feature extraction, and fusion inference, triggering warning commands with a fixed threshold. However, in current cold chain transport scenarios, due to fundamental differences in the sampling frequencies of different sensors (vibration sensors have a much higher sampling frequency than gas concentration sensors), the synchronous waiting mechanism will continuously accumulate inference lags due to the arrival delay of the slowest sensor's data. Furthermore, the fixed threshold denoising method cannot adapt to changes in noise intensity across different transport segments. The forget gate in standard gated loop units will quickly forget deterioration features after a quality degradation event due to subsequent normal data input, resulting in the loss of long-term deterioration history features at the fusion layer. In other words, existing technologies suffer from the technical problem of being unable to achieve real-time fusion inference of long-term quality degradation features under asynchronous arrival conditions from multi-source heterogeneous sensor signals. Summary of the Invention

[0003] In view of this, the present invention provides a method for dynamic evaluation and early warning of the quality of perishable goods transported in frozen containers, which can solve the technical problem in the prior art that the long-range characteristics of quality deterioration cannot be fused and inferred in real time under asynchronous arrival conditions of multi-source heterogeneous sensor signals.

[0004] This invention is implemented as follows: This invention provides a method for dynamic assessment and early warning of the quality of perishable goods transported under refrigeration, comprising the following steps:

[0005] The raw time-series signals from temperature, humidity, vibration and gas concentration sensors are subjected to three-level discrete wavelet decomposition in sliding window mode. Adaptive threshold soft truncation is applied to the detail coefficients of each level to gradually reconstruct the denoised time-series signal and output the signal quality index.

[0006] The four heterogeneous time-series signals after denoising are sent into four independent asynchronous computing pipelines. Each pipeline uses an unlocked ring buffer to pass feature tensors. The forget gate coefficient of the local asymmetric gated loop unit in the pipeline is explicitly coupled with the Arrhenius temperature integral term. When the accumulated temperature integral exceeds the activation energy threshold, the forget gate coefficient is hard-suppressed, forcibly preserving the long-term historical features of quality deterioration.

[0007] The feature tensors output by the four asynchronous pipelines are fed into the hybrid range attention fusion layer. The local sliding window attention submodule and the global anchor attention submodule are computed in parallel. The correlation matrix output by the two is weighted and fused by the weight coefficients determined in real time by the Markov state transition probability.

[0008] The features output by the fusion layer are processed by a dynamic hierarchical early exit mechanism. The three confidence evaluation classifiers judge the quality status in turn. When the confidence exceeds the confidence threshold, the quality score is directly output. Otherwise, it is passed to the next classifier.

[0009] Based on the quality score and the predicted rate of quality deterioration, combined with the deterioration threshold corresponding to the type of goods, a graded early warning instruction is triggered, and the early warning instruction is pushed to the cold chain monitoring terminal in real time.

[0010] The dynamic forget gate adjustment function continuously monitors the gating weight distribution of each asynchronous pipeline, and adaptively adjusts the forget gate decay parameter according to the interval to which the function output value belongs.

[0011] The three-layer discrete wavelet decomposition uses the Dab wavelet basis function, and the adaptive threshold is dynamically generated by multiplying the standard deviation of the approximation coefficient by a learnable scaling factor. The learnable scaling factor is jointly optimized with the main network parameters.

[0012] The signal quality index is calculated from the energy retention ratio of the reconstructed signal. When the signal quality index is lower than the signal quality index threshold, a dynamic backtracking mechanism is triggered to force the feature weights of the current time step to decay to the historical average level.

[0013] The lock-free circular buffer uses a fixed-capacity circular address space as its storage structure. The producer thread and the consumer thread each hold independent write pointers and read pointers. Atomic operations ensure thread safety for pointer updates. The buffer capacity is dynamically adjusted according to the ratio of the sampling frequencies of each sensor.

[0014] The forward path of the locally asymmetric gated loop unit is equipped with a fast decay update gate, and the update coefficient is calculated by normalizing the absolute value of the difference between the current input and the historical hidden state. The feedback path is equipped with a slow accumulation forget gate, and the forget coefficient is explicitly coupled with the Arrhenius temperature integral term. When the accumulated temperature integral exceeds the activation energy threshold, the forget gate coefficient is forced to be reduced to near zero.

[0015] The activation energy threshold is determined by the biochemical reaction activation energy parameter corresponding to the type of goods. The specific value is determined by fitting the Arrhenius equation through laboratory accelerated deterioration experiments, with the goal of minimizing the residual between the measured quality sensory score and the fitted rate curve.

[0016] In the hybrid range attention fusion layer, the local sliding window attention submodule calculates the dense attention weights in the temporal neighborhood with the sliding window step size, and the global anchor attention submodule selects learnable global anchor tokens to participate in the sparse association calculation of all features. The association matrix output by both is fused by weighted summation of the weight coefficients determined in real time by the Markov state transition probability matrix.

[0017] The Markov state transition probability matrix includes the states of the transportation route, such as port loading and unloading, highway trunk lines, cold storage, and the final delivery stage. The state switching probability is estimated by statistically analyzing the switching frequency between states in the historical transportation trajectory. The matrix elements are continuously updated with new records added within the sliding time window as increments.

[0018] In the dynamic hierarchical early exit mechanism, the number of parameters of the three confidence evaluation classifiers increases hierarchically. The first classifier adopts a single-layer linear projection structure, and the second and third classifiers successively add fully connected hidden layers.

[0019] The graded early warning instructions are divided into the following levels: green normal state when the quality score is higher than the normal threshold, yellow attention state when the quality score is between the normal threshold and the deterioration warning threshold, and red warning state when the quality score is lower than the deterioration warning threshold.

[0020] The input variables of the dynamic forget gate adjustment function include the temperature integral accumulation at the current time step, the signal quality index, and the mean forget gate output of the current pipeline. The output is the adjustment intensity index, calculated as follows: , where the coefficient , , The sum is 1.

[0021] When the adjustment intensity index reaches its highest range, the forget gate decay parameter is forcibly reduced to near zero, and the dynamic hierarchical early exit mechanism is triggered to force the high-risk warning level to be output directly from the first classifier.

[0022] The establishment of the training dataset includes: collecting historical data of cold chain transportation; generating augmented samples for a few types of deteriorated samples using a conditional generative adversarial network until the ratio of positive to negative samples is not lower than the lower limit of the positive to negative sample ratio; then applying a pre-processing streaming signal noise reduction pipeline and removing samples whose signal quality index is lower than the lower limit of the signal quality index; and dividing the training set, validation set and test set into layers according to the type of goods.

[0023] The training process employs a dynamic focus loss function, which assigns an adaptive amplification factor to the gradient contribution of minority class degraded samples. The amplification factor is dynamically calculated based on the ratio of positive to negative samples in the current mini-batch. The training loss function consists of a linear superposition of three parts: the mean square error term of the predicted quality score, the mechanism residual penalty term of the Arrhenius rate equation, and the stability regularization term of the feedback weight of the locally asymmetric gated recurrent unit.

[0024] The signal quality index threshold ranges from 0.75 to 0.92; the confidence threshold ranges from 0.82 to 0.95. The value ranges from 0.3 to 0.6. The value range is 0.1 to 0.4. The value range is 0.1 to 0.3; the normal threshold range is 0.70 to 0.88; the deterioration warning threshold range is 0.40 to 0.65; the lower limit of the positive and negative sample ratio is 1:5; the ratio of training set, validation set and test set is 7:1.5:1.5.

