Cold chain temperature and humidity dynamic early warning method and system based on neural network prediction

Abstract

The invention discloses a cold chain temperature and humidity dynamic early warning method and system based on neural network prediction. The method comprises the following steps: collecting state data and environmental parameter data of equipment operation in real time; inputting the data into the trained multi-task prediction model, and outputting a future prediction value; evaluating the credibility of the predicted value, if the credibility is greater than a threshold value, performing prediction after correcting the multi-task prediction model, and if the credibility is less than or equal to the threshold value, performing the next step; and comparing the prediction result with a current actual threshold interval through a dynamic risk scoring algorithm to generate a dynamic risk score, and if the risk score exceeds a dynamic threshold, pushing a grading early warning to a management platform or a user. Compared with a traditional method, the method has the advantages that more time is brought forward to find potential overrun risks, sufficient time is reserved for manual intervention, data sequences are collected regularly based on sensor data collection, and meanwhile, the working conditions of equipment and external public opinion information are comprehensively analyzed to give intelligent early warning information.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of temperature and humidity monitoring, and particularly relates to a cold chain temperature and humidity dynamic early warning method and system based on neural network prediction. BACKGROUND

[0002] The existing cold chain temperature and humidity monitoring technology mainly collects temperature and humidity data in real time through Internet of Things sensors, and performs threshold value judgment based on a preset static threshold value interval (such as 2℃-8℃). In order to realize early warning, the existing technology generally adopts the strategy of narrowing the threshold value interval, for example, narrowing the actual allowed threshold value interval (2℃-8℃) to a more stringent early warning interval (such as 4℃-6℃). When the real-time monitoring data exceeds the early warning interval but does not exceed the actual threshold value, a warning signal is triggered. Narrowing the threshold value interval leads to a too narrow safety boundary, and frequent false alarms are triggered due to environmental fluctuations (such as opening the door to take things, refrigeration cycle) when the equipment is normally running, resulting in a high false alarm rate. Frequent false alarms lead to invalid inspection by the operation and maintenance personnel, and compressing the threshold value interval may force the equipment to run at a high frequency (such as frequent start and stop of the refrigeration system), increase energy consumption, and affect the service life of the equipment. The static threshold value interval cannot be dynamically adjusted according to the equipment state (such as full load / empty load) and environmental conditions (such as seasonal changes), and needs to be frequently calibrated manually. It only relies on current or historical data and cannot predict future temperature and humidity trends, resulting in delayed warning, poor adaptability and predictability. At the same time, traditional temperature and humidity data monitoring mainly uses a single type of sensor and is a post-alarm, which cannot adapt to changes in external environment or working environment. Sensors at different positions are measured in isolation and cannot be comprehensively analyzed to provide accurate cold chain operation environment safety evaluation for users. SUMMARY

[0003] The purpose of the present application is to provide a cold chain temperature and humidity dynamic early warning method and system based on Internet of Things big data analysis and large model prediction, which compresses the safety boundary to reserve buffer time for manual intervention and avoid breaking the actual threshold value when the temperature and humidity changes rapidly.

[0004] The present application is implemented by the following technical scheme: a cold chain temperature and humidity dynamic early warning method based on neural network prediction, characterized by comprising the following steps:

[0005] S1, collecting state data and environmental parameter data of equipment operation in real time;

[0006] S2, inputting the collected data into a trained multi-task prediction model for prediction, and outputting future predicted values;

[0007] S3, evaluating the credibility of the predicted values, if the credibility is greater than a threshold value, correcting the multi-task prediction model and then performing prediction, and if the credibility is less than or equal to the threshold value, performing step S4;

[0008] S4, comparing the predicted value with the current actual threshold interval through a dynamic risk scoring algorithm to generate a dynamic risk score, and if the risk score exceeds the dynamic threshold, a graded early warning is pushed to the management platform or user.

