A cold chain logistics monitoring method and system

Abstract

This invention relates to the field of logistics monitoring technology, and more particularly to a cold chain logistics monitoring method and system. The invention calculates data integrity index values ​​at each moment within a historical time period, i.e., the cumulative value of the product of the effective data percentage of each sensor and its importance weight; obtains threshold exceedance index values ​​at each moment based on pre-processed sensor data and corresponding set numerical ranges; obtains multi-source consistency index values ​​at each moment based on the standard deviation of redundant sensor data, semantic labels, and data changes between adjacent moments; weights and sums the three indices to obtain the confidence weights at each moment; constructs an enhanced input vector using the confidence weights, semantic labels, and data from each sensor; and predicts the pre-processed data and semantic labels of each sensor within a preset time step in the future, based on the enhanced input vectors at each moment within the historical time period, enabling accurate monitoring of cold chain logistics and improving monitoring reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of logistics monitoring, in particular to a cold-chain logistics monitoring method and system. BACKGROUND

[0002] In the cold-chain logistics system, goods such as medicines, vaccines and fresh foods that are highly sensitive to the environment require the key parameters such as temperature and humidity to be stable and controllable throughout the transportation process. The common practice is to install a vehicle terminal on the vehicle to collect real-time vehicle compartment environment data and report it to the cloud platform, and the supervision end (or mobile end) monitors and checks compliance accordingly. With the industry moving from "seeing data" to "predicting risks and leaving traces for auditing", relying solely on uploading and displaying is no longer sufficient to meet the needs of compliance and risk control: the system needs to achieve data consistency between the vehicle end, the cloud platform and the supervision / mobile end (hereinafter referred to as "the three ends"), process traceability and abnormal chain.

[0003] Existing cold-chain logistics future state prediction mostly takes Markov chain, graph model and statistical single trajectory prediction algorithm as the core. Markov chain relies on the strong assumption that the state is determined only by the previous time, making it difficult to capture the cumulative impact of cross-period operations such as cold machine start-stop and door opening to take goods, and the state discretization process will lose key information such as temperature fluctuations. The fixed topology structure of the graph model is difficult to adapt to dynamic scenarios such as temporary sensor failures and environmental mutations, and the conditional probability calculation is limited by the completeness of prior knowledge. The statistical single trajectory algorithm has a sudden drop in prediction accuracy when facing non-stationary disturbances such as drastic temperature changes during loading and unloading and extreme weather due to the stationarity assumption. With the development of deep learning models, by fusing multi-sensor time series data to construct an input vector, using models such as LSTM can capture long-range dependencies and non-linear features, effectively solving the problems of poor adaptability to dynamic scenarios and weak multi-source information fusion capability of traditional methods, and further improving the timeliness and robustness of prediction.

[0004] However, the way of simply concatenating multi-source data to construct an input vector lacks deep modeling of semantic associations of data, such as the causal relationship between cold machine power changes and temperature fluctuations, and the differences in data features at different periods (such as during transportation and at stops for unloading), which are difficult to fully exploit these prior knowledge through model self-learning, resulting in insufficient information density of the input vector, which may still cause prediction bias under complex working conditions. SUMMARY

[0005] Therefore, the technical problem to be solved by the present application is to overcome the defects of the existing cold-chain logistics monitoring method based on cold-chain logistics future state prediction, which simply concatenates multi-source data to construct an input vector, lacks deep modeling of semantic associations of data, and results in poor prediction accuracy of monitoring parameters at future time.

[0006] To solve the above technical problems, the present application provides a cold chain logistics monitoring method, comprising:

[0007] The data integrity index value of each time point in the historical time period is obtained by multiplying the effective data proportion of each sensor in the cold chain vehicle cabin in the historical time period by the importance weight of the sensor, and then accumulating the product results of all sensors.

[0008] Based on the pre-processed data of each sensor at each time point in the historical time period and the corresponding set value interval of the data, the threshold exceeding degree index value of each time point in the historical time period is obtained.

[0009] Based on the standard deviation of the pre-processed data of the redundant sensor measuring the same physical quantity at each time point in the historical time period, the semantic label of the cold chain vehicle cabin at each time point in the historical time period, and the change amount of the pre-processed data of each sensor at adjacent time points in the historical time period, the multi-source consistency index value of each time point in the historical time period is obtained.

[0010] The data integrity index value, threshold exceeding degree index value and multi-source consistency index value of each time point in the historical time period are weighted and summed to obtain the confidence weight of each time point in the historical time period.

[0011] The confidence weight of each time point in the historical time period, the semantic label of the cold chain vehicle cabin and the pre-processed data of each sensor are used to construct an enhanced input vector of each time point in the historical time period.

[0012] Based on the enhanced input vector of each time point in the historical time period, the data of each sensor and the semantic label of the cold chain vehicle cabin in the future preset time step are obtained.

[0013] Preferably, the threshold exceeding degree index value of each time point in the historical time period is obtained based on the pre-processed data of each sensor at each time point in the historical time period and the corresponding set value interval of the data, comprising:

[0014] Based on the pre-processed data of each sensor at each time point in the historical time period and the corresponding set value interval of the data, the exceeding amplitude of each sensor at each time point in the historical time period is calculated, and the formula is:

[0015] ,

[0016] The exceeding amplitude of each sensor at each time point in the historical time period is normalized to obtain the normalized exceeding amplitude of each sensor at each time point in the historical time period.

[0017] Based on the importance weight of each sensor and the normalized exceedance magnitude of each sensor at each moment within the historical time period, the threshold exceedance index value at each moment within the historical time period is calculated using the following formula:

[0018] ,

[0019] in, Within the historical period The threshold at any given time exceeds the degree indicator value. For the first time in the historical period At the [time]th moment The exceedance of each sensor, For the first Each sensor corresponds to the lower limit of a set numerical range. For the first Each sensor corresponds to the upper limit of a set numerical range. For the first time in the historical period At the [time]th moment Data from one sensor, For the first time in the historical period At the [time]th moment Excess amplitude after normalization of each sensor For the first The importance weight of each sensor This represents a time index within a historical period. Indicates the sensor index. This is a collection of sensors for use inside the cold chain vehicle compartment.

[0020] Preferably, the method for obtaining the multi-source consistency index value at each moment in the historical time period based on the standard deviation of the preprocessed data of redundant sensors that measure the same physical quantity as each sensor within the historical time period, the semantic tags of the cold chain compartment at each moment in the historical time period, and the change in the preprocessed data of each sensor at adjacent moments in the historical time period includes:

[0021] The standard deviation of the preprocessed data of redundant sensors that measure the same physical quantity as each sensor within the historical time period is normalized at each moment to obtain the redundant sensor consistency index value of each sensor within the historical time period.

[0022] The importance weight of each sensor is weighted and summed with the consistency index value of each sensor in the historical time period to obtain the consistency index value of redundant sensors at each moment in the historical time period.

[0023] Based on the semantic labels of each moment in the historical time period and its previous moment, determine whether there is an anomaly in the amount of change of the data of each sensor after preprocessing at each moment in the historical time period. If there is, set the cross-modal consistency index value of that moment in the historical time period to 0; if there is no, set the cross-modal consistency index value of that moment in the historical time period to 1.

[0024] The absolute value of the difference between the preprocessed data of each sensor at adjacent moments within the historical time period is normalized to obtain the temporal smoothing consistency index value of each sensor at each moment within the historical time period.

[0025] The importance weight of each sensor is weighted and summed with the temporal smoothing consistency index value of each sensor in the historical time period to obtain the temporal smoothing consistency index value at each moment in the historical time period.

[0026] The weighted summation of the redundant sensor consistency index value, cross-modal consistency index value, and time smoothing consistency index value at each moment within the historical time period yields the multi-source consistency index value at each moment within the historical time period.

[0027] Preferably, the step of obtaining the data from each sensor and the semantic tags of the cold chain vehicle compartment based on the enhanced input vectors at various moments within a historical time period includes:

[0028] Will arrive The confidence weights at each time point within the time period are scaled using a normalization function to obtain... arrive The scaled confidence weights for each time point within the time period; where... This marks the beginning of a historical period. The total number of moments in the historical time period;

[0029] Will arrive The scaled confidence weights at each time step within the time period are multiplied by the learnable confidence correction coefficients corresponding to the forget gate and input gate, respectively, and used as... arrive Forget gate confidence correction term and input gate confidence correction term at each time point within the time period;

[0030] based on arrive The augmented input vector, forget gate confidence correction term, and previous hidden state at each time step within the time period are passed through the forget gate to obtain... arrive The output of the forget gate at each moment within the time period;

[0031] based on arrive The enhanced input vector, the input gate confidence correction term, the hidden state of the previous time in the period, pass through the input gate, to obtain to The input gate output of each time in the period;

[0032] The enhanced input vector, the cell state, the hidden state of the previous time in the period, pass through the output gate, to obtain to The output gate output of each time in the period; The cell state of each time in the period;

[0033] The enhanced input vector, the cell state, the hidden state of the previous time in the period, pass through the output gate, to obtain to The output gate output of each time in the period; Based on the output gate output and the cell state of each time in the period, the hidden state of each time in the period is calculated

[0034] Based on the output gate output and the cell state of each time in the period, the hidden state of each time in the period is calculated to The hidden state of each time in the period; wherein the hidden state of each time represents the data of each sensor at the time; The hidden state of each time and its corresponding semantic label, confidence weight, are used to construct

[0035] The hidden state of each time and its corresponding semantic label, confidence weight, are used to construct The enhanced input vector of the time; wherein , ,is a future preset time step; The hidden state of the time is multiplied by the scaled confidence weight corresponding to the hidden state, and the learnable confidence correction coefficient corresponding to the forget gate and the input gate respectively, as

[0036] The confidence correction term of the forget gate and the confidence correction term of the input gate of the time; Based on the enhanced input vector, the confidence correction term of the forget gate, and the hidden state of the time

[0037] The forget gate output of the time is obtained by passing through the forget gate; Based on the enhanced input vector, the confidence correction term of the input gate, and the hidden state of the time The input gate output of the time is obtained by passing through the input gate;

[0038] The enhanced input vector, the cell state, the hidden state of the previous time in the period, pass through the output gate, to obtain to The output gate output of each time in the period; The enhanced input vector, the cell state, the hidden state of the previous time in the period, pass through the output gate, to obtain

[0039] The output gate output of each time in the period;​​​​ Forget gate output, input gate output, candidate cell state at any given time The cell state at any given time is obtained through cell state updates. Cellular state at any given moment;

[0040] Will Time-dependent augmented input vector, cell state, The hidden state at time t is obtained through the output gate. The output gate outputs at any given time;

[0041] based on The output gate output and cell state at each time point are calculated. The hidden state at any given moment; The hidden state at a given moment is the data from each sensor within a future preset time step;

[0042] Based on the data from each sensor at each moment within a future preset time step, semantic tags are generated for the cold chain compartment at each moment within the future preset time step.

