Cold chain transportation temperature control monitoring method and system based on artificial intelligence

Abstract

The invention discloses a cold chain transportation temperature control monitoring method and system based on artificial intelligence. The method comprises the following steps: firstly, acquiring attribute data of a transportation subject and state data of a temperature control parameter; extracting preliminary coding features by using an encoder pre-trained for different data classes; scene association fusion is carried out on the features through processing units of all monitoring scenes, and scene special information is highlighted; mining a time sequence conduction relation between the scenes by combining a time sequence correlation model; and finally, fusing the scene correlation characteristics and the time sequence correlation characteristics, and outputting a temperature control abnormity monitoring result of each scene. According to the method, through multi-source data fusion, scene targeted analysis and time sequence conduction mining, accurate identification and prospective early warning of cold chain temperature control abnormity are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based cold chain transportation temperature control monitoring method and system. Background Art

[0002] Cold chain transportation is crucial for ensuring the quality of perishable goods like fresh produce and pharmaceuticals. The stability of its temperature control system directly impacts cargo safety. Traditional cold chain temperature monitoring often relies on threshold alarms based on single sensors. This has the following shortcomings: It focuses solely on the real-time values ​​of single parameters, such as temperature and current, and fails to fully integrate transport entity attributes with multi-source status data, resulting in one-sided feature extraction. Furthermore, the monitoring scenarios are isolated, and the temporal transmission relationships between abnormal events are not explored, making it difficult to achieve forward-looking warnings. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for monitoring temperature control in cold chain transportation based on artificial intelligence.

[0004] In a first aspect, an embodiment of the present invention provides an artificial intelligence-based cold chain transportation temperature control monitoring method, comprising:

[0005] Obtain the subject attribute data corresponding to the transport subject and the parameter status data corresponding to the temperature control parameter;

[0006] Performing preliminary feature extraction on the subject attribute data and the parameter state data according to a plurality of pre-trained encoders to obtain preliminary coding features output by each of the pre-trained encoders; each of the pre-trained encoders is pre-trained for different data types;

[0007] Performing scene association fusion on the preliminary coding features output by each of the pre-trained encoders according to the scene processing units corresponding to the multiple monitoring scenes to obtain scene association features output by each of the scene processing units;

[0008] Performing time series correlation feature mining on the preliminary coding features output by each pre-trained encoder according to a time series correlation model to obtain time series correlation features output by the time series correlation model; the monitoring scenario is used to characterize temperature control abnormal events occurring in the temperature control parameters of the transport entity; and a time series conduction relationship exists between the plurality of monitoring scenarios;

[0009] A full-scene prediction calculation is performed based on the scene association features and the time series association features to obtain the temperature control anomaly monitoring results corresponding to each monitoring scene.

[0010] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is configured to execute the method described in the first aspect.

[0011] Compared to existing technologies, the present invention offers the following advantages: The disclosed AI-based cold chain transportation temperature control monitoring method and system utilizes the following: acquiring the transport entity's attribute data and temperature control parameter status data; extracting preliminary encoding features using encoders pre-trained for different data types; integrating features through scene correlation and fusion using processing units for each monitoring scenario, highlighting scene-specific information; mining the temporal transmission relationships between scenarios using a temporal correlation model; and finally fusing scene correlation features with temporal correlation features to output temperature control anomaly monitoring results for each scenario. This method achieves accurate identification and proactive early warning of cold chain temperature control anomalies through multi-source data fusion, scene-specific analysis, and temporal transmission mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0013] Figure 1 A schematic flow chart of the steps of the artificial intelligence-based cold chain transportation temperature control monitoring method provided in an embodiment of the present invention;

[0014] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0016] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0017] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of the artificial intelligence-based cold chain transportation temperature control monitoring method provided in an embodiment of the present disclosure. The artificial intelligence-based cold chain transportation temperature control monitoring method is introduced in detail below.