[0009] Preferably, the training of the multi-task prediction model comprises the following steps:

[0010] (1) Obtain corresponding historical data, including temperature and humidity data, equipment operating status, and environmental parameters;

[0011] (2) Preprocess the historical data, including data cleaning, feature engineering, standardized data, and conversion to supervised learning format;

[0012] (3) Divide the preprocessed data into training set and test set, use the training set data to train the LSTM model, and use the test set data to predict and evaluate the trained LSTM model.

[0013] Preferably, the feature engineering comprises:

[0014] a. Obtain the trend slope sequence of the historical data, which contains the slope, peak value, trough value and data number;

[0015] b. According to the first derivative of the temperature curve, it can be expressed as Δt(n) = T(n) - T(n-1) in the discrete temperature sequence, combined with the downtime of the equipment, to predict the next development trend of the temperature curve, where Δt(n) is the temperature change rate at time n, and T(n) is the temperature at time n;

[0016] c. The second derivative of the temperature curve is the change trend of the temperature change rate Δt ' (n) = Δt(n) - Δt(n-1) = T(n) - 2T(n-1) + T(n-2), when Δt ' (n) > 0, the temperature change rate is accelerating, when Δt ' (n) < 0, the temperature change rate is slowing down, when Δt ' (n) = 0, the temperature changes at a constant rate; Δt ' (n) and feature Δt(n) are advantageous input conditions for predicting the inflection point of the temperature curve;

[0017] d. By performing Fourier transform (DFT) on the discrete temperature data, extract its frequency domain features to understand the periodic fluctuations of the temperature, defrosting period, and frequency spectrum caused by abnormal interference;

[0018] e. Compressor vibration and current data features: extract the compressor vibration frequency features, and then add the current size data to obtain the refrigeration efficiency and compressor working condition characteristics;

[0019] f. Device correlation features: including the relationship between door opening time and temperature rise in different environments, the probability of abnormality is judged by the rate of temperature change combined with relevant feature data.

[0020] Preferably, the probability of abnormality is judged by the rate of temperature change combined with relevant feature data includes:

[0021] 1) The rate of change increases or the difference from the previous rate increases, the probability of abnormality increases;

[0022] 2) Check the access control record, if the access control state is open, the probability of abnormality increases;

[0023] 3) Find the latest compressor power history, combined with the current external temperature, if it is high, the probability of abnormality increases.

[0024] Preferably, the reliability of the predicted value is evaluated by the following formula:

[0025]

[0026] Where h is the length of the predicted data sequence, T i is the measured data, is the predicted value; when the RMSE is greater than the threshold value, the multi-task prediction model needs to be corrected.

[0027] Preferably, the formula of the dynamic risk score algorithm is:

[0028]

[0029] Where Threshold max is the upper limit of the threshold interval, Threshold min is the lower limit of the threshold interval, P t +ΔT is the predicted value, is the rate of change, α is the scene prediction proximity weight coefficient, β is the scene change rate weight coefficient, Risk_Score is the risk score value, P t is the predicted temperature value, ΔT is the prediction time span.

[0030] Preferably, the grading standard of the graded warning is:

[0031] First-level warning: Risk Score ≥70%·S;

[0032] Second-level warning: Risk_Score≥S;

[0033] Where S represents the dynamic threshold, S=x0*x1*x2*······*x n, the initial threshold value x0 of the dynamic threshold value is 1, x n represents the nth adjustment factor coefficient.

[0034] Preferably, the adjustment factors include device high load, door opening frequency, door opening duration, prediction value error, and extreme external environment.

[0035] The system for implementing the cold chain temperature and humidity dynamic early warning method based on neural network prediction comprises:

[0036] The Internet of Things data acquisition module adopts corresponding sensors deployed on devices to collect corresponding data and transmit the data to the forwarding node through a low-power protocol such as ZigBee or LoRa.

[0037] The prediction module adopts a trained multi-task prediction model to predict 30 / 60-minute temperature and humidity prediction values in the future.