[0043] Preferably, the based Time-based augmentation of input vector, forget gate confidence correction term, The hidden state at a given time is obtained through the forget gate. The forget gate output at time step is given by the following formula:

[0044] ,

[0045] in, for The output of the forget gate at any moment, It is the Sigmoid activation function. This is the first weight matrix of the forget gate. This is the second weight matrix of the forget gate. For the bias term of the forget gate, for The boosted input vector at time step, for The hidden state at any given moment. for The forgetting gate confidence correction term at any given time. The learnable confidence correction coefficient corresponds to the forget gate. for Confidence weights after time scaling.

[0046] Preferably, the based Time-based boosting input vector, input gate confidence correction term, The hidden state at time 1 is obtained through the input gate. The input gate output at time t is given by the formula:

[0047] ,

[0048] in, for Input gate output at any time, It is the Sigmoid activation function. This is the first weight matrix of the input gate. This is the second weight matrix of the input gate. For the bias term of the input gate, for The boosted input vector at time step, for The hidden state at any given moment. for Input gate confidence correction term at time step. This is the learnable confidence correction coefficient corresponding to the input gate. for Confidence weights after time scaling.

[0049] Preferably, after calculation arrive After the hidden state at each time point within the time period, The hidden states at time t and their corresponding semantic labels and confidence weights are used to construct... time The enhanced input vector of parallel branches; where, The number of parallel branches, The weight matrices of the forget gate, input gate, and output gate corresponding to each parallel branch are the same;

[0050] Will The scaled confidence weights corresponding to the hidden states at time t are multiplied by the learnable confidence correction coefficients corresponding to the forget gate and input gate, respectively, and used as... Forget gate confidence correction term and input gate confidence correction term at different times;

[0051] based on Time-based augmentation of input vector, forget gate confidence correction term, The hidden state at any given moment, through The forget gate with parallel branches yields... time The forget gate outputs a parallel branch;

[0052] based on Time-based boosting input vector, input gate confidence correction term, The hidden state at any given moment, through The input gate of the parallel branches yields... time point the input gate output of the i-th parallel branch;

[0053] obtain time point the forget gate output, the input gate output, the candidate cell state, the cell state at the time point, by cell state update, obtain time point the cell state of the i-th parallel branch;

[0054] obtain the enhanced input vector at the time point, the cell state, the hidden state at the time point, by the output gate of the i-th parallel branch, obtain time point the output gate output of the i-th parallel branch;

[0055] based on the output gate output and the cell state of the i-th parallel branch at the time point, calculate the hidden state of the i-th parallel branch at the time point;

[0056] based on the hidden state of the i-th parallel branch at the time point, obtain the probability weight corresponding to the i-th parallel branch;

[0057] based on the hidden state of the i-th parallel branch at the time point and the probability weight corresponding to the i-th parallel branch, screen each parallel branch, and take the hidden state of the i-th parallel branch at the time point as the data of each sensor within a future preset time step.

[0058] based on the data of each sensor at each time point within a future preset time step, generate the semantic label of the cold chain vehicle cabin at each time point within the future preset time step.

[0059] Preferably, the screening of each parallel branch based on the hidden state of the i-th parallel branch at the time point and the probability weight corresponding to the i-th parallel branch comprises: respectively take the average value of the product of the scaled confidence weight and the probability weight corresponding to the hidden state of the i-th parallel branch at the time point as the confidence score of the i-th parallel branch;

[0060] respectively take the average value of the product of the scaled confidence weight and the probability weight corresponding to the hidden state of the i-th parallel branch at the time point as the confidence score of the i-th parallel branch;

[0061] respectively take the average value of the product of the scaled confidence weight and the probability weight corresponding to the hidden state of the i-th parallel branch at the time point as the confidence score of the i-th parallel branch; respectively take the average value of the product of the scaled confidence weight and the probability weight corresponding to the hidden state of the i-th parallel branch at the time point as the confidence score of the i-th parallel branch;

[0062] ​​The difference between 1 and the risk event score of each parallel branch is taken as the risk assessment index value of each parallel branch;

[0063] Based on the time of each parallel branch corresponding to reaching the target area and the acceptable maximum duration, the efficiency index value of each parallel branch is obtained, and the formula is:

[0064] ,

[0065] Among them, is the efficiency index value of the i-th parallel branch, is the time of the i-th parallel branch corresponding to reaching the target area, is the starting transportation time corresponding to the i-th parallel branch, is the acceptable maximum duration; The ratio of the number of times that the hidden state of each parallel branch meets the regulatory requirements at the t-th moment to the total number of regulatory requirements is taken as the compliance index value of each parallel branch; The confidence score, risk assessment index value, efficiency index value, and compliance index value of each parallel branch are weighted and summed to obtain a comprehensive score of each parallel branch;

[0066] Based on the comprehensive score of each parallel branch, a pruned branch set is obtained;

[0067] Through a dynamic programming algorithm, the optimal branch in the pruned branch set is obtained, and the hidden state of the optimal branch at the t-th moment is taken as the data of each sensor within a future preset time step;

[0068] Based on the data of each sensor at each moment within a future preset time step, the semantic label of the cold chain vehicle cabin at each moment within the future preset time step is generated.

[0069] Preferably, after obtaining the data of each sensor and the semantic label of the cold chain vehicle cabin within the future preset time step, the data of each sensor and the semantic label of the cold chain vehicle cabin within the future preset time step, the comprehensive score of each parallel branch, the confidence score of each parallel branch, the risk assessment index value, the efficiency index value, the compliance index value, the corresponding weight, the pruning parameter, the dynamic programming algorithm parameter, and the optimal branch index are taken as the original data;

[0070]

[0071] Preferably, after obtaining the data of each sensor and the semantic label of the cold chain vehicle cabin within the future preset time step, the data of each sensor and the semantic label of the cold chain vehicle cabin within the future preset time step, the comprehensive score of each parallel branch, the confidence score of each parallel branch, the risk assessment index value, the efficiency index value, the compliance index value, the corresponding weight, the pruning parameter, the dynamic programming algorithm parameter, and the optimal branch index are taken as the original data;

[0072] ​​​​The data of each sensor of the parallel branch is hashed with the semantic label of the cold chain vehicle cabin and the optimal branch index, and the generated hash value is used as a leaf node together with the comprehensive score of each parallel branch within a future preset time step, the confidence score of each parallel branch, the risk assessment index value, the efficiency index value, the compliance index value, the corresponding weight, the pruning parameter and the dynamic programming algorithm parameter to generate the Merkle root hash of the original data through the Merkle tree structure; when the original data is called, whether the original data is tampered with is verified based on the Merkle root hash of the original data.

[0073] The application further provides a cold chain logistics monitoring system, comprising:

[0074] The application further provides a computer program product, comprising a computer program and a storage medium, wherein the storage medium is used to store the computer program, and the computer program is used to realize the steps of the cold chain logistics monitoring method.

[0075] Compared with the prior art, the above technical scheme of the application has the following beneficial effects:

[0076] The cold chain logistics monitoring method and system can improve the sensitivity of data loss by weighting and aggregating the proportion of effective data of each sensor to construct a data integrity index, which quantifies data availability and highlights the contribution of key sensor signals through the importance weight of the sensor. In the aspect of abnormal risk perception, the threshold exceeding degree index is used to realize fine risk prediction. The deviation amplitude of each sensor data from the preset normal interval is calculated, and after normalization processing to eliminate the difference in the magnitude of different parameters, the threshold exceeding degree index is constructed by weighting and summing the importance weight of the sensor, which can dynamically reflect the abnormal fluctuation of different parameters and enable the model to capture potential over-standard trends that have a greater impact on cold chain safety in advance. In the aspect of multi-source data consistency verification, the standard deviation of redundant sensor data is calculated to measure the dispersion degree of multiple sensors measuring the same physical quantity, which directly reflects the data consistency in the spatial dimension. The cross-modal semantic label verification combines the real-time working condition label of the cold chain vehicle (such as "loading and unloading" and "high-speed transportation") to verify whether the sensor data change matches the current scene. The data change amount at adjacent time is calculated by the time sequence fluctuation amplitude of the sensor data to evaluate the data smoothness in the time dimension. The confidence weight formed by weighting the three is combined with the data of each sensor and the semantic label of the cold chain vehicle cabin to form an enhanced input vector. This input design retains the time sequence characteristics of the monitoring data, incorporates data quality evaluation results through the confidence weight, and supplements the working condition background information with the semantic label, so that the prediction model can adaptively focus on high-quality and consistent effective signals and actively weaken the interference of abnormal or low-confidence data. Finally, a more robust and more practical transportation condition prediction of the future data of each sensor and the semantic label of the cold chain vehicle cabin is realized, which effectively meets the safety monitoring needs in complex and dynamic scenarios of cold chain logistics.