[0025] This invention constructs an asynchronous parallel multi-channel temporal fusion model based on biochemical mechanism constraints. Four independent asynchronous computation pipelines are connected by a lock-free circular buffer, enabling parallel feature extraction from sensor data with different sampling frequencies without waiting for a unified clock synchronization. This fundamentally eliminates the inference latency accumulated in synchronous batch processing frameworks due to waiting for the slowest sensor. The forget gate coefficient of the locally asymmetric gated loop unit is explicitly coupled with the Arrhenius temperature integral term, forcibly suppressing the forget gate when the accumulated temperature integral exceeds the activation energy threshold, ensuring that deteriorated long-range historical features are not dissipated by subsequent normal data. The hybrid range attention fusion layer fuses locally dense attention and globally sparse attention in parallel, allowing the model to simultaneously perceive the impact of short-term temperature fluctuations and long-distance transportation segment switching on quality. In summary, this invention solves the technical problem mentioned in the background art of being unable to achieve real-time fusion inference of long-range features of quality deterioration under asynchronous arrival conditions from multi-source heterogeneous sensor signals. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a line graph showing the change of the forgetting gate coefficient of each sensor pipeline with transportation time.

[0028] Figure 3 This is a line graph showing how the local and global weight coefficients of the mixed-range attention fusion layer change with the state switching of the transportation segment.

[0029] Figure 4 This is a graph showing the change in cargo quality score over transportation time and the distribution of warning levels.

[0030] Figure 5This is a line graph showing the variation of the regulation intensity index output by the dynamic forgetting gate regulation function with transport time.

[0031] Figure 6 This is a distribution chart showing the sample diversion ratio of the three confidence assessment classifiers at each stage of transportation. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0033] like Figure 1 The diagram shown is a flowchart of a method for dynamic assessment and early warning of the quality of perishable goods transported under frozen conditions, provided by the present invention. This method includes the following steps:

[0034] S01. Perform three-level discrete wavelet decomposition on the raw time-series signals from temperature, humidity, vibration and gas concentration sensors in sliding window mode, apply adaptive threshold soft truncation processing to the detail coefficients of each level, gradually reconstruct the denoised time-series signal, and output the signal quality index.

[0035] S02. The four heterogeneous time-series signals after noise reduction are sent into four independent asynchronous computing pipelines. Each pipeline transmits the feature tensor through an unlocked ring buffer. The forget gate coefficient of the local asymmetric gated loop unit in the pipeline is explicitly coupled with the Arrhenius temperature integral term. When the accumulated temperature integral exceeds the cargo activation energy threshold, the forget gate coefficient is forcibly suppressed to retain the long-term historical characteristics of quality deterioration.

[0036] S03. The feature tensors output by the four asynchronous pipelines are fed into the hybrid range attention fusion layer. The local sliding window attention submodule and the global anchor attention submodule are computed in parallel. The correlation matrix output by the two is weighted and fused by the weight coefficients determined in real time by the Markov state transition probability.

[0037] S04. The features output by the fusion layer are processed by the dynamic hierarchical early exit mechanism. The three confidence evaluation classifiers judge the quality status in turn. Samples with obviously normal or obviously abnormal status are directly output as quality scores at the first classifier, and the remaining samples are passed to the subsequent classifiers step by step.

[0038] S05. Based on the quality score and the predicted rate of quality deterioration, combined with the deterioration threshold corresponding to the type of goods, a graded early warning instruction is triggered. The early warning level is determined by the interval to which the quality score belongs and the rate of deterioration. The early warning instruction is pushed to the cold chain monitoring terminal in real time.

[0039] S06. The dynamic forget gate adjustment function continuously monitors the gating weight distribution of each asynchronous pipeline. Based on the interval to which the function output value belongs, the forget gate decay parameter of the biochemical mechanism-constrained asynchronous parallel multi-channel time-series fusion model is adaptively adjusted to maintain the balance between inference accuracy and computational resource consumption.

[0040] The three-layer discrete wavelet decomposition uses the Darbu wavelet basis function to ensure time-frequency localization. The first layer decomposition splits the original signal into an approximate coefficient sequence and a detail coefficient sequence. An adaptive threshold soft truncation is applied to the detail coefficient sequence. The threshold is dynamically generated by multiplying the standard deviation of the approximate coefficients by a learnable scaling factor, which is jointly optimized with the main network parameters during the training of the biochemically constrained asynchronous parallel multi-channel time-series fusion model. The second and third layers repeat the above decomposition process, forming a coefficient pyramid with three levels of frequency resolution. After denoising, inverse wavelet reconstruction is performed on the approximate coefficient sequences of each layer to restore the denoised time-series signal. The signal quality index is calculated from the energy retention ratio of the reconstructed signal. When the signal quality index is lower than the experimentally calibrated threshold, the dynamic backtracking mechanism within the biochemically constrained asynchronous parallel multi-channel time-series fusion model is triggered, forcibly decaying the feature weights of the current time step to the historical average level to prevent low-quality signals from contaminating subsequent inference. The suggested threshold range for the signal quality index is 0.75–0.92. The specific value is determined by collecting historical data from multiple batches of cold chain sensors under different noise intensities, statistically analyzing the relationship between energy loss and inference error, using the energy retention ratio corresponding to the error inflection point as a candidate threshold value, and confirming it through at least three rounds of cross-validation experiments. The entire pre-processing streaming signal denoising algorithm updates incrementally in a sliding window mode, without waiting for the entire data segment to arrive, maintaining millisecond-level processing latency, and seamlessly integrating with the subsequent asynchronous parallel lock-free buffer mechanism. The technical advantages of the aforementioned adaptive denoising algorithm for streaming signals based on wavelet multi-scale decomposition are as follows: Traditional batch processing denoising methods cannot be executed when the signal has not arrived completely, while sliding window streaming wavelet decomposition allows denoising to proceed synchronously with data arrival, eliminating the accumulation of delays introduced by waiting for complete data segments; adaptive threshold soft truncation automatically compensates for the differences in signal and noise characteristics between different sensors and different transportation segments, avoiding the double degradation problem of over-truncating effective features on low-noise signals or under-truncating features on high-noise signals with a fixed threshold; the mechanism of feeding the signal quality index back to the gating unit enables the network to have an endogenous perception capability for signal anomalies, fundamentally suppressing the interference of poor input on the accuracy of quality prediction.

[0041] The lock-free circular buffer is a data queue with a fixed-capacity circular address space as its storage structure. The producer thread and the consumer thread each hold independent write pointers and read pointers. The absolute value of the difference between the two pointers does not exceed the buffer capacity. Data writing and reading do not require mutex locks. The thread safety of pointer updates is ensured through atomic operations, thereby eliminating the thread synchronization waiting delay caused by traditional lock mechanisms.

[0042] The locally asymmetric gated loop unit is a structural modification of the standard gated loop unit. Its forward path incorporates a fast decay update gate, with the update coefficient calculated by normalizing the absolute value of the difference between the current input and the historical hidden state, giving the unit high sensitivity to input mutations. The feedback path incorporates a slow accumulation forgetting gate, with the forgetting coefficient explicitly coupled to the Arrhenius temperature integral term. When the accumulated temperature integral exceeds the activation energy threshold, the forgetting gate coefficient is forcibly suppressed to near zero, ensuring the unit retains long-range memory of deterioration characteristics after a quality degradation event, preventing rapid forgetting due to normal data in subsequent time steps. The activation energy threshold is determined by the biochemical reaction activation energy parameters corresponding to the type of goods. It is recommended to refer to the activation energy range of common goods in the standard cold chain quality kinetics database (usually [missing information - likely a database name]). The activation energy value of specific goods was confirmed through multiple experimental iterations by fitting the Arrhenius equation through laboratory accelerated deterioration experiments (gradient heating, humidity control, and vibration loading) with the goal of minimizing the residual between the measured sensory quality score and the fitted rate curve.