[0038] The early warning decision module generates a dynamic risk score by using a dynamic trend scoring algorithm according to the prediction value and the current actual threshold interval, and pushes a graded early warning to a management platform or a user when the risk score exceeds a dynamic threshold.

[0039] The storage medium of the application stores a program, and the program is executed by a processor to implement the cold chain temperature and humidity dynamic early warning method based on neural network prediction.

[0040] The computing device of the application comprises a processor and a memory for storing a program executable by the processor, and the processor executes the program stored in the memory to implement the cold chain temperature and humidity dynamic early warning method based on neural network prediction.

[0041] Compared with the prior art, the application has the following beneficial effects:

[0042] 1. Reduce the false positive rate: replace the mechanical compression threshold with the predicted future trend to reduce the false positive rate.

[0043] 2. Early warning time: the application compresses the safety boundary to reserve a buffer time for manual intervention, avoids breaking the actual threshold when the temperature and humidity change rapidly, and compared with the traditional method, the application can find the potential over-limit risk more time in advance to reserve sufficient time for manual intervention.

[0044] 3. Dynamic adaptability: the model automatically learns the feature differences of different scenes (such as medical refrigerator door opening and transportation vehicle vibration environment), and does not need manual calibration.

[0045] 4. Resource conservation: the system in the application mainly relies on sensor data acquisition timing to collect data sequences, and gives intelligent early warning information by comprehensively analyzing the working condition of the device itself and external public opinion information, reduces invalid early warning, and reduces the operation and maintenance cost of the device (mainly the cost of employing people). BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of an embodiment of a cold chain temperature and humidity dynamic early warning method based on neural network prediction.

[0047] Figure 2 A flowchart of training of an embodiment of a multi-task prediction model.

[0048] Figure 3 A data graph collected in a certain time period for an embodiment.

[0049] Figure 4 A temperature change data graph in a certain time period for an embodiment.

[0050] Figure 5 A comparison curve graph of actual value and predicted value for an embodiment. DETAILED DESCRIPTION

[0051] In order for those skilled in the art to better understand the present application, the following will further describe the present application in conjunction with the drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0052] The present application predicts the future temperature and humidity change trend by analyzing big data of Internet of Things combined with deep learning model without changing the actual threshold interval, and dynamically generates early warning signal.

[0053] As shown in Figure 1 The cold chain temperature and humidity dynamic early warning method based on neural network prediction of the present embodiment comprises the following steps:

[0054] S1, real-time collection of state data and environmental parameter data of equipment operation.

[0055] The obtained data includes, but is not limited to, temperature and humidity data of the monitored subject (if humidity needs to be collected, humidity collection is also required), temperature data of the external environment, compressor power (which can be indirectly judged by current), access control record (if it is a cold storage, refrigerator, etc. with installed door magnetic sensor), and weather information, power department early warning information and notification information, etc.

[0056] S2. Input the collected data into the trained multi-task prediction model and input the corresponding prediction value.

[0057] The training of the multi-task prediction model in this embodiment includes the following steps:

[0058] (1) Obtain relevant historical data. The historical data in this embodiment includes fused historical temperature and humidity data, equipment operating status (such as compressor frequency, access control records), environmental parameters (such as temperature and humidity data), and cold chain equipment information collection (equipment usage data).

[0059] (Information such as address, manufacturer, year of manufacture, and compressor power used).

[0060] (2) Preprocessing the historical data, including standardizing the data and converting it to a supervised learning format. The preprocessing in this embodiment includes:

[0061] Data cleaning: Remove outliers (such as transient sensor failure data, brief network disconnections, etc.).

[0062] Feature engineering:

[0063] a. For historical data collected by the device, calculate the trend slope of the data. The trend slope sequence includes the slope, peak value, trough value and number of data entries; (0.22, 2.1, 6.5, 20) indicates that the temperature rises from 2.1℃ to 6.5℃ in 20 sampling periods, with an average rise slope of 0.22℃, where 2.1℃ and 6.5℃ are the inflection points of the temperature curve.