[0077] In order to solve the problem that the existing method can only output a single path or a single future state in future state prediction, lack of modeling ability of multi-trajectory distribution under complex environment in cold chain transportation, leading to risk warning lag or even failure, and difficult to meet the stringent requirements of cold chain logistics on abnormal chain. The present application improves the LSTM model, multiplies the scaled confidence weight with the learnable confidence correction coefficient corresponding to the forgetting gate and the input gate respectively, as the forgetting gate confidence correction term and the input gate confidence correction term, and integrates into the gate calculation process. This design allows the model to dynamically adjust the information processing strategy according to the data quality, so that the influence of the information is retained when the confidence is high, and the contribution of the information is weakened when the confidence is low, thereby improving the robustness to abnormal / missing data. On the basis of obtaining reliable hidden state, multiple future branches are developed in parallel, and each parallel branch represents a possible future trajectory. The main parameters are shared to ensure the consistency of each branch in time sequence, physical logic, such as different trajectories conforming to the temperature regulation characteristics of the cold chain refrigeration system, avoiding the dispersion and disorder of multi-trajectory distribution, reducing the parameter redundancy, and reducing the calculation cost in the reasoning stage. The differential design of branch bias term allows each branch to flexibly cover different future scenarios caused by complex factors such as road condition changes, loading and unloading operations, and slight fluctuations of equipment, breaking the limitations of traditional single path prediction, realizing comprehensive modeling of cold chain transportation multiple possibilities, improving the single branch prediction accuracy through data quality perception, and covering complex scenarios through multi-branch parallel, effectively solving the risk warning lag problem, and providing reliable technical support for the stringent monitoring requirements of cold chain logistics abnormal chain. BRIEF DESCRIPTION OF DRAWINGS

[0078] In order to make the content of the present application more easily understood, the following will further explain the present application in detail according to the specific embodiments of the present application and in conjunction with the drawings, in which:

[0079] Figure 1 It is a system architecture diagram of a cold chain logistics monitoring method of the present application.

[0080] Figure 2 It is a schematic diagram of a data acquisition and semantic label generation unit.

[0081] Figure 3 It is a schematic diagram of a cloud future state superposition prediction unit.

[0082] Figure 4 It is a schematic diagram of a decision collapse and compliance certificate generation unit.

[0083] Figure 5 It is a schematic diagram of an end-cloud-end consistency and degradation checking unit.

[0084] Figure 6is a monitoring flowchart of vaccine cold chain transportation.

[0085] Figure 7 is a monitoring flowchart of cold chain truck fleet joint distribution scenarios. DETAILED DESCRIPTION

[0086] The present application will be further described below with reference to the drawings and specific embodiments, so that those skilled in the art can better understand the present application and implement it. The embodiments are not limiting to the present application.

[0087] Referring to Figure 1 , the embodiment one provides a cold chain logistics monitoring method, comprising:

[0088] The embodiment collects multi-source data related to the environment and the transportation state in real time, generates semantic labels and confidence weights automatically after filtering, aligning and fusing processing, and finally outputs an enhanced input vector of each time in a historical time period, which provides a key input for subsequent prediction and decision models, and ensures that the system can perform real-time and reliable environment monitoring and risk warning, and provides an interpretable input for the prediction model.

[0089] The collected multi-source data includes but is not limited to: temperature , humidity , location , speed , cabin door state , task phase and other data. The temperature and humidity are provided by the vehicle cabin environment sensor, the location information and speed are jointly provided by the GPS module and the inertial navigation unit, the cabin door state is collected by the reed switch or Hall sensor, and the task phase information is issued by the transportation scheduling system (such as "loading", "transporting", "unloading", etc.). Among them, is the temperature of the cold chain vehicle cabin at the th time in the historical time period, is the humidity of the cold chain vehicle cabin at the th time in the historical time period, is the location of the cold chain transportation vehicle at the th time in the historical time period, is the horizontal coordinate of the cold chain transportation vehicle at the th time in the historical time period, is the vertical coordinate of the cold chain transportation vehicle at the th time in the historical time period, is the speed of the cold chain transportation vehicle at the th time in the historical time period, is the cabin door state of the cold chain transportation vehicle at the th time in the historical time period.

[0090] The data collected by each sensor at each time in the historical time period is preprocessed to obtain preprocessed data of each sensor at each time in the historical time period, including:

[0091] For analog sensor data such as temperature and humidity that are easily affected by device jitter, Kalman filtering or moving average filtering algorithm is used for smoothing processing to eliminate sensor noise and instantaneous jitter; for example, the abnormal short-time jump of the temperature sensor can be smoothed as the real change trend.

[0092] Abnormal data (such as temperature less than -50℃ or greater than 150℃) or missing data beyond the physical reasonable range are regarded as invalid data, and are filled in by redundant sensors or interpolation method.

[0093] Based on the preprocessed data of each sensor at each time in the historical time period, semantic labels are generated for each time in the historical time period according to the set rules The semantic labels are used to describe the current transportation state, which facilitates the subsequent model to understand the specific situation of transportation. The rules for generating semantic labels are usually defined according to the working state of the vehicle and the output value of the sensor. These labels can help the system understand the current state of the vehicle and provide context information for subsequent prediction and decision-making, for example:

[0094] Based on the door state sensor (magnetic switch) data and GPS positioning data, when the door is opened (magnetic switch outputs high level) and the GPS positioning coordinates fall within the preset warehouse geographic fence, the semantic label is "loading"; based on the vehicle speed sensor data and door state data, when the door is closed (magnetic switch outputs low level) and the speed is greater than 5 km / h, the semantic label is "in transit"; based on the temperature sensor data and the regulatory threshold, when the temperature exceeds the regulatory threshold (such as 8℃, the duration can be configured, and the default is 10 minutes), the semantic label is "over temperature"; based on the GPS positioning data and the spatial distance of the preset transportation path, when the vehicle deviates from the planned path by more than 2 kilometers, the semantic label is "deviation".

[0095] Based on the corresponding semantic labels and preprocessed data at each time in the historical time period, the confidence weight of each time in the current sliding event is generated. The confidence considers data integrity, threshold exceeding degree and multi-source consistency. The specific scheme is as follows:

[0096] ​Step S1: Multiply the percentage of valid data (the ratio of valid data to the total data) of each sensor in the cold chain compartment within the historical time period by the importance weight of that sensor. Then, sum the products for all sensors to obtain the data integrity index value for each moment within the historical time period. The data integrity index measures the completeness and validity of data collection and is defined as the historical time period. ( In this embodiment, The percentage of valid data within 10 seconds is calculated using the following formula:

[0097] ,

[0098] in, Within the historical period Data integrity index value at any given time. Indicates time, This is a collection of sensors for the cold chain vehicle compartment. For the first time in the historical period Total number of valid data from each sensor For the first time in the historical period Total number of data from each sensor For the first time in the historical period The percentage of effective data from each sensor For the first The importance weight of each sensor Indicates the sensor index. This refers to a historical period.

[0099] For example, if the effective data percentage of the temperature sensor in the cold chain truck compartment during a historical period is 0.95, and the importance weight of the temperature sensor is 0.5; the effective data percentage of the position sensor is 0.85, and the importance weight of the position sensor is 0.3; and the effective data percentage of the humidity sensor is 0.98, and the importance weight of the humidity sensor is 0.2, then... .

[0100] Step S2: Based on the preprocessed data of each sensor at each moment within the historical time period and the corresponding set numerical range, obtain the threshold exceedance index value for each moment within the historical time period. This index measures the degree to which the sensor data at the current moment exceeds the set numerical range. Unlike simply judging whether it exceeds the limit, this embodiment sets the numerical range. Set as a regulatory threshold range, and use quantitative calculations to reflect the compliance of data by exceeding the allowed range, including:

[0101] Based on the preprocessed data of each sensor at each moment within a historical time period and the corresponding set numerical range, the exceedance range of each sensor at each moment within the historical time period is calculated using the following formula:

[0102] ,

[0103] represents the amplitude of the th time point in the historical time period exceeding the regulatory range.

[0104] The normalized exceeding amplitude of each sensor at each time point in the historical time period is obtained by normalizing the exceeding amplitude of each sensor at each time point in the historical time period, and the formula is:

[0105] ,

[0106] The threshold exceeding degree index value at each time point in the historical time period is calculated based on the importance weight of each sensor and the normalized exceeding amplitude of each sensor at each time point in the historical time period, and the formula is:

[0107] ,

[0108] Since the confidence should increase as the compliance increases, the final entering index is defined as the complementary form of the normalized exceeding amplitude, when , when , .

[0109] wherein, is the threshold exceeding degree index value at the th time point in the historical time period, is the exceeding amplitude of the th sensor at the th time point in the historical time period, is the lower limit value of the set value interval corresponding to the th sensor, is the upper limit value of the set value interval corresponding to the th sensor, is the data of the th sensor at the th time point in the historical time period, is a truncation function to limit the output result to [0, 1], is the maximum allowed exceeding amount of the th sensor, is the normalized exceeding amplitude of the th sensor at the th time point in the historical time period, is the importance weight of the th sensor, Indicates an index representing a moment within a historical time period. Indicates the sensor index. This is a collection of sensors for use inside the cold chain vehicle compartment.

[0110] Step S3: Based on the standard deviation of the preprocessed data of redundant sensors that measure the same physical quantity with each sensor within the historical time period, the semantic tags of the cold chain compartment at each time point within the historical time period, and the changes in the preprocessed data of each sensor at adjacent times within the historical time period, obtain the multi-source consistency index value at each time point within the historical time period, including:

[0111] The standard deviation of the preprocessed data from redundant sensors measuring the same physical quantity as each sensor within the historical time period is normalized at each moment to obtain the redundancy sensor consistency index value for each sensor within the historical time period. The formula is as follows:

[0112] ,

[0113] The redundancy sensor consistency index value at each moment in the historical time period is obtained by weighting and summing the importance weight of each sensor with the consistency index value of each sensor in the historical time period. The formula is as follows:

[0114] ,

[0115] Based on the semantic labels of each moment within a historical time period and its previous moment, it is determined whether there are any anomalies in the changes of the preprocessed data of each sensor at each moment within the historical time period. If so, the cross-modal consistency index value at that moment within the historical time period is set accordingly. If it is 0, then set the cross-modal consistency index value at that moment within the historical time period to 0. =1;

[0116] By establishing a mapping rule base for semantic label transformation and expected data change, that is, based on the characteristics of cold chain transportation scenarios, pre-setting reasonable change trends of sensor data under different semantic label switching scenarios, and performing consistency verification of sensor data change at each moment based on the mapping rule base, if the change of all sensors conforms to the expected trend under the current semantic label switching scenario, then cross-modal consistency is determined to be normal; if the change of any key sensor deviates from the expected trend, then cross-modal consistency is determined to be abnormal. The value is 0; for example, after the hatch is opened, the temperature at the current moment should increase compared to the previous moment. If the data does not show this trend, it is recorded as 0; if it meets the expectation, it is recorded as 1.