[0043] The hybrid range attention fusion layer simultaneously operates a local sliding window attention submodule and a global anchor attention submodule. The local submodule calculates dense attention weights within the temporal neighborhood using a sliding window with a step size of 6–10. The suggested step size is determined through experiments analyzing the quality deterioration timescales of different product categories. The global submodule selects 6–10 learnable global anchor tokens to participate in the sparse association calculation of all features. The suggested range for the number of anchors is also determined through ablation experiments on the validation set aiming for Pareto optimality between inference latency and fusion accuracy. The association matrices output by both are fused through weighted summation. The weight coefficients are determined in real-time by the Markov state transition probability of the current transportation segment of the goods. This Markov state transition probability matrix is ​​statistically estimated from historical cold chain transportation segment switching records and continuously updated with new transportation records after system deployment.

[0044] The dynamic hierarchical early exit mechanism involves deploying three confidence evaluation classifiers after the fusion layer. Each classifier outputs a confidence distribution vector for the quality state. When the highest confidence exceeds a confidence threshold, the sample exits inference directly at the current classifier and outputs a quality score; otherwise, it is passed to the next classifier layer. The suggested range for the confidence threshold is 0.82–0.95, and its specific value is determined through grid search experiments on the test set aiming for a joint optimization of the false positive rate and average inference latency.

[0045] The biochemical mechanism-constrained asynchronous parallel multi-channel temporal fusion model is based on deep learning. Its specific structure is as follows: The model input layer receives four heterogeneous data streams: a temperature temporal tensor (dimension is the number of time steps multiplied by the number of temperature sensor nodes), a humidity temporal tensor, a vibration spectrum tensor, and a gas concentration scalar sequence. After wavelet denoising in the pre-streaming signal denoising pipeline, the four input streams enter four completely independent asynchronous computation pipelines. Each pipeline has no synchronization waiting barriers and uses a lock-free circular buffer to transfer feature tensors. The buffer capacity determines the maximum allowable asynchronous time step offset between pipelines. The buffer capacity allocation is dynamically adjusted based on the ratio of the sampling frequencies of each sensor. Pipelines with higher sampling frequencies are allocated larger buffer slots to absorb the rate difference between the arrival of high-frequency data bursts and the processing rate of low-frequency pipelines, preventing frame loss due to buffer overflow. The recommended range for the number of buffer slots is 32–256, estimated by the product of the ratio of the sampling frequencies of each sensor and the target maximum asynchronous offset duration. Each pipeline contains a locally asymmetric gated recurrent unit. The recommended hidden state dimension range is 64–256. The allocation of hidden state dimensions for each pipeline is proportional to the estimated information entropy of the corresponding sensor signal. Pipelines for sensors with higher information entropy are allocated more neurons to ensure the feature representation capacity of high-information signals. The feature tensors output from the four pipelines enter the mixed-range attention fusion layer. The recommended number of attention heads in the layer ranges from 4 to 12. The memory allocation of the attention heads is dynamically scaled during inference based on the product of the current batch size and the sequence length. When the product exceeds the experimentally calibrated upper limit of memory usage, the number of anchors in the global anchor attention submodule is automatically pruned downward from the upper limit to prevent memory overflow. The training loss function of the model consists of a linear superposition of three parts: the mean squared error term of the predicted quality score, the mechanism residual penalty term of the Arrhenius rate equation, and the stability regularization term of the feedback weights of the locally asymmetric gated recurrent units. The weight coefficients of these three terms are recommended to be determined through validation set parameter tuning experiments within the ranges of [1.0, 1.0], [0.1, 0.5], and [0.01, 0.1], respectively. A dynamic hierarchical early exit mechanism is implemented in the output layer. The number of parameters in the three confidence evaluation classifiers increases hierarchically. The first classifier uses a single-layer linear projection structure to ensure the inference speed of the fast exit path. The second and third classifiers successively add one or two fully connected hidden layers, with a recommended parameter increase ratio of 2 to 4 times.The steps for establishing the training dataset for the biochemical mechanism-constrained asynchronous parallel multi-channel time-series fusion model specifically include: collecting historical cold chain transportation data, including raw sensor records of temperature, humidity, vibration, and gas concentration, and corresponding final-state quality scores of goods (sensory scores, microbial detection results, or freshness indicators measured by instruments); statistically analyzing the ratio of deteriorated samples to normal samples in the historical data and recording the ratio of positive to negative samples; for a minority of deteriorated samples, using a conditional generative adversarial network as the generator, with the goods type label and deterioration degree classification label as conditional inputs, generating a model that corresponds to the previous model. Augmented samples with consistent statistical characteristics of real deterioration time series are used until the ratio of positive to negative samples is not less than 1:5. The conditional generative adversarial network is trained with real deteriorated samples as positive samples and generated samples as negative samples, using gradient penalty regularization until the discriminator can no longer distinguish between real and generated samples. A pre-processing streaming signal denoising pipeline is applied to all samples to output the denoised time series and signal quality index, and samples with signal quality index below the experimental calibration lower limit are removed. The dataset is divided into hierarchical groups according to the type of goods, and the recommended ratio of training set, validation set and test set is 7:1.5:1.5. The specific steps for training the biochemical mechanism-constrained asynchronous parallel multi-channel temporal fusion model include: initializing the pre-denoising pipeline with pre-trained wavelet scaling factor parameters, and randomly initializing the main network parameters with a normal distribution; replacing the standard cross-entropy loss with a dynamic focus loss function, and assigning an adaptive amplification factor to the gradient contribution of minority class degraded samples. The amplification factor is dynamically calculated based on the ratio of positive to negative samples in the current mini-batch, with a larger amplification factor for a higher ratio, and a suggested range of 1.0 to 5.0; supervising the difference between the corruption rate predicted by the network and the theoretically derived value of the equation with the mechanism residual penalty term of the Arrhenius rate equation, setting the penalty term weight to a smaller value in the early stage of training to allow the network to freely explore the feature space, and linearly increasing it in the later stage of training to strengthen the mechanism constraint; adopting a layer-by-layer learning rate decay strategy, with the learning rate of the pipeline layer closer to the input being lower than that of the fusion layer, and the learning rate of the fusion layer being lower than that of the output layer, with a suggested ratio of 1:2:4; and recommending a total of 80 to 200 training rounds, specifically terminated when the mean square error of the validation set quality score does not decrease significantly for 10 consecutive rounds.

[0046] The dynamic focus loss function is an extension of the standard focus loss. By dynamically calculating the minority class gradient amplification factor in real-time, the ratio of positive to negative samples is statistically analyzed in each training mini-batch. This amplification factor adapts to the degree of class imbalance rather than using a fixed hyperparameter, thus maintaining continuous attention to minority class samples even when the class ratio fluctuates between training batches, thereby improving the recall rate for early minor quality anomalies. The conditional generative adversarial network (GAN) uses the product type label and the degradation level label as conditional inputs to generate augmented samples consistent with the statistical characteristics of the actual degradation time series. Its technical effect is that, under the condition that degradation samples are extremely scarce in historical data, it compensates for class imbalance at the data distribution level, preventing the decision boundary of the classification model from being severely shifted due to the overwhelming advantage of the number of normal samples. This improves the model's ability to identify scarce abnormal events without relying on the manual collection of a large number of real degradation samples.