[0064] like Figure 3 As shown, this embodiment calculates the trend slope of the data (divided into upward trend and downward trend) over a certain period of time (usually the period from the time of the first data collection to the time of the peak / trough data collection).

[0065] b. Based on the first derivative of the temperature curve In a discrete temperature sequence, it can be represented as: Δt(n) = T(n) - T(n-1). Combined with the compressor's downtime, the next trend of the temperature curve can be predicted; where Δt(n) is the rate of temperature change at time n, and T(n) is the temperature at time n.

[0066] c. Based on the second derivative of the temperature curve The trend of the rate of temperature change Δt in a discrete temperature sequence ' (n) = Δt(n) - Δt(n-1) = T(n) - 2T(n-1) + T(n-2), which can be interpreted as the second derivative of the temperature curve. When Δt ' When (n) > 0, the rate of temperature change is accelerating (e.g., accelerating heating or decelerating cooling). When Δt '(n) < 0, the temperature rate of change is slowing down (e.g. decelerating warming or accelerating cooling). When Δt ' (n) = 0, the temperature changes at a constant rate (linearly). The characteristic Δt ' (n) together with the characteristic Δt(n) can predict the inflection point of the temperature curve (e.g. the temperature rebound trend after the refrigeration system stops).

[0067] d. By performing a Fourier transform (DFT) on the discrete temperature data, the frequency domain features are extracted to understand the periodic fluctuations of the temperature, the defrosting period, and the frequency spectrum caused by abnormal interference.

[0068] e. Compressor vibration and current data features: Extract the compressor vibration frequency features, and then add the current size data to comprehensively obtain the refrigeration efficiency and compressor working condition characteristics. For example, the compressor works for 30 minutes and stops (determined according to the previous historical data), and according to the current temperature and the upward trend rate of change, the highest point of the upward trend is obtained.

[0069] a. Monitor the external environment temperature meter data sequence outside the monitoring area, and the sampling frequency is consistent with that of the internal temperature sequence of the monitoring area;f. Equipment association features, the relationship between the door opening time and the temperature rise in different environments. Figure 4 For a customer, the refrigerator door is not closed well (virtually closed) for about half an hour (the customer sets the refrigerator temperature to 4.5℃, and the safe interval is 2℃-8℃, and the normal temperature fluctuation is 3℃-5.5℃). Generally, by increasing or maintaining the previous rate of temperature change, combined with relevant feature data, the probability of abnormality is determined:

[0070] The change rate is accelerated or the difference from the previous rate is increased, and the probability of abnormality is increased;

[0071] Check the access control record, if the access control state is open, the probability of abnormality is increased;

[0072] Find the latest compressor power history, combined with the current external temperature, if it is high, the probability of abnormality is increased.

[0073] (3) Use the preprocessed historical data to train a multi-task prediction model.

[0074] The multi-task prediction model of the embodiment is trained using a long short-term memory network (LSTM) based on a recurrent neural network (RNN), but is not limited to only using this kind of neural network learning and prediction technology. Through the fusion of historical temperature and humidity data, equipment running state (such as compressor frequency, access control record), multi-modal joint prediction is performed.

[0075] According to different customer equipment, the corresponding training learning is customized, that is, each device of each customer corresponds to a training. The training process is as follows:

[0076] 1) According to the requirements of data acquisition, a multi-dimensional time sequence, that is, a multi-element time sequence, is formed. These sequences are used as the input of the multi-layer LSTM to enable the model to learn;

[0077] 2) The preprocessed data is divided into a training set (80%) and a test set (20%). Since the refrigeration mode (the compressor stops working when the set temperature is reached) of refrigerators, refrigerators, cold chain vehicles, etc. is a regular periodic fluctuation, if there is no abnormal situation, the temperature in the next 30-60 minutes can be predicted using the data in the past period. The sampling interval can be uniformly set according to the response characteristics of the refrigeration equipment. During the prediction analysis, according to the actual test, 1-5 minutes is generally used as the sampling interval of the temperature data. Other types of sensors need to be set according to the characteristics of the collected signals and the requirements of the Shannon sampling theorem;

[0078] 3) The training set is used to train the LSTM model.