[0117] The absolute value of the difference between the pre-processed data of each sensor at adjacent time points in the historical time period is normalized to obtain the time smoothing consistency index value of each sensor at each time point in the historical time period, and the formula is:

[0118] ,

[0119] The importance weight of each sensor is weighted and summed with the time smoothing consistency index value of each sensor in the historical time period to obtain the time smoothing consistency index value at each time point in the historical time period, and the formula is:

[0120] ,

[0121] The redundant sensor consistency index value, the cross-modal consistency index value, and the time smoothing consistency index value at each time point in the historical time period are weighted and summed to obtain the multi-source consistency index value at each time point in the historical time period , and the calculation formula is: , is the weight of the redundant sensor consistency index, is the weight of the cross-modal consistency index, is the weight of the time smoothing consistency index.

[0122] The multi-source consistency index measures the consistency between data from different sensors or different time dimensions, and can reflect the credibility of the overall system data.

[0123] wherein, is the redundant sensor consistency index value of the i-th sensor at the j-th time point in the historical time period, is the pre-processed data of the i-th redundant sensor measuring the same physical quantity as the j-th sensor at the j-th time point in the historical time period, is a truncation function to limit the output result to [0, 1], , is the average value of the pre-processed data of the i-th redundant sensor measuring the same physical quantity as the j-th sensor at the j-th time point in the historical time period, is the maximum allowed variance of the j-th sensor, is the total number of redundant sensors measuring the same physical quantity as the j-th sensor, is the time smoothing consistency index value of the j-th sensor at the i-th time point in the historical time period, is the average value of the pre-processed data of the i-th redundant sensor measuring the same physical quantity as the j-th sensor at the j-th time point in the historical time period, is the maximum allowed variance of the j-th sensor, is the total number of redundant sensors measuring the same physical quantity as the j-th sensor, is the time smoothing consistency index value of the j-th sensor at the i-th time point in the historical time period, is the time smoothing consistency index value of the j-th sensor at the i-th time point in the historical time period, ​​​​​a redundancy sensor consistency index value of the moment, a data of the first sensor in the historical time period, a data of the first sensor in the historical time period, a data of the first sensor in the historical time period, a data of the first sensor in the historical time period, a data of the first sensor in the historical time period, a data of the first sensor in the historical time period, an absolute value, a maximum change amplitude allowed by the first sensor, a maximum change amplitude allowed by the first sensor, a time smoothing consistency index value of the first sensor in the historical time period, a time smoothing consistency index value of the first sensor in the historical time period, a time smoothing consistency index value of the first sensor in the historical time period, a time smoothing consistency index value of the first sensor in the historical time period, a time smoothing consistency index value of the first sensor in the historical time period, an importance weight of the first sensor, an importance weight of the first sensor, an index of a moment of a historical time period, an index of a sensor, a sensor set in a cold chain vehicle cabin.

[0124] Step S4: Weighted sum of the data integrity index value, the threshold exceeding degree index value and the multi-source consistency index value of each moment in the historical time period to obtain a confidence weight of each moment in the historical time period, the formula is:

[0125]

[0126] In order to ensure that the three indexes can be weighted calculated, all indexes need to be normalized to [0, 1] before entering the confidence calculation, 1 represents more reliable, a confidence weight of the first moment in the historical time period. a confidence weight of the first moment in the historical time period.

[0127] Step S5: The data of each sensor in each moment in the historical time period and the semantic label of the cold chain vehicle cabin, the confidence weight and the semantic label of the cold chain vehicle cabin are used to construct elements of an enhanced input vector of each moment in the historical time period; wherein the enhanced input vector of the first moment in the historical time period is , The data of each sensor, the vector not only contains original numerical information, but also carries semantic label and confidence weight, which can not only describe the objective state of the cold chain environment, but also reflect its reliability, thereby providing higher explainability and robustness for subsequent prediction and decision-making.

[0128] ​​Step S6: Based on the enhanced input vectors at each moment in the historical time period, obtain the data of each sensor and the semantic labels of the cold chain compartment within the future preset time step.

[0129] In this embodiment, optionally, the enhanced input vectors at each moment in the historical time period are passed through an LSTM model to obtain the data of each sensor and the semantic labels of the cold chain compartment within a future preset time step.

[0130] In this embodiment, preferably, based on the enhanced input vectors at each moment within a historical time period, the data from each sensor and the semantic labels of the cold chain vehicle compartment within a future preset time step are obtained by improving the LSTM model (Long Short-Term Memory with Time-Superposition and Tagging, LSTM-TT), including:

[0131] Historical time period ( arrive The confidence weight at each moment within a given time period is scaled using a normalization function to obtain the scaled confidence weight at each moment within the historical time period. The calculation formula is as follows:

[0132] ,

[0133] in, For the first time in the historical period Confidence weights scaled up at each time point This marks the beginning of a historical period. The total number of moments in the historical time period;

[0134] In specific applications, the Sigmoid function or power transformation can also be used. Calibration is performed to enhance sensitivity to boundary conditions.

[0135] Will arrive The scaled confidence weights at each time step within the time period are multiplied by the learnable confidence correction coefficients corresponding to the forget gate and input gate, respectively, and used as... arrive Forget gate confidence correction term and input gate confidence correction term at each time point within the time period;

[0136] based on arrive The augmented input vector, forget gate confidence correction term, and previous hidden state at each time step within the time period are passed through the forget gate to obtain... arrive The output of the forget gate at each moment within the time period;

[0137] The LSTM-TT model possesses temporal memory properties; its hidden state and cell state at each time step are derived recursively from the state at the previous time step. The initial hidden state (e.g., at the start of the task) is... and cell state Typically initialized as a zero vector or trainable parameters. The hidden state at each recursive time step comes from the model's calculation of the augmented input vector from the previous time step. It is a dynamic encoding of historical information by the model, carrying the environmental state, semantic information, and confidence features from previous time steps, providing temporal context for the prediction at the current time step.

[0138] based on arrive The augmented input vector, input gate confidence correction term, and previous hidden state at each time step within the time period are passed through the input gate to obtain... arrive The input gate output at each time point within the time period;

[0139] Will arrive The forget gate output, input gate output, candidate cell state, and previous cell state at each time step within the time period are obtained through cell state updates. arrive Cell state at each moment within the time period;

[0140] Will arrive The augmented input vector, cell state, and previous hidden state at each time step within the time period are passed through the output gate to obtain... arrive The output gate outputs at each time point within the time period;

[0141] based on arrive The output gate output and cell state at each time point within the time period are calculated. arrive The hidden state at each moment within the time period; where the hidden state at each moment represents the data of each sensor at that moment;

[0142] Based on the data from each sensor at each moment within a future preset time step, semantic tags for the cold chain compartment at each moment within the future preset time step are generated.

[0143] by The hidden states at time t and their corresponding confidence weights are used to construct... The boosted input vector at time t; where, , Predetermine the time step for the future;

[0144] Will The scaled confidence weight corresponding to the hidden state at the moment is multiplied by the learnable confidence correction coefficient corresponding to the forget gate and the input gate respectively, as The forget gate confidence correction term and the input gate confidence correction term at the moment t;

[0145] Based on The enhanced input vector, the forget gate confidence correction term, and the input gate confidence correction term at the moment t, The hidden state at the moment t, through the forget gate, obtains The forget gate output at the moment t, the formula is:

[0146] ,

[0147] Among them, The forget gate output at the moment t, Sigmoid activation function, The first weight matrix of the forget gate, The second weight matrix of the forget gate, The bias term of the forget gate, The enhanced input vector at the moment t, The hidden state at the moment t, The forget gate confidence correction term at the moment t, The learnable confidence correction coefficient corresponding to the forget gate, The scaled confidence weight at the moment t. Based on The enhanced input vector, the input gate confidence correction term, and the input gate confidence correction term at the moment t, The hidden state at the moment t, through the input gate, obtains The input gate output at the moment t, the formula is:

[0148] , Among them, The input gate output at the moment t,

[0149] Sigmoid activation function, The first weight matrix of the input gate, The second weight matrix of the input gate,

[0150] The bias term of the input gate, The enhanced input vector at the moment t, The hidden state at the moment t, The input gate confidence correction term at the moment t, The learnable confidence correction coefficient corresponding to the input gate, The scaled confidence weight at the moment t. , Among them, The input gate output at the moment t, ​​​Input gate confidence correction term at time step. The input gate corresponds to the learnable confidence correction coefficient. for Confidence weights after time scaling.

[0151] Will Forget gate output, input gate output, candidate cell state at any given time The cell state at any given time is obtained through cell state updates. The cell state at time t is given by the formula:

[0152] ,

[0153] in, for Cellular state at any given moment For Hadamard element-wise multiplication, for The state of candidate cells at any given time. for The state of a cell at any given moment.

[0154] The process of updating the candidate cell state is as follows: Time-based boosting input vector The hidden state at time t is obtained by calculating the candidate cell states. The candidate cell state at time t is given by the formula: ,in, For the tanh function, This is the first weight matrix for the candidate cell states. This is the second weight matrix for the candidate cell states. This is a bias term for the candidate cell state;

[0155] Will Time-dependent augmented input vector, cell state, The hidden state at time t is obtained through the output gate. The output of the gate at time is calculated using the following formula: ,in, for The output gate outputs at any time. This is the first weight matrix of the output gate. This is the second weight matrix of the output gate. This is the bias term for the output gate.

[0156] based on The output gate output and cell state at each time point are calculated. The hidden state at time t is given by the formula: , The hidden state of the moment is the data of each sensor within the future preset time step. Based on the data of each sensor at each moment within the future preset time step, the semantic label of the cold chain vehicle cabin at each moment within the future preset time step is generated by setting a preset rule or by a pre-trained deep learning model. By introducing a confidence correction term, the gating state no longer depends only on the input and the historical hidden state, but also dynamically considers the data reliability, making the prediction more robust.