[0018] Step S201, obtaining the subject attribute data corresponding to the transport subject and the parameter status data corresponding to the temperature control parameter;

[0019] Step S202: performing preliminary feature extraction on the subject attribute data and the parameter state data according to a plurality of pre-trained encoders to obtain preliminary coding features output by each of the pre-trained encoders; each of the pre-trained encoders is pre-trained for different data types;

[0020] Step S203, performing scene association fusion on the preliminary coding features output by each of the pre-trained encoders according to the scene processing units corresponding to the multiple monitoring scenes, to obtain scene association features output by each of the scene processing units;

[0021] Step S204: performing time series correlation feature mining on the preliminary coding features output by each of the pre-trained encoders according to a time series correlation model to obtain time series correlation features output by the time series correlation model; the monitoring scenario is used to characterize temperature control abnormal events occurring in the temperature control parameters of the transport entity; and a time series conduction relationship exists between the plurality of monitoring scenarios;

[0022] Step S205 , performing full-scenario prediction calculation based on the scenario-related features and the time series-related features to obtain a temperature control anomaly monitoring result corresponding to each monitoring scenario.

[0023] In the embodiment of the present invention, for example, this embodiment takes the server of a fresh cold chain logistics enterprise as the execution entity, and combines its actual cold chain transportation scenario (such as a refrigerated truck transporting fresh strawberries) to describe in detail the specific implementation process of the cold chain transportation temperature control monitoring method based on artificial intelligence.

[0024] As the data processing core, the server first needs to collect two types of key data from multiple heterogeneous data sources: the transport entity's attribute data and the status data of temperature control parameters. Taking a refrigerated truck (transport entity) transporting fresh strawberries as an example: Attribute data: The server obtains the refrigerated truck's inherent attribute information through the enterprise resource planning (ERP) system, including vehicle type (e.g., medium-sized van refrigerated truck), manufacturer (a well-known cold chain equipment brand), year of manufacture (2020), rated load (2 tons), refrigeration unit model (e.g., Thermo King T-800), and historical maintenance records (whether the compressor has been replaced in the past 12 months, whether the temperature sensor has been calibrated). This data reflects the transport entity's basic capabilities and reliability. For example, the refrigeration unit model directly correlates to its cooling efficiency, while historical maintenance records indirectly reflect the equipment's current operating status. Parameter status data: The server collects dynamic temperature control data in real time through onboard Internet of Things (IoT) devices. This includes: temperature sensor data: six temperature sensors installed at different locations in the vehicle compartment (e.g., front, middle, and rear), collecting temperature values ​​every 30 seconds (e.g., the current temperature at each measurement point is 2.3°C, 2.5°C, and 2.4°C); humidity sensor data: the real-time humidity value in the vehicle compartment (e.g., 85%); refrigeration equipment operating data: compressor current (e.g., 15A), voltage (220V), operating mode (e.g., "forced cooling" or "energy saving"), and cumulative operating time (e.g., the current run time for this trip is 5 hours); environmental interference data: the number of door openings (e.g., the door has been opened 3 times during this trip), the duration of each door opening (e.g., the first door opening lasted 2 minutes, the second door opening lasted 3 minutes), and the ambient temperature (e.g., the current temperature in the area being passed is 30°C). The server receives real-time data streams from IoT devices via the MQTT protocol and retrieves static attribute data, such as historical maintenance records, from the enterprise database. These two types of data are ultimately integrated into a structured input dataset for subsequent feature extraction.