[0047] Wherein, the dynamic forgetting gate adjustment function is denoted as The input variables include the cumulative temperature integral at the current time step. (Unit: kJ / mol), Signal Quality Index (Dimensionless, range of values ​​is) and the average output of the forget gate in the current pipeline. (Dimensionless, range of values ​​is) The function output is the adjustment intensity index. (Dimensionless), the calculation formula is: ;in The activation energy threshold (unit: kJ / mol) for the cargo, coefficient , , All three are dimensionless, and their sum is 1. Specific values ​​are determined through a grid search experiment on the validation set with the objective of minimizing the mean square error of the quality score. Suggested ranges are as follows: , , .when When the forget gate decay parameter remains unchanged, the network maintains normal inference; when At that time, the forget gate decay parameter is in step size Adjustments should be made towards reduction to enhance the preservation of historical deterioration characteristics; when At that time, the forget gate decay parameter is in step size Forced compression in the direction of reduction, while simultaneously triggering the global anchor attention submodule of the mixed-range attention fusion layer to expand the number of anchors upward by 1-2, in order to enhance the perception of global abnormal patterns; when At this point, the forget gate attenuation parameter is forcibly reduced to near zero, triggering a dynamic hierarchical early exit mechanism that forces the high-risk warning level to be output directly from the first classifier, without being passed to subsequent classifiers. The step size and interval boundaries are obtained as follows: On a test set containing multiple batches of cold chain transportation data for different types of goods, the step size range is progressively scanned, with the adjustment magnitude of the forget gate attenuation parameter as the independent variable and the mean square error of quality score prediction over the subsequent 10 time steps as the dependent variable. The step size corresponding to the inflection point of error decrease is taken as the lower bound of the suggested range, and the step size corresponding to the second rise point of error is taken as the upper bound of the suggested range. This is confirmed by more than three rounds of independent repeated experiments.

[0048] In the Markov state transition probability matrix, the transportation segment state definition includes port loading and unloading, highway trunk line, cold storage idling, and last delivery segment. The state switching probability is estimated by statistically analyzing the switching frequency between each state in the historical transportation trajectory. The matrix element update is based on the new record added within the sliding time window, ensuring that the probability estimation continuously adapts to changes in the operation mode.

[0049] The tiered early warning system is classified based on quality score ranges: a green "normal" status indicates a quality score above the experimentally calibrated normal threshold; a yellow "concern" status indicates a quality score between the normal and deterioration warning thresholds; and a red "warning" status indicates a quality score below the deterioration warning threshold. Both the normal and deterioration warning thresholds are determined through a combined experiment of sensory evaluation and instrument-measured freshness indices for the corresponding product type. A suggested range for the normal threshold is 0.70–0.88, and for the deterioration warning threshold is 0.40–0.65. Specific values ​​are verified through multiple batches of product experiments after mapping the inflection point of the combined distribution curve of sensory evaluation and instrument measurement results to the quality score coordinate axis.

[0050] The technical advantages of the biochemical mechanism-constrained asynchronous parallel multi-channel temporal fusion model are as follows: Four completely independent asynchronous channels are connected by a lock-free circular buffer, allowing sensor data with different sampling frequencies to be processed in parallel without unified clock synchronization, fundamentally eliminating the latency accumulated in the traditional synchronous batch processing framework due to waiting for the slowest sensor; The forward fast decay update gate of the locally asymmetric gated recurrent unit enables the unit to respond instantly to input mutations, and the feedback slow accumulation forget gate, through explicit coupling with the Arrhenius equation, makes the network's memory decay for quality degradation events constrained by thermodynamic mechanisms rather than purely data-driven, ensuring the legitimacy of the prediction results at the physical mechanism level; The parallel operation of locally dense attention and globally sparse attention in the hybrid range attention fusion layer enables the model to simultaneously capture the impact of short-term temperature fluctuations and long-distance transportation segment switching on quality; The dynamic hierarchical early exit mechanism concentrates computing resources on samples with ambiguous states, and samples with clear states complete the judgment quickly with extremely low inference computation, thus avoiding unnecessary computing power waste while ensuring millisecond-level response requirements.

[0051] Optionally, the present invention also provides a computer-implemented system form, forming a dynamic assessment and early warning system for the quality of frozen transport of perishable goods. The system includes an edge computing node equipped with a temperature sensor, a humidity sensor, a vibration sensor, and a gas concentration sensor, and a computer communicatively connected to the edge computing node. The computer is provided with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they execute the above-mentioned method for dynamic assessment and early warning of the quality of frozen transport of perishable goods, and push graded early warning instructions to the cold chain monitoring terminal in real time.

[0052] The specific implementation of step S01 is as follows: For the raw time-series signals from four sensors (temperature, humidity, vibration, and gas concentration), a three-level discrete wavelet decomposition is performed step-by-step using a sliding window mode. The Darbu wavelet is selected as the wavelet basis function to ensure time-frequency localization characteristics. The first level of decomposition splits the raw signal within the current window into an approximate coefficient sequence and a detail coefficient sequence. An adaptive threshold soft truncation is applied to the detail coefficient sequence. The threshold is dynamically generated by multiplying the standard deviation of the approximate coefficients by a learnable scaling factor. The learnable scaling factor is jointly optimized with the main network parameters during model training, thereby automatically adjusting the threshold according to the noise intensity of different sensors and different transportation segments. The second and third levels repeat the above decomposition process, forming a coefficient pyramid with three levels of frequency resolution. After denoising, inverse wavelet reconstruction is performed on the approximate coefficient sequences of each level to restore the denoised time-series signal. The signal quality index is calculated from the energy retention ratio of the reconstructed signal, with a reference threshold range of 0.75 to 0.92. When the signal quality index is lower than this threshold, a dynamic backtracking mechanism is triggered, forcibly decaying the feature weights of the current time step to the historical average level to prevent low-quality signals from interfering with subsequent inference. The entire noise reduction process is updated step by step in a sliding window mode, without waiting for the entire data segment to arrive, maintaining millisecond-level processing latency, and seamlessly connecting with the subsequent asynchronous parallel lock-free buffer mechanism.

[0053] The specific implementation of step S02 is as follows: The four heterogeneous timing signals after denoising are sent to four completely independent asynchronous computing pipelines. Each pipeline has a lock-free circular buffer with a fixed-capacity circular address space as the storage structure. The producer thread and consumer thread hold independent write pointers and read pointers respectively. The thread safety of pointer updates is ensured through atomic operations. The number of buffer slots is in the reference range of 32 to 256 and is dynamically allocated according to the ratio of the sampling frequencies of each sensor. Pipelines with higher sampling frequencies are allocated more slots to absorb the rate difference between burst data and the processing rate of low-frequency pipelines. Each pipeline has a locally asymmetric gated loop unit. The fast decay update gate coefficient of its forward path is calculated by normalizing the absolute value of the difference between the current input and the historical hidden state, so that the unit has high response sensitivity to input mutations. The slow accumulation forget gate coefficient of the feedback path is explicitly coupled with the Arrhenius temperature integral term. When the accumulated temperature integral exceeds the cargo activation energy threshold, the forget gate coefficient is forced to be reduced to near zero, forcibly retaining the long-term historical characteristics of quality deterioration. The activation energy threshold is determined by referencing the standard cold chain quality kinetics database and fitting the Arrhenius equation through accelerated deterioration experiments; it typically ranges from 40 to 100. The reference range for the hidden state dimension of each pipeline is 64 to 256, which is proportionally distributed to the estimated information entropy value of the corresponding sensor signal.

[0054] The specific implementation of step S03 is as follows: The feature tensors output by the four asynchronous pipelines enter the mixed-range attention fusion layer. Within the layer, the local sliding window attention submodule calculates dense attention weights within the temporal neighborhood using a sliding window with a step size of 6-10, capturing the impact of short-term temperature fluctuations on quality. The global anchor attention submodule selects 6-10 learnable global anchor tokens to participate in the sparse association calculation of all features, capturing the impact of long-distance transportation segment switching on quality. The association matrices output by both are weighted and fused using weight coefficients determined in real-time by the Markov state transition probability. The Markov state transition probability matrix is ​​statistically estimated from historical cold chain transportation segment switching records, including states such as port loading / unloading, highway trunk lines, cold storage idling, and final delivery stages. Matrix elements are continuously updated incrementally with new records added within the sliding time window. The reference range for the number of attention heads is 4-12. When the product of the batch size and sequence length exceeds the upper limit of GPU memory usage, the number of anchors in the global anchor attention submodule is automatically pruned to prevent GPU memory overflow.