[0157] The existing method relies on Markov chain, graph model or single trajectory prediction algorithm based on statistics for future state prediction. Such methods can only output a single path or a single future state, and lack the ability to model the multi-trajectory distribution in the complex environment of cold chain transportation. When the transportation environment changes suddenly (such as sudden deviation, refrigerator failure), single trajectory prediction often cannot cover multiple possibilities, resulting in insufficient early warning. If the existing multi-trajectory prediction method is used, such as MDN (mixed density network), MultiPath++, there are obvious shortcomings: first, such methods usually generate multiple trajectories through probability distribution in the model output layer, and still take single trajectory prediction error as the core optimization target in the training stage, without parallel learning of multiple trajectories from the bottom architecture, resulting in distribution dispersion of generated multiple trajectories - some trajectories are disconnected from the physical laws of cold chain (such as sudden temperature drop after refrigerator failure), which is low in practicality; second, a large number of samples (such as Gaussian mixture sampling of MDN) or beam search (such as path search of MultiPath++) are needed in the reasoning stage to obtain effective trajectories, which not only increases the calculation delay (difficult to meet the millisecond-level response requirement of cold chain real-time monitoring), but also may miss key risk trajectories (such as extreme temperature fluctuation trajectories) due to sampling randomness; third, the generated multiple trajectories lack strong correlation with the cold chain scene, and cannot clearly indicate the working condition background corresponding to the trajectory, making it difficult for operation personnel to quickly judge the priority of the trajectory and delay risk disposal.

[0158] Therefore, the present application realizes the bottom optimization of multi-trajectory prediction by reconstructing the LSTM unit architecture. A time superposition layer (TSL) is introduced in the LSTM unit to parallelly expand The model employs multiple parallel branches, each representing a possible future evolution path. Each branch shares the backbone parameters, with only the bias terms privatized to capture the multimodal distribution structure. Finally, each branch is normalized using softmax to obtain probability weights, thus forming the probability distribution of the future trajectory. This approach breaks the limitations of traditional models where the output layer generates multiple trajectories. During the training phase, the model forms parallel future branches during temporal feature learning—each parallel branch corresponds to a potential cold chain operating state (e.g., "normal refrigeration branch," "inefficient refrigeration branch," "sudden door opening branch"). Simultaneously, each branch shares gating parameters (weight matrices for the forget gate, input gate, and output gate), avoiding the trajectory logic confusion caused by independent branch training and ensuring consistency in physical characteristics such as "temperature change rate" and "parameter correlation patterns," conforming to the actual operating logic of cold chain transportation. Furthermore, real-time semantic tags for cold chain vehicles (e.g., "high-speed transportation," "stopping and unloading") and confidence weights at each time point are deeply embedded in the branch calculation process, enabling each parallel branch to clearly associate the "scenario-data quality-prediction result" relationship, significantly improving the interpretability of multiple trajectories. This design fundamentally avoids reliance on sampling or beam search during the inference phase, providing operators with more accurate and interpretable decision-making support. The specific solution is as follows:

[0159] In this embodiment, preferably, after calculation... arrive After the hidden state at each time point within the time period, The hidden states at time t and their corresponding semantic labels and confidence weights are used to construct... time The enhanced input vectors of the parallel branches; the weight matrices of the forget gate, input gate, and output gate corresponding to each parallel branch are the same; in this embodiment, let... ;

[0160] Will The scaled confidence weights corresponding to the hidden states at time t are multiplied by the learnable confidence correction coefficients corresponding to the forget gate and input gate, respectively, and used as... Forget gate confidence correction term and input gate confidence correction term at different times;

[0161] based on Time-based augmentation of input vector, forget gate confidence correction term, The hidden state at any given moment, through The forget gate with parallel branches yields... time The forget gate outputs multiple parallel branches; where, The weight matrices of the forget gate, input gate, and output gate corresponding to each parallel branch are the same;

[0162] based on Time-based boosting input vector, input gate confidence correction term, The hidden state at any given moment, through The input gate of the parallel branches yields... time Input gate output with parallel branches;

[0163] Will time Forget gate output, input gate output, candidate cell state, and parallel branch. The cell state at any given time is obtained through cell state updates. time Cell states with parallel branches;

[0164] Will Time-dependent augmented input vector, cell state, The hidden state at any given moment, through The output gate of the parallel branch obtains time The output gate has multiple parallel branches;

[0165] based on At any given moment, the output gate of each parallel branch, the cell state, and the calculation are performed. The hidden state of each parallel branch at any given time;

[0166] based on The hidden state of each parallel branch at any given time is determined, and the probability weight corresponding to each parallel branch is obtained.

[0167] based on Given the hidden state of each parallel branch at each time step and the probability weight corresponding to that branch, filter each parallel branch and select the branch corresponding to the hidden state at each time step. The hidden state at a given moment serves as the data from each sensor within a future preset time step.

[0168] based on Hidden state of each parallel branch at time step , Obtain the probability weights corresponding to each parallel branch; parallel branches at each time step The hidden state itself represents the parallel branch at that moment. The predicted values ​​are the data from each sensor, but they are not directly involved in the softmax calculation; among them, for Time-of-flight parallel branch The hidden state;

[0169] Will The hidden state of each parallel branch at time t is calculated by a linear layer or a small network at the end of each parallel branch , The confidence score of each parallel branch , , The confidence score of each parallel branch The corresponding trainable parameters

[0170] The confidence score of each parallel branch is normalized by softmax to obtain the corresponding probability weight of each parallel branch, and the formula is as follows: , , is the index of the parallel branch The confidence score of each parallel branch , The corresponding probability weight of each parallel branch , which means the probability of the occurrence of the first future trajectory or the credibility weight of the first trajectory mode, and the weight satisfies , which reflects the confidence distribution of the model in all possible future evolution modes. The final multi-trajectory prediction result can be represented as ,

[0171] , represents The vector composed of the corresponding sensor parameters and semantic labels of the parallel branch at time t , which reflects the judgment of the model on different future evolution paths, rather than the normalization of the predicted value itself. In this embodiment, specifically, to ensure the accuracy, coverage, diversity and probability calibration of multi-trajectory prediction, a composite loss function is used in the training stage of the LSTM-TT model, and the composite loss function is as follows:

[0172] ,

[0173] ,

[0174] Among them, is the composite loss function, is the error between the predicted trajectory (the predicted data of each sensor and the semantic label of the cold chain vehicle cabin within the preset future time step) and the true trajectory (the true data of each sensor and the semantic label of the cold chain vehicle cabin within the preset future time step), is used to improve the trajectory coverage, such as imposing constraints on the minimum average deviation minADE coverage and top-k coverage, is used to avoid trajectory overlap, such as DTW / Fréchet distance regularization, ​Used to improve the accuracy of probability distribution calibration, such as desired calibration error ECE or temperature scaling. These are the weights of the loss function.

[0175] In this embodiment, preferably, the basis The hidden state and its corresponding probability weight for each parallel branch at each time step are used to filter each parallel branch, including:

[0176] Each The average of the products of the scaled confidence weight and the probability weight corresponding to the hidden state of each parallel branch at time step 1 is used as the confidence score of each parallel branch, as shown in the formula:

[0177] ,

[0178] in, For parallel branches The confidence score, For time index, for Time-of-flight parallel branch The scaled confidence weights corresponding to the hidden states. for Time-of-flight parallel branch The probability weights.

[0179] The confidence score of each parallel branch can comprehensively consider the input confidence and the model output probability, and measure the reliability of the trajectory-dependent data and the prediction.

[0180] Will The average number of times the data from each sensor in the hidden state of each parallel branch exceeds its threshold range is used as the risk event score for each parallel branch.

[0181] The difference between 1 and the risk event score of each parallel branch is used as the risk assessment index value of each parallel branch. The calculation formula is:

[0182] ,

[0183] in, For parallel branches Risk assessment indicator values, For indicator functions, for Time-of-flight parallel branch The data from each sensor corresponding to the hidden state. This represents the lower limit of the threshold range corresponding to the data from each sensor. This represents the upper limit of the threshold range corresponding to the data from each sensor.

[0184] For example, if the temperature is within its safe range, then the temperature contributes zero risk at that moment; if it exceeds the threshold range, it is recorded as a risk event. The value range is [0,1]. The smaller the value, the lower the risk, and it can measure the risk of the trajectory violating the cold chain constraints in the future.

[0185] Based on the arrival time to the target region and the maximum acceptable duration for each parallel branch, the efficiency index value for each parallel branch is obtained, using the following formula:

[0186] ,

[0187] in, For the first The efficiency index value of parallel branches. For the first The arrival time of each parallel branch to the target region. For the first The starting transport time corresponding to each parallel branch. The maximum acceptable duration;

[0188] The efficiency index value of each parallel branch is used to quantify transportation efficiency, energy consumption, or arrival time. It is calculated by subtracting the ratio of the predicted actual transportation time of each parallel branch to the maximum allowable time calculated from the start time from 1.

[0189] Will The ratio of the number of times the hidden state of each parallel branch satisfies regulatory requirements to the total number of regulatory requirements is used as the compliance indicator value for each parallel branch. The formula is as follows:

[0190] ,

[0191] in, For parallel branches Compliance indicator values, The total number required by regulations, For parallel branches Number of times regulatory requirements are met The closer it is to 1, the more compliant it is.

[0192] The system determines whether a trajectory complies with drug administration and cold chain food regulations by defining constraints (regulatory requirements). The judgment logic can be modeled based on actual standard clauses or can be freely set and can be completed automatically. For example:

[0193] If the transport temperature for frozen foods is defined as not exceeding -18℃, and the transport temperature for refrigerated foods is defined as being maintained between 0℃ and 10℃, then the formula for determining this is: If the temperature exceeds the specified range, it is determined that it does not meet the regulatory requirements, representing a branch whether a particular regulatory requirement item is met, for parallel branches corresponding temperature.

[0194] If the transport vehicle is set not to deviate from the planned route by more than 2 km, and must record trajectory information throughout the journey, if the transport vehicle deviates by more than 2 km or there is a loss of position data, it is determined that it does not meet the regulatory requirements.

[0195] The confidence score, risk assessment index value, efficiency index value, and compliance index value of each parallel branch are weighted and summed to obtain a comprehensive score for each parallel branch, and the formula is:

[0196]

[0197] wherein, the comprehensive score of the parallel branch, is a weight coefficient.