[0079] 4) The test set is used to predict and evaluate the trained LSTM model.

[0080] The embodiment adopts the preprocessed historical data to train a multi-task prediction model, and outputs the future 30 / 60 minute temperature and humidity prediction values (different future time prediction, only need to set different predict_steps values). The trained model is to predict the future data, which is to make a prediction under normal conditions, and does not add additional factors such as access control and compressor power for calculation. The model compares the predicted data with the dynamic threshold (calculated according to the dynamic threshold rule) to determine whether to issue a warning.

[0081] The method of the application can train different models for medical refrigerators, refrigerators, cold chain transport vehicles, etc.

[0082] S3, since the monitoring process of the cold chain equipment is a long-time data acquisition, the credibility of the prediction value needs to be evaluated. If the credibility is greater than the threshold, the multi-task prediction model needs to be corrected and then predicted. If the credibility is less than or equal to the threshold, the next step is performed.

[0083] In the embodiment, the root mean square difference is used as the judgment of the prediction result In the implementation process, when the RMSE is greater than 0.5℃, the credibility of the prediction result cannot be used for early warning, otherwise a large number of false positives will be caused. At this time, the model needs to be modified (fine-tune the parameters) and restarted for model training, for example: according to the data collection activity area (South China), the summer time is long, fine-tune the time series length, and the outdoor temperature, humidity (rainy), adjust the input parameters, and restart the model training. Wherein h is the prediction data sequence length, T i is the measured data, is the predicted data.

[0084] S4, according to the closeness of the prediction result to the current actual threshold interval (such as 2℃-8℃), a dynamic risk score is generated to avoid false positives caused by fixed compression threshold, and if the risk score exceeds the dynamic threshold, an early warning signal is generated.

[0085] The embodiment generates a dynamic risk score through a dynamic risk score algorithm:

[0086]

[0087] Wherein, Threshold max is the upper limit of the threshold interval, Threshold min is the lower limit of the threshold interval, Pt+ΔT is the predicted value, is the change rate, α is the scene prediction closeness weight coefficient, β is the scene change rate weight coefficient, Risk_Score is the risk value, P t is the predicted temperature value, and ΔT is the prediction time span (for example, 30 minutes).

[0088] When the risk score exceeds the dynamic threshold, a graded early warning is pushed to the management platform or user (short message, email, etc.). The grading standard of the embodiment is as follows:

[0089] First: Risk Score ≥70%·S;

[0090] Second: Risk_Score≥S;

[0091] Wherein S represents the dynamic threshold, S=x0*x1*x2*······*x n , the initial threshold x0 of the dynamic threshold is 1. The embodiment lists six adjustment factors, and the adjustment rules of the dynamic threshold are shown in Table 1.

[0092] Table 1 Adjustment rules of dynamic threshold

[0093]

[0094] S5, the above collected related data, prediction value and risk value are stored and form a report.

[0095] The system for implementing the cold chain temperature and humidity dynamic early warning method based on neural network prediction comprises:

[0096] The Internet of Things data acquisition module deploys sensors on corresponding devices to collect corresponding data and transmit the data to the forwarding node (gateway) through ZigBee, MQTT, HTTPS or LoRa low-power protocols.

[0097] The prediction module uses the trained multi-task prediction model to predict the temperature and humidity prediction values in the next 30 / 60 minutes.

[0098] The early warning decision module generates a dynamic risk score using the dynamic trend scoring algorithm according to the prediction results and the current actual threshold interval, and pushes a graded early warning to the management platform or the user when the risk score exceeds the dynamic threshold.