[0055] The specific implementation of step S04 is as follows: The features output by the fusion layer are processed by a dynamic hierarchical early exit mechanism. Three confidence evaluation classifiers are arranged sequentially, and each classifier outputs a confidence distribution vector of the quality state. The first classifier adopts a single-layer linear projection structure to ensure the inference speed of the fast exit path; the second and third classifiers successively add one or two fully connected hidden layers, with the parameter increase ratio ranging from 2 to 4 times. When the highest confidence output by a classifier exceeds the confidence threshold (reference range 0.82 to 0.95), the sample directly exits the inference at the current classifier and outputs a quality score; otherwise, it is passed to the next classifier, so that computational resources are concentrated on samples with ambiguous states, and samples with clear states are judged quickly with extremely low computational cost.

[0056] The specific implementation of step S05 is as follows: Based on the quality score and predicted rate of quality deterioration output in step S04, combined with the deterioration threshold corresponding to the type of goods, a graded early warning instruction is triggered. A green "normal" state is indicated when the quality score is higher than the normal threshold (reference range 0.70–0.88); a yellow "concern" state is indicated when the quality score is between the normal threshold and the deterioration warning threshold (reference range 0.40–0.65); and a red "warning" state is indicated when the quality score is lower than the deterioration warning threshold. Both the normal threshold and the deterioration warning threshold are determined through a joint experiment using sensory scores and instrument-measured freshness indicators corresponding to the type of goods. This is verified and confirmed through multiple batches after mapping the inflection point of the joint distribution curve to the quality score coordinate axis. The early warning instruction is pushed to the cold chain monitoring terminal in real time.

[0057] The specific implementation of step S06 is as follows: using a dynamic forgetting gate adjustment function. Continuously monitor the gating weight distribution of each asynchronous pipeline, with input variables including the temperature integral accumulation at the current time step. (unit Signal quality index and the average output of the forget gate in the current pipeline The calculation formula is ,in The activation energy threshold, coefficient , , The sum is 1, and the suggested ranges are respectively , , .when When the forget gate decay parameter remains at its current value, it remains at its current value; when... When the forgetting gate decay parameter is adjusted in a decreasing direction with a step size of 0.05 to 0.10, When the forget gate decay parameter is forcibly compressed with a step size of 0.10 to 0.20, the number of global anchor points is simultaneously expanded upwards by 1 to 2; when At that time, the forget gate decay parameter is forcibly reduced to near zero, and the dynamic hierarchical early exit mechanism is triggered to force the high-risk warning level to be output directly from the first classifier.

[0058] It should be noted that the key technologies of this invention include: connecting four independent asynchronous computing pipelines with a lock-free circular buffer, enabling sensor data with different sampling frequencies to proceed in parallel without unified clock synchronization, eliminating the latency accumulated in the synchronous batch processing framework due to waiting for the slowest sensor; explicitly coupling the Arrhenius temperature integral term and the forget gate coefficient to constrain the decay behavior of the forget gate, so that the retention of deterioration history features is driven by thermodynamic mechanisms rather than pure data distribution, ensuring from the perspective of physical legitimacy that long-range deterioration features do not dissipate due to subsequent normal data; running local dense attention and global sparse attention in parallel with a hybrid range attention fusion layer, and using Markov state transition probabilities for real-time weighted fusion, enabling the model to simultaneously perceive short-term fluctuations and long-range segment switching. These three technologies together form a complete technical synergy to ensure real-time fusion inference of long-range features of quality deterioration under asynchronous arrival conditions.

[0059] It's important to note that during cold chain transportation, when multiple types of goods are en route simultaneously and samples of quality deterioration from each batch are extremely scarce in historical data, the model's recall rate for early minor quality anomalies will be severely low due to the overwhelming dominance of normal samples. This causes the decision boundary to shift significantly towards the normal class, preventing the model from triggering timely warnings in the early stages of quality deterioration and creating a systemic risk of missed detections. The reason for this technical problem is that quality deterioration events are low-probability events in actual cold chain transportation. In historical data, the ratio of normal samples to deteriorated samples may be tens to one or even higher. Under these conditions, the standard cross-entropy loss function's gradient contribution to the majority class of normal samples dominates, causing the model parameter updates to consistently favor the normal class feature space. The gradient contribution of the minority class of deteriorated samples is overwhelmed due to the extremely small base, ultimately resulting in the model's ineffective training in representing deterioration features. Common solutions to this problem include oversampling or undersampling the training dataset to balance the class ratio, or using a weighted loss function with fixed class weights to assign higher weights to the minority class. However, oversampling directly replicates minority class samples, failing to increase sample diversity, and the model is still prone to overfitting to the finite distribution of existing deteriorated samples. Undersampling discards a large number of normal samples, reducing the model's ability to model normal states. Fixed class weights cannot adaptively adjust when the class ratio fluctuates between training batches, and when the class ratio in a mini-batch happens to deviate from the global ratio, the gradient amplification effect deviates from expectations. This invention effectively solves this technical problem by using a conditional generative adversarial network with product type labels and deterioration severity classification labels as conditional inputs to generate augmented samples consistent with the statistical characteristics of the actual deterioration time series. This compensates for class imbalance at the data distribution level, ensuring that the generated samples cover the diverse distribution of real deteriorated samples rather than simply replicating them. Simultaneously, a dynamic focus loss function is used to dynamically calculate the minority class gradient amplification factor by statistically analyzing the ratio of positive and negative samples in each training mini-batch. This amplification factor adaptively changes with the class ratio of the current mini-batch, maintaining continuous attention to minority class samples even when the class ratio fluctuates between batches. This dual approach of data layer and loss layer synergistically improves the model's recall rate for early minor quality anomalies.

[0060] Specifically, the principle of this invention is:

[0061] First, traditional synchronous batch processing frameworks require all sensors to complete data acquisition and transmit it to subsequent modules at the same time step. In multi-sensor systems with significantly different sampling frequencies, this inevitably leads to buffer accumulation of high-frequency sensor data or forced interpolation of low-frequency sensor data, both of which introduce additional latency or information distortion. This invention replaces the synchronous waiting barrier with a lock-free circular buffer. The producer thread and consumer thread each hold independent read / write pointers, ensuring thread safety through atomic operations. This allows for complete decoupling and parallel operation of the sensor pipelines. Burst data from high-frequency sensors is absorbed by a larger-capacity buffer slot, and data from low-frequency sensors does not block the progress of high-frequency pipelines, thus eliminating synchronous waiting latency in a physical mechanism.

[0062] Secondly, quality degradation events have an irreversible cumulative effect over time. The forget gate of a standard gated loop unit is purely data-driven. If the temperature returns to normal in subsequent time steps after a degradation event, the forget gate coefficient will naturally rise with changes in the input distribution, leading to a rapid dilution of the degradation history characteristics. This invention explicitly couples the forget gate coefficient with the Arrhenius temperature integral term. The Arrhenius equation describes the exponential growth relationship between the reaction rate and the cumulative temperature integral. When the cumulative integral exceeds the activation energy threshold of the goods, it indicates that the quality degradation reaction has irreversibly exceeded the critical point. At this point, the forget gate coefficient is forcibly suppressed to near zero, forcibly preserving the degradation history characteristics. This makes the model's memory decay behavior constrained by thermodynamic mechanisms rather than solely determined by data distribution, thus ensuring the reliable preservation of long-range degradation characteristics at the level of physical legitimacy.

[0063] Furthermore, the cold chain transportation process simultaneously involves two types of quality influencing factors with significant differences in time scale: short-term local temperature fluctuations and long-distance transportation segment switching. Models with a single attention range struggle to capture both types of patterns. In the hybrid range attention fusion layer of this invention, the local sliding window attention submodule focuses on dense correlation calculation within the temporal neighborhood, while the global anchor attention submodule achieves sparse correlation calculation across the entire sequence through a small number of learnable anchors. The correlation matrices output by both are fused in real time with Markov state transition probabilities, enabling the model to adaptively adjust the contribution ratio of the two types of attention under different transportation segment states. This ensures the ability to perceive both short-term fluctuations and long-distance switching within the same inference path.