[0025] To extract effective features from heterogeneous data, the server uses multiple encoders pre-trained for different data types to perform preliminary feature extraction on the integrated main attribute data and parameter state data. First, the server uses a feature mapping module to unify the format and transcode the raw data. For main attribute data (such as "refrigeration unit model" and "historical maintenance record"), the feature mapping module converts them into numerical features. For example, "Thermo King T-800" is mapped to a predefined model code (such as 001), and "Compressor replaced within the last 12 months" is mapped to a binary feature (1 for yes, 0 for no). For textual maintenance notes (such as "Expansion valve replaced on May 10, 2023"), a word embedding (Word2Vec) model is used to convert them into low-dimensional vectors. For parameter state data (such as temperature time series and compressor current), the feature mapping module organizes the time series data into time windows (such as the temperature value of the last hour, a total of 120 time points) and maps discrete state data (such as "strong cooling mode") into one-hot encoding. The server then uses a multi-source integration module to fuse the subject attribute features (such as vehicle reliability) with the parameter state features (such as temperature fluctuations) to avoid information loss in a single data dimension. For example, the feature vector corresponding to the vehicle's "refrigeration unit model" (reflecting theoretical cooling capacity) and the feature vector corresponding to the "average temperature over the past hour" (reflecting actual cooling performance) are concatenated into a high-dimensional vector (e.g., 256 dimensions), forming a multi-source integrated feature. Next, the server inputs the multi-source integrated feature into multiple pre-trained encoders. These encoders have been pre-trained for different data types (such as structured attribute data and time-series sensor data) to learn common feature representations. The structured attribute encoder uses a multi-layer perceptron (MLP) architecture. During the pre-training phase, it is trained on a large amount of historical vehicle attribute data (e.g., the models and maintenance records of 100,000 refrigerated trucks) to learn implicit features such as "vehicle reliability" and "refrigeration capacity." For example, given the input data for a vehicle's "manufacturer," "year of manufacture," and "maintenance records," the output is a 64-dimensional feature vector, one dimension of which may represent the probability of reduced cooling efficiency due to equipment aging. Time Series Sensor Encoder: This encoder uses a long short-term memory (LSTM) or time series convolutional network (TCN) architecture. During the pre-training phase, it uses millions of time series data points, such as temperature and current, to learn features such as "temperature fluctuation periodicity" and "abnormal compressor current patterns." For example, it inputs a temperature time series (120 points) from the last hour and outputs a 64-dimensional feature vector, one dimension of which may represent the "intensity of the temperature exceeding the standard trend." Ultimately, each pre-trained encoder outputs preliminary encoding features for its data class (e.g., a 64-dimensional vector for the structured attribute encoder and a 64-dimensional vector for the time series sensor encoder). These features retain the core information of the original data and lay the foundation for subsequent scene fusion and time series analysis.

[0026] Abnormal temperature control events (monitoring scenarios) that may occur in cold chain transportation are diverse and targeted, such as "continuous high temperature causing strawberries to rot", "refrigeration equipment failure causing temperature rise", "frequent opening and closing of doors causing temperature fluctuations", etc. These scenarios need to be analyzed in combination with different data features, so the server configures a scene processing unit for each monitoring scenario to correlate and fuse the preliminary encoding features output by the pre-trained encoder. Taking the "refrigeration equipment failure" scenario (the first monitoring scenario) and the "temperature continues to exceed the upper limit" scenario (the second monitoring scenario) as examples, there is a temporal conduction relationship between the two scenarios: equipment failure (the first scenario) may cause the temperature to continue to exceed the standard (the second scenario). The fusion process of the scene processing unit is as follows: First, the server dynamically allocates the weights (first routing weights) of each pre-trained encoder according to the scenario requirements through the first dynamic routing component corresponding to each monitoring scenario. For example, the "refrigeration equipment failure" scenario requires focusing on analyzing the compressor operation data (time series sensor data) and the equipment's historical maintenance records (structured attribute data). Therefore, the dynamic routing component will assign a higher weight (such as 0.7) to the time series sensor encoder and a lower weight (such as 0.3) to the structured attribute encoder. The "temperature continues to exceed the upper limit" scenario requires comprehensive temperature time series trends (time series sensor data) and vehicle cooling capacity (structured attribute data). Therefore, the dynamic routing component may assign balanced weights to the two (such as 0.5 and 0.5). Subsequently, the server performs weighted fusion on the preliminary encoding features of each pre-trained encoder based on the first routing weight to obtain the first scene fusion feature for each scene. For example, the fusion feature of the "refrigeration equipment failure" scenario is: [0.7×{time series sensor encoding feature}+0.3×{structured attribute encoding feature}] Finally, the scene processing unit performs deep correlation fusion on the first scene fusion feature. For example, for the "refrigeration equipment failure" scenario, the processing unit may include an attention mechanism sub-module to further highlight key features such as "abnormal fluctuations in compressor current" and "historical compressor replacement frequency", and output the scene-related features of the scenario (such as a 128-dimensional vector), which concentrates on reflecting information directly related to the equipment failure.