[0198] wherein, in order to ensure that indicators of different dimensions can be weighted and integrated in the scoring function, all indicators need to be normalized before entering the scoring function: for indicators that are proportional in themselves (such as , ), the output is naturally in the interval [0, 1], and no additional normalization is required; for original value type indicators (such as transportation time, energy consumption, etc.), Min-Max normalization is required: , is the output after Min-Max normalization, is the input data, is the minimum value of the input data, is the maximum value of the input data, and after normalization, all indicator values fall within the range [0, 1], ensuring the comparability and stability of the scoring calculation.

[0199] In order to reduce redundant trajectories and improve decision-making efficiency, the present application adopts a "two-stage strategy" based on the comprehensive score of each parallel branch to obtain a pruned branch set, including: removing trajectory branches with a comprehensive score below a set comprehensive score threshold , DTW (Dynamic Time Warping) or Fréchet distance measurement, if the similarity is higher than a set similarity threshold (such as 0.8), only the one with the highest comprehensive score is retained.

[0200] Through dynamic programming (DP) algorithm, in the pruned set, search for the trajectory branch with the lowest cost and the highest score to obtain the optimal branch in the pruned branch set; ​​

[0201] The transfer cost function for each parallel branch The calculation formula is:

[0202] ,

[0203] in, For parallel branch temperature changes, For the parallel branch speed variation, Penalties for violating cold chain thresholds, These are the weighting coefficients.

[0204] Choose the trajectory branch with the lowest transition cost and the highest overall score. The parallel branch with the lowest cost and highest overall score is selected as the optimal branch. This is the optimal branch index.

[0205] The optimal branch corresponds to The hidden state at each moment serves as the data from each sensor within a future preset time step. Based on the data from each sensor at each moment within the future preset time step, semantic tags for the cold chain compartment at each moment are generated.

[0206] Compared with traditional single-trajectory prediction methods, this model can not only generate multiple possible future evolution paths, but also dynamically adjust the prediction results by combining semantic and confidence information, thereby improving robustness and interpretability.

[0207] In this embodiment, preferably, after obtaining the data of each sensor and the semantic tags of the cold chain vehicle compartment within a future preset time step, the data within the future preset time step are... The data from each sensor in each parallel branch, the semantic tags of the cold chain compartment, the comprehensive score of each parallel branch, the confidence score of each parallel branch, the risk assessment index value, the efficiency index value, the compliance index value and their corresponding weights, the pruning parameters, the dynamic programming algorithm parameters, and the optimal branch index are used as the original data.

[0208] Within the preset time step in the future The data from each sensor in each parallel branch, along with the semantic tags of the cold chain vehicle compartment and the optimal branch index, are hashed. The generated hash value, along with the comprehensive score of each parallel branch, the confidence score of each parallel branch, the risk assessment index value, the efficiency index value, the compliance index value and their corresponding weights, pruning parameters, and dynamic programming algorithm parameters within a preset future time step, are used as leaf nodes. Through a Merkle tree structure, a Merkle root hash (compliance certificate) of the original data is generated. When the original data is accessed, the Merkle root hash of the original data is used to verify whether the original data has been tampered with.

[0209] ,

[0210] wherein, is the Merkle root hash of the moment original data, is a Merkle tree structure, is a hash processing;

[0211] The compliance certificate can be stored in the cloud or the blockchain for subsequent regulatory verification, traceability audit and three-end consistency check. The certificate not only contains the final result, but also encapsulates the whole process logic. It can not only check whether the optimal result is consistent, but also verify whether the discarded track score and elimination process are reasonable through sampling, so as to realize the transparency and traceability of the whole process.

[0212] The existing method of blockchain or log storage usually only stores the hash of the final result or key field, and cannot trace back the prediction process. The regulatory party cannot understand "why other trajectories are discarded". The present application encapsulates the future state set, the comprehensive function, the pruning / DP parameter and the optimal trajectory index, generates a root hash using a Merkle tree, and forms a compliance certificate. The regulatory end can sample and verify the discarded trajectory to confirm that the scoring and elimination process is reasonable.

[0213] As Figure 2 , 3 , 4, 5, Figure 2 is a schematic diagram of a data acquisition and semantic label generation unit, Figure 3 is a schematic diagram of a cloud future state superposition prediction unit, Figure 4 is a schematic diagram of a decision collapse and compliance certificate generation unit, Figure 5 is a schematic diagram of an end-cloud-end consistency and degradation check unit.

[0214] The graceful degradation of the existing method is common in the IoT field, but it is usually "only part of the data is retained", it is not bound to the cold chain regulations, and it also lacks certificate support. The three-end of the present application ensures consistency based on certificate comparison. When the check fails or the network is abnormal, only the minimum compliance fields required by the regulations (temperature, location, timestamp, semantic label, confidence and certificate digest) are transmitted, and an error code is attached. After the network is restored, the backfill is automatically executed and the idempotency is guaranteed. Even if the network is interrupted, the regulatory end can still obtain the minimum compliance data, and there will be no audit gap, ensuring the compliance of "continuous chain". The specific scheme is as follows:

[0215] In this embodiment, preferably, through the data acquisition and semantic label generation unit arranged on the vehicle end, multi-source data such as temperature, humidity, position, speed and hatch state are collected. After filtering and alignment processing, semantic labels and confidence weights are automatically generated to form enhanced input vectors at each moment in the historical time period;

[0216] The future state prediction unit in the cloud uses an improved LSTM model to realize multi-trajectory future state distribution prediction, thereby improving prediction robustness and interpretability.

[0217] The decision collapse and compliance certificate generation unit on the vehicle side, the cloud side, and the supervision side scores and optimizes the multi-trajectory results, selects the optimal branch, obtains the data of each sensor and the semantic label of the cold chain vehicle cabin within the preset time step in the future, and at the same time, packages and generates a hash certificate for the optimal branch trajectory set, the comprehensive score of each branch, and the optimization logic, thereby ensuring traceability and evidence credibility throughout the process.

[0218] The end-cloud-end consistency and degradation checking unit on the vehicle side, the cloud side, and the supervision side compares the certificates between the three ends, thereby realizing data consistency checking. When the network is abnormal or the checking fails, only the minimum compliance fields (temperature, location, timestamp, semantics, confidence, and certificate digest) are transmitted, ensuring compliance data continuity, and automatically supplementing after network recovery, thereby ensuring long-term consistency.

[0219] In each data cycle, the vehicle side, the cloud side, and the supervision side all hold compliance certificates. If the compliance certificates of the three ends are completely consistent, it indicates that the data has not been tampered with and is synchronized successfully, and the three ends all show complete results (the optimal branch, the branch set digest, and the compliance certificate). If the results of the three ends are inconsistent, or the network indicators are abnormal (such as end-to-end delay > 30 seconds, packet loss rate > 5%, and certificate checking failure), the degradation mechanism is triggered.

[0220] In the degradation mode, only the minimum compliance fields of the key sensors and the semantic label of the cold chain vehicle cabin (monitoring parameters) at each time of the optimal branch, the compliance certificate, and the error reason code are transmitted, thereby ensuring that the cold chain transportation process meets the minimum requirements of supervision review. The key sensors and the semantic label of the cold chain vehicle cabin, the compliance certificate, and the error reason code are: wherein, is the compliance certificate, is the error reason code (01: hash inconsistency; 02: data missing; 03: communication timeout), and non-key fields (such as humidity and energy consumption indicators) are suspended to reduce bandwidth consumption. In this way, the supervision end can still receive the minimum compliance data that is continuously linked under extreme conditions.

[0221] When the network is restored, the system automatically triggers the supplement process:

[0222] First, the missing interval detection is carried out, and the sequence number and timestamp of each data are used to locate the missing data during the abnormal period. Then, data completion is carried out, the missing optimal branch and compliance credentials are retransmitted, and the three ends are compared again to ensure consistency. If the same data appears repeatedly due to retransmission, the system avoids repeated archiving through sequence number and hash check.

[0223] Even if the network is interrupted, the regulatory end can still see the temperature, location, timestamp and other key fields, avoiding violations due to data loss. The transparency is enhanced, and the degraded data is accompanied by error codes and credential summaries, so that the regulatory party can trace "why the data is incomplete". After the network is restored, the system will automatically complete the complete data and credentials to avoid permanent gaps. The minimum compliance field directly corresponds to the requirements of the drug administration bureau and the food cold chain standard, ensuring that it still meets the regulatory review in extreme cases.

[0224] Based on example one, this example two respectively takes vaccine cold chain transportation and a cross-provincial cold chain truck team joint distribution scene as an example to illustrate:

[0225] As shown in Figure 6 , the Figure 6 is a monitoring flowchart for vaccine cold chain transportation. The transportation vehicle is from "drug warehouse → distribution center → vaccination point", and the temperature, humidity and transportation state are strictly monitored throughout the journey. The drug administration department requires:

[0226] The temperature of the transport cabin must be maintained at 2℃-8℃;

[0227] Any node must be able to trace back the complete environmental data and decision-making process;

[0228] Even if the network is interrupted, the regulatory end can still see the temperature, location, timestamp and other key fields, avoiding violations due to data loss. The transparency is enhanced, and the degraded data is accompanied by error codes and credential summaries, so that the regulatory party can trace "why the data is incomplete". After the network is restored, the system will automatically complete the complete data and credentials to avoid permanent gaps. The minimum compliance field directly corresponds to the requirements of the drug administration bureau and the food cold chain standard, ensuring that it still meets the regulatory review in extreme cases.

[0229] The cold chain vehicle terminal is equipped with temperature and humidity sensors, GPS, hatch sensors and task scheduling interfaces. The collection frequency of temperature and humidity is set to 1 Hz, and the collection frequency of GPS positioning data is set to 0.1 Hz. Then, in the data preprocessing stage, Kalman filter algorithm is used to eliminate jitter error of temperature and humidity sensors caused by external interference, and GPS interpolation method is used to fill in the positioning data to ensure data continuity. On this basis, semantic label generation is performed, and corresponding labels are triggered according to multi-dimensional data combination: when the hatch is opened and the vehicle position is in the warehouse area, the "loading" label is generated; when the hatch is closed and the vehicle speed is greater than 5 km / h, the "in transit" label is generated; when the temperature exceeds 8℃ for 10 minutes, the "over temperature" label is generated; when the actual driving track deviates from the planned route by more than 2km, the "deviation" label is generated; finally, the confidence is calculated, and the confidence is calculated according to the data integrity, threshold exceeding degree and multi-source data consistency. For example, if the confidence is 0.92, it means that the semantic label determination result is highly reliable, and an enhanced input vector is generated.