[0064] Finally, the dynamic hierarchical early exit mechanism concentrates computational resources on samples with ambiguous states, while samples with clear states exit directly at the first classifier. The dynamic forget gate adjustment function continuously monitors the distribution of gate weights and adaptively adjusts the decay parameters, enabling the entire system to maintain inference accuracy while ensuring millisecond-level response. Overall, it forms a complete technical closed loop for real-time fusion inference of long-range features with deteriorating quality under asynchronous arrival conditions.

[0065] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0066] The specific implementation method of step S01 is as follows.

[0067] The original time-series signals come from four types of sensors: temperature, humidity, vibration, and gas concentration. The original discrete signals are denoted as follows: Using Darbrook wavelet basis functions to Perform three-level discrete wavelet decomposition. layer( Decomposition of the current approximate coefficient sequence (in (Splitting into an approximate coefficient sequence) With detail coefficient sequence The decomposition formula is as follows:

[0068] ;

[0069] ;

[0070] In the formula, These are the coefficients of the Dab wavelet low-pass filter. For the corresponding high-pass filter coefficients, This is the discrete-time index of the current layer's output. For the convolution sum index. For the ... Layer detail factor Apply adaptive threshold soft truncation, threshold By the Standard deviation of layer approximation coefficients With learnable scaling factor Jointly determined:

[0071] ;

[0072] In the formula, The learnable scaling factor (dimensionless) has an empirical value of 1.0 and is jointly optimized with the main network parameters during model training. By the Layer approximation coefficient sequence Statistical calculations show that its dimensions are the same as the original signal. same, Dimensions and Since they are the same, the dimensions on both sides of the equation are identical. (Soft truncation function) The definition is as follows:

[0073] ;

[0074] In the formula, For symbolic functions, For absolute value operations, To ensure that the truncated coefficients do not produce negative shifts, after the three-level decomposition and soft truncation are completed, inverse wavelet reconstruction is performed on the approximate coefficient sequences of each level to restore the denoised time-series signal. The inverse reconstruction is the standard inverse process of Dabu wavelet decomposition. Signal quality index The definition is as follows:

[0075] ;

[0076] In the formula, the molecule is the reconstruction signal. The total energy, with the original signal as the denominator. The total energy, both have the same dimensions. It is a dimensionless ratio. .when When the value falls below the experimentally calibrated threshold, a dynamic backtracking mechanism is triggered, and the feature weights at the current time step decay to the historical average level to prevent low-quality signals from contaminating subsequent inference. The entire decomposition and reconstruction process proceeds incrementally in a sliding window mode, maintaining millisecond-level processing latency.

[0077] The specific implementation method of step S02 is as follows.

[0078] The four denoised signals enter independent asynchronous computation pipelines. Feature tensors are passed between pipelines using lock-free circular buffers. The producer thread writes pointers and the consumer thread reads pointers through atomic operations, eliminating the need for mutex locks and thus eliminating thread synchronization latency. Each pipeline contains a locally asymmetric gated loop unit, which updates the gate coefficients by rapidly decaying forward. The calculation formula is as follows:

[0079] ;

[0080] In the formula, To update the gate weight matrix (dimensionless). To update the gate bias vector (dimensionless). For the current time step The input vector, This is the hidden state vector from the previous time step. For logistic activation function, This is the numerically stable term (dimensionless), and the default value is... , Let L be the vector L2 norm; the fractional term on the right is the normalization mapping factor (dimensionless), which makes L be the vector L2 norm. It exhibits high sensitivity to input mutations, and both sides are dimensionless quantities. The feedback slowly accumulates the forgetting gate coefficient. With Arrhenius temperature integral term Explicit coupling, calculated as follows:

[0081] ;

[0082] In the formula, This is the forget gate weight matrix (dimensionless). This is the forget gate bias vector (dimensionless). This is the coupling strength coefficient (dimensionless), with an empirical value of 0.9. This is the cumulative temperature integral at the current time step. The activation energy threshold for the cargo is the same for both (both units are...). Therefore The ratio is dimensionless, and the entire expression on the right-hand side is a dimensionless quantity. dimensionless Truncate the ratio to less than 1 to ensure ;when hour, It is forced to be suppressed to near zero, and the long-term historical characteristics of deterioration are forcibly preserved. Accelerated deterioration experiments were conducted on goods under gradient temperature increase, humidity control, and vibration loading. The goal was to minimize the residual between the measured sensory quality score and the fitted rate curve. The Arrhenius equation was fitted to determine the activation energy range (40–100 kcal / kg) of common goods in the standard cold chain quality kinetics database. The specific value was confirmed through multiple experimental iterations.

[0083] The specific implementation method of step S03 is as follows.

[0084] The feature tensors output from the four pipelines enter the mixing range attention fusion layer, and the local sliding window attention submodule within the layer operates with a step size. (Suggested range 6-10, determined through experiments analyzing the time scale of quality deterioration for different product categories) Calculate the dense attention weight matrix within the time neighborhood. Global Anchor Attention Submodule Selection 6–10 learnable global anchor tokens (suggested range 6–10, determined through ablation experiments on the validation set aiming for Pareto optimality of inference latency and fusion accuracy) ( (Assuming the feature dimension) participates in sparse association calculation, outputting an association matrix. The weighted fusion formula for the two is as follows:

[0085] ;

[0086] In the formula, To merge the correlation matrix, (Dimensionless) represents the current time step. By Markov state transition probability matrix The fusion weighting coefficients are determined in real time. and Since they have the same dimensions, therefore Its dimensions are consistent with both. Markov state transition probability matrix. elements Indicates the status of the transportation segment Transition to state The probability, where the transportation route status includes port loading and unloading, main highways, cold storage, and the last stage of delivery, etc. It is estimated based on the frequency of state transitions in historical transportation trajectories and continuously updated with new records added within a sliding time window. Take the weighted value of the transition probability corresponding to the current state row.

[0087] The specific implementation method of step S04 is as follows.

[0088] The output features of the fusion layer are sequentially passed through three confidence evaluation classifiers, the first of which... One classifier ( Output the confidence distribution vector of the quality status. ,in This represents the number of quality status categories. When... Exceeding the confidence threshold (The suggested range is 0.82–0.95, determined through grid search experiments on the test set with the joint optimization of false positive rate and average inference delay as the objective.) When the sample is at the [missing value] ... The inference process terminates directly at each classifier and outputs the quality score. Otherwise, pass it to the first... The process continues with multiple classifiers, up to the third classifier which forces a specific output. The first classifier uses a single-layer linear projection structure, while the second and third classifiers each add one or two fully connected hidden layers, with the number of parameters typically increasing by a factor of 2 to 4.

[0089] The specific implementation method of step S05 is as follows.

[0090] The tiered early warning system is based on quality scores. (Dimensionless) and predicted deterioration rate Jointly determined: when When it is in a green normal state; when At this time, it is in a yellow alert status; when The current state is a red alert. In the formula, This is the normal threshold (dimensionless), with a recommended range of 0.70–0.88. The deterioration warning threshold (dimensionless) is recommended to be in the range of 0.40–0.65. Both thresholds are mapped from the inflection point of the joint distribution curve of sensory scores and instrument-measured freshness indicators corresponding to the product type to the quality score coordinate axis, and verified through experiments on multiple batches of products. Warning commands are pushed to the cold chain monitoring terminal in real time.

[0091] The specific implementation method of step S06 is as follows.