[0099] As a specific embodiment, the temperature and humidity sensor is deployed in a medical refrigerator to collect temperature data in the refrigerator, and the collected temperature and humidity data are input into the trained deep learning model to predict the temperature in the refrigerator and issue an early warning.

[0100] Through analysis of the feature data, the temperature abnormality reason is synchronously given: for example, when the temperature fluctuation period is normal, the compressor working condition and the external environment temperature are also normal, the temperature early warning gives that the refrigerator load is too low, and the temperature is expected to exceed the range after 20 minutes; when the current temperature data change characteristics are consistent with the case where the door is not closed, it is given that the door is not closed, and the temperature will exceed the required range after 15 minutes.

[0101] Actual application example

[0102] Scenario: Through long-term extraction of feature data of the refrigerator by the system, the feature data is input into the LSTM, and the subsequent 12 data are output from the LSTM. Through learning of the data for half a year, the parameters of the LSTM are determined. As the prediction model of the refrigerator. If the current temperature is 3.42℃ in the subsequent actual detection, the temperature in the next 60 minutes is predicted and it is judged whether to issue an early warning, the data is collected every 5 minutes, the monitoring interval is 2℃-8℃, and the dynamic threshold (taken in the last one hour) is 1.

[0103] The present embodiment obtains 12 temperature data after the current temperature data through the learned and trained prediction model, and if there is temperature abnormal data in the 12 temperature records, the system will issue an early warning. At the same time, the prediction model matches the existing data characteristics with the features extracted in the feature engineering, and further gives the early warning reason. Figure 5 As shown in the comparison curve of the actual value and the predicted value (dotted line).

[0104] The predicted temperature and humidity values are subjected to risk assessment through a dynamic wind row scoring algorithm to avoid misjudgment caused by a fixed compression threshold, and when the risk score exceeds the dynamic threshold, a graded early warning is pushed to a management platform or a user.

[0105] The preferred embodiments of the application disclosed above are only used to help explain the application. The preferred embodiments do not describe all the details and limit the application to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of the specification. The specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the application, so that those skilled in the art can well understand and utilize the application. The application is limited only by the claims and their full scope and equivalents.

Claims

1. A method for dynamic early warning of temperature and humidity in the cold chain based on neural network prediction, characterized in that, Includes the following steps: S1. Real-time acquisition of equipment operating status data and environmental parameter data; S2. Input the collected data into the trained multi-task prediction model for prediction, and output the future predicted value; S3. Evaluate the reliability of the predicted value. If the reliability is greater than the threshold, the multi-task prediction model is corrected and prediction is performed again. If the reliability is less than or equal to the threshold, proceed to step S4. S4. The predicted value is compared with the current actual threshold range through a dynamic risk scoring algorithm to generate a dynamic risk score. If the risk score exceeds the dynamic threshold, a graded warning will be pushed to the management platform or the user.

2. The method for dynamic early warning of cold chain temperature and humidity based on neural network prediction according to claim 1, characterized in that, Training a multi-task prediction model includes the following steps: (1) Obtain relevant historical data, including temperature and humidity data, equipment operating status, and environmental parameters; (2) Preprocess historical data, including data cleaning, feature engineering, data standardization, and conversion into supervised learning format; (3) Divide the preprocessed data into training set and test set. Use the data in the training set to train the LSTM model, and use the data in the test set to predict and evaluate the trained LSTM model.