[0230] The enhanced input vector is uploaded to the cloud to improve the LSTM model, and four different branches are generated:

[0231] For example: trajectory 1: temperature remains 5-7℃, probability 65%;

[0232] Trajectory 2: temperature rises slightly to 8.5℃, probability 20%;

[0233] Trajectory 3: temperature rises sharply to 10℃ (over temperature), probability 10%;

[0234] Trajectory 4: temperature drops to 4℃, probability 5%.

[0235] In the cloud, different trajectory branches are comprehensively scored, branches with comprehensive score lower than the pruning threshold 0.3 are removed, redundant trajectories similar to trajectory 1 (temperature stable, probability 65%) are filtered out through Fréchet distance, and the optimal branch with the lowest risk and the highest compliance is selected by dynamic programming algorithm; then, in the certificate generation link, the branch trajectory set, score function parameters, pruning parameters, DP parameters and optimal branch index are packaged, and the root hash is generated by using Merkle tree. The certificate is stored in the cloud and synchronized to the supervision end, providing support for subsequent sampling verification.

[0236] During the edge-cloud-edge consistency and degradation verification phase, the three ends simultaneously receive the optimal branch and compliance credentials, independently calculate hashes, and compare them. If they match, the complete result is displayed, including the optimal branch, a summary of the entire branch trajectory, and the compliance credentials. If inconsistencies or anomalies occur (such as network interruption lasting more than 30 seconds or packet loss exceeding 5%), the system automatically switches to degradation mode, transmitting only the minimum compliance fields. In this case, the regulatory end can still view temperature and location information, ensuring uninterrupted compliance processes. Once the network is restored, the system triggers a recovery mechanism to fill in the missing trajectory set and complete credentials. After the three ends re-compare the compliance credentials and confirm consistency, archiving is completed, effectively preventing compliance vulnerabilities.

[0237] like Figure 7 As shown, Figure 7 This is a monitoring flowchart for a joint cold chain delivery scenario. In a cross-provincial joint cold chain delivery scenario, multiple cold chain vehicles depart from different warehouses and need to transport vaccines, medicines, and fresh food to several distribution centers and hospitals within a limited time. Unlike single-vehicle transportation, the fleet dispatch center requires unified monitoring and coordinated scheduling of the operational status of multiple vehicles to avoid situations where "some vehicles arrive early and wait" or "some vehicles are delayed, causing a break in the overall delivery chain." Simultaneously, the drug regulatory authorities require the ability to conduct unified random checks of data vouchers in multi-vehicle scenarios to ensure that the environmental parameters, predicted trajectories, and compliance decision-making processes of each vehicle are traceable. Even if a vehicle experiences a communication interruption, the regulatory end must receive its minimum compliance fields in real time.

[0238] Each refrigerated truck is equipped with a temperature and humidity sensor, GPS module, door sensor, and task interface to collect real-time information on temperature, humidity, location, speed, door status, and task stage.

[0239] The temperature and humidity data are collected at a frequency of 1 Hz, while the GPS data is collected at a frequency of 0.2 Hz. During the data preprocessing stage, Kalman filtering is used to smooth the temperature and humidity data, and GPS data is supplemented by interpolation and inertial navigation fusion, while ensuring the alignment of timestamps of data from multiple vehicles. Semantic label generation follows specific rules: a "loading" label is generated when the hatch is open and the vehicle is in the warehouse; a "transporting" label is generated when the hatch is closed and the speed is > 5 km / h; a "temperature exceeding 8℃" label is generated when the temperature is > 8℃ for 10 minutes; and a "deviation from the planned path" label is generated when the deviation is more than 2 km. The system also calculates the confidence level for each data point based on data integrity, consistency, and the degree to which the threshold is exceeded, for example, a confidence level of 0.88.

[0240] In the cloud future state superposition prediction link, the enhanced input vectors of all vehicles are uploaded to the cloud, and the improved LSTM model is used for future state prediction. Unlike single vehicle prediction, this model simultaneously expands the future trajectory branches of multiple vehicles in the time superposition layer (TSL), and introduces a "coordination constraint loss" in the training stage to ensure that the prediction results of multiple vehicles maintain time consistency at key nodes such as highway exits and distribution center entrances. Taking the two-vehicle cooperative prediction as an example, trajectory 1 of vehicle A is stable at 6℃ (probability 70%), and trajectory 2 is raised to 9℃ (probability 20%). Trajectory 1 of vehicle B is stable at 7℃ (probability 65%), and trajectory 2 is delayed by 15 minutes due to deviation (probability 25%). The system also explicitly labels the vehicle ID and task stage in the branch label when outputting the trajectory, which facilitates subsequent decision-making to consider both individual vehicles and the overall situation of the vehicle fleet.

[0241] In the decision collapse and compliance certificate generation phase, the cloud first scores the multi-vehicle prediction results. In the scoring function of each vehicle trajectory, the coordination consistency factor of the vehicle at the key node is used to constrain the arrival time window of multiple vehicles at the inspection station or distribution center. Then pruning and optimization are performed. Branches with a score below the pruning threshold of 0.25 are removed, redundant trajectories with a similarity greater than 0.85 are removed by DTW distance, and a candidate set with a wider coverage is retained. Then, the dynamic programming method is used to select the overall optimal branch, which ensures compliance of individual vehicles while improving overall efficiency of the vehicle fleet. Finally, the compliance certificate is generated, which encapsulates the multi-vehicle trajectory set, scoring function parameters, pruning and DP parameters, and coordination constraint parameters, and uses Merkle tree to generate root hash. This certificate can prove the optimal branch selection of each vehicle and encapsulate the coordination constraint logic of multiple vehicles.

[0242] In terms of end-cloud-end consistency and degradation checking, the scheduling center, cloud platform, and regulatory platform receive the optimal branch and compliance certificate and compare the root hash: if the comparison is consistent, the complete results are displayed, including the optimal trajectory of each vehicle, the summary of the branch trajectory set, and the decision certificate; if the comparison is inconsistent or there is an abnormal link for a vehicle, it enters the degradation mode, only transmitting the minimum compliance field of the vehicle, and the regulatory interface will identify "compliance minimum guarantee mode"; when the network recovers, the system triggers the backfill mechanism to supplement the missing multi-vehicle trajectory set and complete certificate, and ensures long-term consistency through idempotency checking.

[0243] Embodiment three provides a cold chain logistics monitoring system, comprising:

[0244] The memory is used to store a computer program, and the processor is used to execute the computer program to realize the steps of the cold chain logistics monitoring method.

[0245] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0246] The present application is described in reference to the flowchart and / or block diagrams of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0247] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0248] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0249] Obviously, the above-described embodiments are only examples and are not intended to limit the present application. Based on the above description, those skilled in the art can make other variations and modifications of the present application without deviating from the scope of the present application. Therefore, the present application should not be limited by the above embodiments.

Claims

1. A cold chain logistics monitoring method, characterized in that, The method comprises the following steps: The data integrity index value of each time point in the historical time period is obtained by multiplying the effective data proportion of each sensor in the cold chain vehicle cabin in the historical time period by the importance weight of the sensor, and then accumulating the product results of all sensors; The threshold exceeding degree index value of each time point in the historical time period is obtained based on the preprocessed data of each sensor at each time point in the historical time period and the set value interval corresponding to the data; The multi-source consistency index value of each time point in the historical time period is obtained based on the standard deviation of the preprocessed data of the redundant sensor measuring the same physical quantity as each sensor at each time point in the historical time period, the semantic label of the cold chain vehicle cabin at each time point in the historical time period, and the change amount of the preprocessed data of each sensor at adjacent time points in the historical time period; The confidence weight of each time point in the historical time period is obtained by weighted summation of the data integrity index value, the threshold exceeding degree index value and the multi-source consistency index value of each time point in the historical time period; The enhanced input vector of each time point in the historical time period is constructed based on the confidence weight of each time point in the historical time period, the semantic label of the cold chain vehicle cabin and the preprocessed data of each sensor; The data of each sensor and the semantic label of the cold chain vehicle cabin in the future preset time step are obtained based on the enhanced input vector of each time point in the historical time period, which comprises: Will arrive The confidence weights at each time point within the time period are scaled using a normalization function to obtain... arrive The scaled confidence weights for each time point within the time period; where... This marks the beginning of a historical period. The total number of moments in the historical time period; Will arrive The scaled confidence weights at each time step within the time period are multiplied by the learnable confidence correction coefficients corresponding to the forget gate and input gate, respectively, and used as... arrive Forget gate confidence correction term and input gate confidence correction term at each time point within the time period; Based on to the enhanced input vector of each time within the period, the forget gate confidence correction term, the hidden state of the previous time, through the forget gate, obtain to the forget gate output of each time within the period; based on to the enhanced input vector of each time in the period, the input gate confidence correction term, and the hidden state of the previous time through the input gate, to obtain to the input gate output of each time in the period; will be described below. to the forget gate output, the input gate output, the candidate cell state, the cell state of the previous time point of each time point in the time period, through the cell state update, the cell state of each time point in the time period is obtained. to the cell state of each time point in the time period; Will arrive The augmented input vector, cell state, and previous hidden state at each time step within the time period are passed through the output gate to obtain... arrive The output gate outputs at each time point within the time period; based on arrive The output gate output and cell state at each time point within the time period are calculated. arrive The hidden state at each moment within the time period; where the hidden state at each moment represents the data of each sensor at that moment; In the hidden state of the time instant and its corresponding semantic label, confidence weight, construct the enhanced input vector of the time instant; wherein, , preset time steps in the future; Will The scaled confidence weights corresponding to the hidden states at time t are multiplied by the learnable confidence correction coefficients corresponding to the forget gate and input gate, respectively, and used as... Forget gate confidence correction term and input gate confidence correction term at different times; based on the enhanced input vector of the time instant, the forget gate confidence correction term, the hidden state of the time instant, by the forget gate, to obtain the forget gate output of the time instant; based on the enhanced input vector of the time moment, the input gate confidence correction term, the hidden state of the time moment passes through the input gate, to obtain the input gate output of the time moment; will be described below. the forget gate output, the input gate output, the candidate cell state, the cell state at the time t, the cell state at the time t+1, and the cell state at the time t+2. the cell state at the time t, the cell state at the time t+1, and the cell state at the time t+2. will be described below. the enhanced input vector, the cell state, the hidden state at the time step, through the output gate, to obtain the output gate output at the time step; Based on the output gate output, cell state, at the time instant, to calculate the hidden state at the time instant; the hidden state at the time instant is the data of each sensor within a preset time step in the future. The semantic label of the cold chain vehicle cabin at each time point in the future preset time step is generated based on the data of each sensor at each time point in the future preset time step.