[0092] Dynamic forgetting gate adjustment function The input variables include the temperature integral accumulation at the current time step. (unit Signal quality index (Dimensionless) and the average output of the current pipeline forget gate (dimensionless, The function output is the adjustment intensity index. (Dimensionless), the calculation formula is as follows:

[0093] ;

[0094] In the formula, , , All are dimensionless coefficients, and the sum of the three is 1. Suggested ranges are as follows: , , The goal is to determine the minimum mean square error of the quality score through a grid search experiment on the validation set. It is a dimensionless ratio. and All of them are dimensionless quantities, therefore Since it is a dimensionless quantity, the dimensions are consistent on both sides of the equation. When When the forget gate decay parameter remains unchanged, the network maintains normal inference; when At this time, the forgetting gate decay parameter is adjusted to decrease in step size of 0.05–0.10 to enhance the retention of historical deterioration characteristics; when At this time, the forget gate decay parameter is forcibly compressed with a step size of 0.10 to 0.20, while simultaneously triggering the global anchor attention submodule to expand the number of anchors upward by 1 to 2, in order to enhance the perception of global abnormal patterns; when At this point, the forget gate attenuation parameter is forcibly reduced to near zero, triggering a dynamic hierarchical early exit mechanism that forces the high-risk warning level to be output directly from the first classifier. The step size values ​​and interval boundaries were determined by progressively scanning the step size range of 0.01–0.25 on a test set containing multiple batches of cold chain transportation data for different types of goods. The step size was set as the independent variable (the adjustment magnitude of the forget gate attenuation parameter) and the mean square error of the quality score prediction over the subsequent 10 time steps was set as the dependent variable. The step size value corresponding to the error decrease inflection point was taken as the lower bound of the suggested range, and the step size value corresponding to the second error increase point was taken as the upper bound of the suggested range. This was confirmed through at least three rounds of independent repeated experiments.

[0095] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0096] This embodiment uses the actual transportation process of a batch of chilled fresh fish goods from a main highway to the final delivery point as the test scenario. One each of a temperature sensor, humidity sensor, vibration sensor, and gas concentration sensor were selected, with a data collection duration of 72 hours. The sensor sampling frequencies were 1... 0.5 10 and 0.2 The activation energy parameter of the cargo was confirmed to be 68 after fitting the Arrhenius equation through accelerated degradation experiments. The normal threshold is set to 0.80, the degradation warning threshold is set to 0.55, the signal quality index threshold is set to 0.82, and the confidence threshold is set to 0.88. Take 0.45, Take 0.30, Take 0.25.

[0097] In step S01, the four original sensor signals are progressively processed using a sliding window of 256 sampling points, performing three-level Darbu wavelet decomposition. The adaptive threshold for each level is dynamically generated by the product of the standard deviation of the corresponding level's approximation coefficients and a learnable scaling factor. After three-level inverse wavelet reconstruction, a denoised time-series signal and signal quality index are output. In this embodiment, the original signal noise intensity of the vibration sensor on the highway trunk section is relatively high, and the corresponding signal quality index drops to 0.79 within a certain time window, below the set threshold of 0.82, triggering a dynamic backtracking mechanism that decays the feature weights of that time step to the historical average level. The statistical results of the signal quality indices of the four sensors are shown in Table 1.

[0098] Table 1. Statistical Table of Signal Quality Index of Four Sensors

[0099]

[0100] In step S02, the four denoised signals are input into four independent asynchronous computation pipelines. Buffer slots are allocated based on the sampling frequency ratio: 128 slots for the vibration sensor pipeline, 64 slots for the temperature sensor pipeline, 32 slots for the humidity sensor pipeline, and 32 slots for the gas concentration sensor pipeline. There are no synchronization barriers between pipelines, and atomic operations ensure thread safety for read / write pointers. The hidden state dimension of the local asymmetric gated loop unit within each pipeline is allocated based on the information entropy estimate: 192 for the vibration sensor pipeline, 128 for the temperature sensor pipeline, 96 for the humidity sensor pipeline, and 64 for the gas concentration sensor pipeline. During the transportation segment from hour 36 to hour 48, the accumulated temperature integral reaches 71. More than 68 The activation energy threshold is set, and the forgetting gate coefficient of the corresponding pipeline is forcibly suppressed to near zero, preserving the historical characteristics of quality deterioration during that period. The trend of the forgetting gate coefficient of each pipeline over time is as follows: Figure 2 As shown in the figure, the hard suppression process of the forget gate coefficient after the temperature integral accumulation exceeds the activation energy threshold is clearly presented.

[0101] In step S03, the feature tensors output from the four pipelines enter the mixing range attention fusion layer, with the number of attention heads set to 8. The local sliding window step size is set to 8, and the number of global anchor tokens is initially set to 8. The Markov state transition probability matrix is ​​estimated statistically from historical transportation records. In this embodiment, the goods undergo state transitions from highway trunk line to cold storage and then to the final delivery stage. The elements of the transition probability matrix are shown in Table 2.

[0102] Table 2 Markov State Transition Probability Matrix

[0103]

[0104] After 60 hours, the goods enter the final stage of delivery. The adjustment of the corresponding state transition probability increases the weight of the global anchor attention submodule, enhancing the model's ability to perceive the impact of long-distance route switching. The weight coefficients of the mixed-range attention fusion layer change with the state switching of the transportation route as follows: Figure 3 As shown.

[0105] In step S04, the dynamic hierarchical early exit mechanism sequentially activates three confidence assessment classifiers. Statistics show that during the normal transportation phase from hour 0 to hour 30, approximately 74% of the time-step samples reached the confidence threshold of 0.88 at the first classifier and exited directly, maintaining a low inference delay. During the quality deterioration phase from hour 36 to hour 48, the proportion of samples exiting directly at the first classifier decreased to 31%, with more samples being passed to the second and third classifiers for finer judgment. The sample splitting ratios for the three classifiers at each stage are shown in Table 3.

[0106] Table 3. Sample splitting ratio for each stage of the classifier

[0107]

[0108] As the quality deteriorates, more samples enter the subsequent classifiers, and computing resources are automatically concentrated on samples with ambiguous states. This diversion trend shows that the dynamic hierarchical early exit mechanism can reduce computing overhead when the state is clear and ensure inference accuracy when the state is ambiguous.

[0109] In step S05, a graded early warning is triggered based on the quality score and the predicted rate of quality deterioration. For example... Figure 4As shown, in this embodiment, the quality score curve changes with transportation time, exhibiting a trend of initial stability, rapid decline in the middle stage, and eventual stabilization. From hour 0 to hour 30, the quality score remains between 0.83 and 0.91, in a green normal state. At hour 36, the quality score drops to 0.62, entering a yellow concern state, triggering a yellow warning command and pushing it to the cold chain monitoring terminal. At hour 42, the quality score further decreases to 0.48, below the deterioration warning threshold of 0.55, triggering a red warning command. The rapid decline in the quality score in the middle stage is due to the fact that after the accumulated temperature integral continuously exceeds the activation energy threshold, the forgetting gate is forcibly suppressed, the historical characteristics of deterioration are fully preserved and fully utilized in the fusion layer, and the model's predictive response to quality deterioration becomes more accurate as the accumulated characteristics strengthen. The rate of decline of the quality score curve is consistent with the actual process of biochemical deterioration.

[0110] In step S06, the dynamic forget gate adjustment function adaptively adjusts the forget gate decay parameter of each channel. During hours 36 to 48, the temperature integral accumulation... 71 Signal quality index The mean is 0.85, and the mean of the forget gate output is... The moderating intensity index is 0.38. It is 0.71, which is at the level of Within the specified interval, the forget gate decay parameter is forcibly compressed in a decreasing direction with a step size of 0.15, while the number of global anchor points is expanded from 8 to 9, enhancing the perception of global anomalous patterns. The change of the adjustment intensity exponent over time is as follows: Figure 5 As shown in the figure, the regulation intensity index rapidly rises to the 0.71 range after 36 hours and remains there until the temperature integral growth rate slows down after the transportation segment switches to the final delivery segment. The regulation intensity index gradually falls back to around 0.45, and the forget gate decay parameter automatically recovers to the normal regulation step range, which reflects the dynamic regulation function's ability to balance inference accuracy and computational resource consumption.