3. The method for dynamic early warning of cold chain temperature and humidity based on neural network prediction according to claim 2, characterized in that, Feature engineering includes: a. Obtain the trend slope sequence of historical data, wherein the trend slope sequence includes the slope, peak value, trough value and number of data entries; b. Based on the first derivative of the temperature curve, it can be expressed as Δt(n) = T(n) - T(n-1) in the discrete temperature sequence. Combined with the downtime of the equipment, the next trend of the temperature curve is predicted, where Δt(n) is the rate of temperature change at time n, and T(n) is the temperature at time n. c. The second derivative of the temperature curve, and the trend of the rate of temperature change Δt in the discrete temperature sequence. ' (n)=Δt(n)-Δt(n-1)=T(n)-2T(n-1)+T(n-2), when Δt ' When (n) > 0, the rate of temperature change is accelerating, and when Δt ' When (n) < 0, the rate of temperature change is slowing down, and when Δt ' When (n) = 0, the temperature changes at a constant rate; Δt ' (n) and feature Δt(n) are favorable input conditions for predicting the inflection point of the temperature curve; d. By performing Fourier transform (DFT) on discrete temperature data, its frequency domain features are extracted to understand the periodic fluctuations of temperature and the defrosting cycle, as well as the spectrum caused by abnormal interference; e. Compressor vibration and current data characteristics: Extract the compressor vibration frequency characteristics, and then add current magnitude data to obtain the refrigeration efficiency and compressor operating characteristics; f. Equipment-related characteristics: including the relationship between door opening time and temperature rise under different environments. By combining the rate of temperature change with relevant characteristic data, the probability of anomalies can be determined.

4. The method for dynamic early warning of cold chain temperature and humidity based on neural network prediction according to claim 3, characterized in that, By analyzing the rate of temperature change and combining it with relevant characteristic data, the probability of an anomaly can be determined, including: 1) If the rate of change accelerates or the difference from the previous rate increases, the probability of an anomaly increases; 2) Check the access control records. If the access control status is "open," the probability of an anomaly increases. 3) Check the recent compressor power history and combine it with the current external temperature. If it is too high, the probability of an anomaly will increase.

5. The method for dynamic early warning of cold chain temperature and humidity based on neural network prediction according to claim 1, characterized in that, The reliability of the predicted values ​​is evaluated using the following formula: Where h is the length of the predicted data sequence, T i These are actual measured data. This is a predicted value; When the RMSE is greater than the threshold, a multi-task prediction model is needed for correction.

6. The method for dynamic early warning of cold chain temperature and humidity based on neural network prediction according to claim 1, characterized in that, The formula for the dynamic risk scoring algorithm is: Among them, Threshold max Threshold is the upper limit of the threshold interval. min P is the lower limit of the threshold interval. t +ΔT is the predicted value. Let α be the rate of change, β be the weighting coefficient for the scene prediction proximity, and Risk_Score be the risk score. t The predicted temperature value is given by ΔT, which represents the prediction time span.

7. The method for dynamic early warning of cold chain temperature and humidity based on neural network prediction according to claim 6, characterized in that, The grading criteria for tiered early warning are as follows: Level 1 Warning: Risk Score ≥70%·s; Level 2 warning: Risk_Score ≥ S; Where S represents the dynamic threshold, S = x0*x1*x2*······*x n The initial threshold x0 of the dynamic threshold is 1, x n This represents the coefficient of the nth adjustment factor.

8. The method for dynamic early warning of cold chain temperature and humidity based on neural network prediction according to claim 7, characterized in that, Adjustment factors include high equipment load, frequency of door opening, duration of door opening, prediction error, and extreme external environment.

9. A system for implementing the dynamic early warning method for cold chain temperature and humidity based on neural network prediction as described in claim 1, characterized in that, include: The IoT data acquisition module uses sensors deployed on corresponding devices to collect relevant data and transmits it to the forwarding node via low-power protocols such as ZigBee or LoRa. The prediction module uses a trained multi-task prediction model to predict temperature and humidity values ​​for the next 30 / 60 minutes. The early warning decision module generates a dynamic risk score based on the predicted value and the current actual threshold range using a dynamic risk scoring algorithm. When the risk score exceeds the dynamic threshold, a tiered early warning is pushed to the management platform or the user.

10. A computer device, characterized in that, include: A memory and a processor, and a computer program stored in the memory, which, when executed on the processor, implements the method as described in any one of claims 1 to 8.

Images