2. The cold chain monitoring method of claim 1, wherein, The threshold exceeding degree index value of each time point in the historical time period is obtained based on the preprocessed data of each sensor at each time point in the historical time period and the set value interval corresponding to the data, which comprises: The exceeding amplitude of each sensor at each time point in the historical time period is calculated based on the preprocessed data of each sensor at each time point in the historical time period and the set value interval corresponding to the data, and the formula is: ， The exceeding amplitude of each sensor at each time point in the historical time period is normalized to obtain the normalized exceeding amplitude of each sensor at each time point in the historical time period; The threshold exceeding degree index value of each time point in the historical time period is calculated based on the importance weight of each sensor and the normalized exceeding amplitude of each sensor at each time point in the historical time period, and the formula is: ， in, Within the historical period The threshold at any given time exceeds the degree indicator value. For the first time in the historical period At the [time]th moment The exceedance of each sensor, For the first Each sensor corresponds to the lower limit of a set numerical range. For the first Each sensor corresponds to the upper limit of a set numerical range. For the first time in the historical period At the [time]th moment Data from one sensor, For the first time in the historical period At the [time]th moment Excess amplitude after normalization of each sensor For the first The importance weight of each sensor This represents a time index within a historical period. Indicates the sensor index. This is a collection of sensors for use inside the cold chain vehicle compartment.

3. The cold chain monitoring method of claim 1, wherein, The multi-source consistency index value of each time point in the historical time period is obtained based on the standard deviation of the preprocessed data of the redundant sensor measuring the same physical quantity as each sensor at each time point in the historical time period, the semantic label of the cold chain vehicle cabin at each time point in the historical time period, and the change amount of the preprocessed data of each sensor at adjacent time points in the historical time period, which comprises: The redundant sensor consistency index value of each sensor in the historical time period is obtained by normalizing the standard deviation of the preprocessed data of the redundant sensor measuring the same physical quantity as each sensor at each time point in the historical time period; The redundant sensor consistency index value of each time point in the historical time period is obtained by weighted summation of the importance weight of each sensor and the consistency index value of each sensor in the historical time period. determining whether the change amount of the pre-processed data of each sensor at each time in the historical time period is abnormal based on the semantic labels of each time and its previous time in the historical time period, if there is, setting the cross-modal consistency index value of the historical time period at this time to 0, if not, setting the cross-modal consistency index value of the historical time period at this time to 1; normalizing the absolute value of the difference of the pre-processed data of each sensor at adjacent times in the historical time period to obtain the time smoothing consistency index value of each sensor at each time in the historical time period; weighting and summing the importance weight of each sensor and the time smoothing consistency index value of each sensor in the historical time period to obtain the time smoothing consistency index value of each time in the historical time period; weighting and summing the redundant sensor consistency index value, the cross-modal consistency index value and the time smoothing consistency index value of each time in the historical time period to obtain the multi-source consistency index value of each time in the historical time period.

4. The cold chain monitoring method of claim 1, wherein, The method comprises the following steps: The enhanced input vector at the moment, the forgetting gate confidence correction term, The hidden state at the moment, and the forgetting gate output at the moment are obtained through the forgetting gate. The forgetting gate output at the moment is obtained through the forgetting gate. ， wherein, is the forget gate output at time step t, is a Sigmoid activation function, is a first weight matrix for the forget gate, is a second weight matrix for the forget gate, is a bias term for the forget gate, is the augmented input vector at time step t, is the hidden state at time step t, is the forget gate confidence correction term at time step t, is a learnable confidence correction coefficient for the forget gate, is the scaled confidence weight at time step t.

5. The cold chain monitoring method of claim 1, wherein, The method comprises the following steps: The enhanced input vector at the moment, the input gate confidence correction term, The hidden state at the moment, through the input gate, to obtain The input gate output at the moment, the formula is: ， wherein, is the input gate output at time step t, is a Sigmoid activation function, is a first weight matrix for the input gate, is a second weight matrix for the input gate, is a bias term for the input gate, is the augmented input vector at time step t, is the hidden state at time step t, is the input gate confidence correction term at time step t, is a learnable confidence correction coefficient for the input gate, is the scaled confidence weight at time step t.

6. The cold chain monitoring method of claim 1, wherein, After the hidden state of each time point in the time period is calculated to The hidden state of each time point and its corresponding semantic label, confidence weight, are used to construct The hidden state of each time point and its corresponding semantic label, confidence weight, are used to construct The hidden state of each time point and its corresponding semantic label, confidence weight, are used to construct The hidden state of each time point and its corresponding semantic label, confidence weight, are used to construct The hidden state of each time point and its corresponding semantic label, confidence weight, are used to construct The hidden state of each time point and its corresponding semantic label, confidence weight, are used to construct Will The scaled confidence weights corresponding to the hidden states at time t are multiplied by the learnable confidence correction coefficients corresponding to the forget gate and input gate, respectively, and used as... Forget gate confidence correction term and input gate confidence correction term at different times; based on the enhanced input vector of the time instant, a forget gate confidence correction term, the hidden state of the time instant, by the forget gate of the i-th parallel branch, to obtain the output of the forget gate of the i-th parallel branch at the time instant the output of the forget gate of the i-th parallel branch at the time instant based on the enhanced input vector at the time instant, the input gate confidence correction term, the hidden state at the time instant, by the input gate outputs of the T parallel branches; and the time instant the input gate outputs of the T parallel branches. will be described below. time the output of the forget gate, the output of the input gate, the candidate cell state, the cell state at the time t, through the cell state update, to obtain time the cell state of the i-th parallel branch; Will Time-dependent augmented input vector, cell state, The hidden state at any given moment, through The output gate of the parallel branch obtains time The output gate has multiple parallel branches; based on the output gate output of each parallel branch at the time instant, the cell state, the hidden state of each parallel branch at the time instant; based on obtain the probability weight corresponding to each parallel branch according to the hidden state of each parallel branch at the moment; Based on The hidden state of each parallel branch at the moment and the probability weight corresponding to the branch are screened, and the hidden state of the branch screened at the moment is taken as the data of each sensor within a preset time step in the future. The hidden state of the moment is taken as the data of each sensor within a preset time step in the future. generating the semantic label of the cold chain vehicle cabin at each time in the future preset time step based on the data of each sensor at each time in the future preset time step.

7. A cold chain logistics monitoring method according to claim 6, characterized in that, The method comprises the following steps: Screen each parallel branch according to the hidden state of each parallel branch at the moment and the probability weight corresponding to the branch. respectively the average of the product of the scaled confidence weight and the probability weight corresponding to the hidden state of each parallel branch at each time instant as the confidence score of each parallel branch; The mean of the number of sensors whose data corresponding to the hidden state of each parallel branch exceeds the threshold interval at each time instant is taken as the risk event score of each parallel branch. The mean of the number of sensors whose data corresponding to the hidden state of each parallel branch exceeds the threshold interval at each time instant is taken as the risk event score of each parallel branch. taking the difference between 1 and the risk event score of each parallel branch as the risk assessment index value of each parallel branch; obtaining the efficiency index value of each parallel branch based on the time of each parallel branch corresponding to reaching the target area and the acceptable maximum time length, the formula is: ， wherein, is the efficiency index value of the i-th parallel branch, is the time of arrival at the target region corresponding to the i-th parallel branch, is the start transportation time corresponding to the i-th parallel branch, is the maximum acceptable duration, is the parallel branch index;​​​ The ratio of the number of times that the hidden state of each parallel branch meets the regulatory requirements to the total number of regulatory requirements at each time is taken as the compliance index value of each parallel branch. The ratio of the number of times that the hidden state of each parallel branch meets the regulatory requirements to the total number of regulatory requirements at each time is taken as the compliance index value of each parallel branch. weighting and summing the confidence score, the risk assessment index value, the efficiency index value and the compliance index value of each parallel branch to obtain the comprehensive score of each parallel branch; obtaining the pruned branch set based on the comprehensive score of each parallel branch; By means of a dynamic programming algorithm, the optimal branch in the pruned branch set is obtained, and the hidden state corresponding to the optimal branch at the preset time is taken as the data of each sensor within a future preset time step. time step. generating the semantic label of the cold chain vehicle cabin at each time in the future preset time step based on the data of each sensor at each time in the future preset time step.

8. A cold chain logistics monitoring method according to claim 7, characterized in that, After obtaining the data of each sensor in the future preset time step and the semantic label of the cold chain vehicle cabin, the data of each sensor in the future preset time step and the semantic label of the cold chain vehicle cabin are matched The data of each sensor in each parallel branch and the semantic label of the cold chain vehicle cabin, the comprehensive score of each parallel branch, the confidence score of each parallel branch, the risk evaluation index value, the efficiency index value, the compliance index value, the corresponding weight, the pruning parameter, the dynamic programming algorithm parameter and the optimal branch index are taken as the original data. Within the preset time step in the future The data from each sensor in each parallel branch, along with the semantic tags of the cold chain vehicle compartment and the optimal branch index, are hashed. The generated hash value, along with the comprehensive score of each parallel branch within a preset future time step, the confidence score of each parallel branch, the risk assessment index value, the efficiency index value, the compliance index value and their corresponding weights, pruning parameters, and dynamic programming algorithm parameters, are used as leaf nodes. Through a Merkle tree structure, the Merkle root hash of the original data is generated. When the original data is accessed, the Merkle root hash of the original data is used to verify whether the original data has been tampered with.

9. A cold chain logistics monitoring system characterized in that, comprising: a memory for storing a computer program; a processor for executing the computer program to realize the steps of the cold chain logistics monitoring method in any one of claims 1 to 8.

Images