[0111] This invention employs a biochemical mechanism-constrained asynchronous parallel multi-channel time-series fusion model, the specific structure and data flow of which are described below. The model input layer receives four heterogeneous data streams. The temperature time-series tensor dimension is the number of time steps multiplied by the number of temperature sensor nodes. The humidity time-series tensor, vibration spectrum tensor, and gas concentration scalar sequence are independent. After passing through a pre-wavelet denoising pipeline, the four inputs enter four asynchronous computation pipelines. Within each pipeline, a locally asymmetric gated recurrent unit processes the feature tensor step-by-step and passes it to the mixed-range attention fusion layer via a lock-free circular buffer. The feature vector output from the fusion layer is sequentially passed through three confidence evaluation classifiers. Samples meeting the confidence threshold are directly output as quality scores, while other samples are passed layer by layer.

[0112] Alternative neural network structures are described below. The first alternative structure replaces the local asymmetric gated recurrent unit with a temporal convolutional network. The four denoised temporal signals are fed into four independent branches of the temporal convolutional network. Each branch consists of multiple layers of causal dilated convolutions stacked together, with the dilation coefficient increasing layer by layer in powers of 2 to cover a multi-scale temporal receptive field ranging from short-term mutations to long-term trends. The output feature tensor is passed to the mixed-range attention fusion layer through an unlocked circular buffer. The subsequent fusion and output process is completely consistent with the main scheme. The difference from the main scheme is that the retention of long-range memory depends on the receptive field coverage of the dilated convolutions rather than explicit forgetting gate coupling, making it suitable for edge deployment scenarios with extremely high requirements for real-time inference latency. The second alternative structure replaces the local asymmetric gated loop unit with a state-space model. The dynamic evolution of the sensor signal is described by linear state-space equations in the four pipelines. The eigenvalues ​​of the state transition matrix are constrained within the unit circle to ensure long-range memory stability. The Arrhenius temperature integral term is coupled with the biochemical mechanism by modifying the diagonal elements of the state transition matrix. The output hidden state tensor is passed to the mixed-range attention fusion layer through a lock-free circular buffer. Its feature extraction dimension is consistent with the main scheme. The Arrhenius mechanism residual penalty term in the loss function maintains the same convergence objective as the main scheme.

[0113] It should be noted that the variables involved in this invention are explained in detail in Table 4.

[0114] Table 4. Variable Explanation Table

[0115]

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic evaluation and early warning of the quality of perishable goods transported under refrigeration, characterized in that, Includes the following steps: The raw time-series signals from temperature, humidity, vibration and gas concentration sensors are subjected to three-level discrete wavelet decomposition in sliding window mode. Adaptive threshold soft truncation is applied to the detail coefficients of each level to gradually reconstruct the denoised time-series signal and output the signal quality index. The four heterogeneous time-series signals after denoising are sent into four independent asynchronous computing pipelines. Each pipeline uses an unlocked ring buffer to pass feature tensors. The forget gate coefficient of the local asymmetric gated loop unit in the pipeline is explicitly coupled with the Arrhenius temperature integral term. When the accumulated temperature integral exceeds the activation energy threshold, the forget gate coefficient is hard-suppressed, forcibly preserving the long-term historical features of quality deterioration. The feature tensors output by the four asynchronous pipelines are fed into the hybrid range attention fusion layer. The local sliding window attention submodule and the global anchor attention submodule are computed in parallel. The correlation matrix output by the two is weighted and fused by the weight coefficients determined in real time by the Markov state transition probability. The features output by the fusion layer are processed by a dynamic hierarchical early exit mechanism. The three confidence evaluation classifiers judge the quality status in turn. When the confidence exceeds the confidence threshold, the quality score is directly output. Otherwise, it is passed to the next classifier. Based on the quality score and the predicted rate of quality deterioration, combined with the deterioration threshold corresponding to the type of goods, a graded early warning instruction is triggered, and the early warning instruction is pushed to the cold chain monitoring terminal in real time.

2. The method for dynamic evaluation and early warning of the quality of perishable goods transported under frozen conditions according to claim 1, characterized in that, Also includes: The dynamic forget gate adjustment function continuously monitors the gating weight distribution of each asynchronous pipeline, and adaptively adjusts the forget gate decay parameter according to the interval to which the function output value belongs.

3. The method for dynamic evaluation and early warning of the quality of perishable goods transported under frozen conditions according to claim 2, characterized in that, The signal quality index is calculated from the energy retention ratio of the reconstructed signal. When the signal quality index is lower than the signal quality index threshold, a dynamic backtracking mechanism is triggered, which forces the feature weights of the current time step to decay to the historical average level.

4. The method for dynamic evaluation and early warning of the quality of perishable goods transported frozen according to claim 3, characterized in that, The lock-free circular buffer uses a fixed-capacity circular address space as its storage structure. The producer thread and the consumer thread each hold independent write pointers and read pointers. Atomic operations ensure thread safety for pointer updates. The buffer capacity is dynamically adjusted according to the ratio of the sampling frequencies of each sensor.

5. The method for dynamic evaluation and early warning of the quality of perishable goods transported frozen according to claim 4, characterized in that, The forward path of the locally asymmetric gated loop unit is equipped with a fast decay update gate, and the update coefficient is calculated by normalizing the absolute value of the difference between the current input and the historical hidden state. The feedback path is equipped with a slow accumulation forget gate, and the forget coefficient is explicitly coupled with the Arrhenius temperature integral term. When the accumulated temperature integral exceeds the activation energy threshold, the forget gate coefficient is forced to be reduced to near zero.

6. The method for dynamic evaluation and early warning of the quality of perishable goods transported under frozen conditions according to claim 5, characterized in that, The activation energy threshold is determined by the biochemical reaction activation energy parameter corresponding to the type of goods. The specific value is determined by fitting the Arrhenius equation through laboratory accelerated deterioration experiments, with the goal of minimizing the residual between the measured quality sensory score and the fitted rate curve.

7. The method for dynamic evaluation and early warning of the quality of perishable goods transported under frozen conditions according to claim 6, characterized in that, In the hybrid range attention fusion layer, the local sliding window attention submodule calculates the dense attention weights in the temporal neighborhood with the sliding window step size, and the global anchor attention submodule selects learnable global anchor tokens to participate in the sparse association calculation of all features. The association matrix output by both is fused by weighted summation of the weight coefficients determined in real time by the Markov state transition probability matrix.

8. The method for dynamic evaluation and early warning of the quality of perishable goods transported under refrigeration according to claim 7, characterized in that, The Markov state transition probability matrix includes the states of the transportation route, such as port loading and unloading, highway trunk lines, cold storage, and the final delivery stage. The state switching probability is estimated by statistically analyzing the switching frequency between states in the historical transportation trajectory. The matrix elements are continuously updated with new records added within the sliding time window as increments.

9. The method for dynamic evaluation and early warning of the quality of perishable goods transported under refrigeration according to claim 8, characterized in that, In the dynamic hierarchical early exit mechanism, the number of parameters of the three confidence evaluation classifiers increases hierarchically. The first classifier adopts a single-layer linear projection structure, and the second and third classifiers successively add fully connected hidden layers.

10. The method for dynamic evaluation and early warning of the quality of perishable goods transported under frozen conditions according to claim 9, characterized in that, The graded early warning instructions are divided into the following levels: green normal state when the quality score is higher than the normal threshold, yellow attention state when the quality score is between the normal threshold and the deterioration warning threshold, and red warning state when the quality score is lower than the deterioration warning threshold.

Images