[0027] Because there are temporal transmission relationships between monitoring scenarios (e.g., equipment failure, temperature exceeding the standard, and cargo spoilage), the server needs to capture the temporal dependencies of features using a temporal association model. The temporal association model operates as follows: First, the server uses the second dynamic routing component corresponding to each monitoring scenario to determine the weights (second routing weights) for each pre-trained encoder in the temporal analysis. Unlike the scenario processing unit, the temporal association model focuses more on the temporal continuity of features, and therefore may bias weight allocation towards the temporal sensor encoder. For example, in the "equipment failure-temperature exceeding standard" transmission analysis, the second routing weight might be assigned to the temporal sensor encoder and 0.2 to highlight the temporal dependencies of time series data such as temperature and current. The server then performs a weighted fusion of the preliminary encoded features based on the second routing weights, generating a second-scenario fused feature for each scenario (e.g., a 128-dimensional vector). Next, the temporal association model performs temporal association mining on the second-scenario fused feature. In this implementation, the temporal correlation model includes a temporal convolutional correlation module and a temporal nonlinear transformation module: The temporal convolutional correlation module uses a 1D convolutional layer with a sliding window size of 4 (i.e., analyzing data from the last four time steps). The convolution kernel calculates the correlation between features in adjacent time steps. For example, given the second scenario fusion feature sequence (time steps t-3, t-2, t-1, t) for the "equipment failure" scenario, the convolution operation extracts local temporal patterns such as "abnormal current increase from time t-1 to t" and "slight temperature increase from time t-2 to t-1," and outputs a temporal correlation weight for each time step (e.g., reflecting the importance of a time step to subsequent scenarios). The temporal nonlinear transformation module uses a gated recurrent unit (GRU) or an attention mechanism to aggregate the temporal correlation weights. For example, the attention mechanism assigns higher weights to critical time steps (e.g., the initial time t0 of the equipment failure). Ultimately, a global temporal correlation feature (e.g., a 128-dimensional vector) is output, which integrates temporal transmission information between scenarios (e.g., "the temperature will continue to rise within 2 hours after the equipment failure occurs"). Taking "equipment failure - temperature exceeding the standard" as an example, the timing association model may output features indicating that: when the device has a current abnormality (a precursor to failure) at time t0, the temperature will begin to rise at t0+1 hour (t1), and the temperature may exceed the threshold (the strawberry preservation range of -2°C to 5°C) at t0+2 hours (t2), thereby establishing a timing transmission relationship between scenarios.

[0028] The server ultimately needs to combine scenario-related features (reflecting the key factors of the current scenario) and time-series-related features (reflecting the temporal transmission between scenarios) to output the temperature control anomaly probability (monitoring result) for each monitoring scenario. The specific process is as follows: First, the server dynamically assigns weights (third routing weights) to the scenario-related and time-series-related features through the third dynamic routing component corresponding to each monitoring scenario. For example, for the "equipment failure" scenario (early warning scenario), the real-time features of the current device status (scenario-related features) are more dependent, so the scenario-related feature weight is set to 0.7 and the time-series-related feature weight is set to 0.3. For the "temperature continuously exceeding the upper limit" scenario (subsequent derivative scenario), the cumulative effect over time (such as the trend of temperature continuing to rise after a failure occurs) is more dependent, so the time-series-related feature weight is set to 0.7 and the scenario-related feature weight is set to 0.3. The server then performs a weighted fusion of the two types of features based on the third routing weights to obtain a third-scenario fused feature (e.g., a 256-dimensional vector) for each scenario. Finally, the server performs dimension mapping on the third-scenario fused feature through a fully connected module to output the temperature control anomaly probability for each monitoring scenario. For example, when the third fused feature of the "equipment failure" scenario is input, the fully connected layer (output dimension 1) uses a sigmoid activation function to output an abnormality probability (e.g., 0.85, indicating an 85% probability of equipment failure). Similarly, when the third fused feature of the "temperature continuously exceeds the upper limit" scenario is input, the output is an abnormality probability (e.g., 0.92, indicating a 92% probability of the temperature continuously exceeding the upper limit within the next two hours). The server ultimately compares these probability values ​​with a preset threshold (e.g., 0.7). If the threshold is exceeded, an alert is triggered, notifying monitoring personnel to take measures (e.g., inspecting refrigeration equipment or adjusting transportation routes).

[0029] To ensure the accuracy of the above method, the server needs to train the model based on historical data. The training data includes: a historical cold chain temperature control training set: 100,000 transportation records from the past three years, each record containing sample subject attribute data (such as vehicle model, maintenance history) and sample parameter status data (such as temperature time series, compressor current); sample temperature control anomaly annotation: manually annotated abnormal events (such as "On May 10, 2022, a refrigerated truck exceeded the temperature limit due to a compressor failure"). During training, the server couples the initial pre-trained encoder, scene processing unit, and time series correlation model into an initial full-scenario temperature control perception model and optimizes model parameters through multi-task learning: The model's forward propagation takes historical data as input and sequentially calculates preliminary encoding features, scene-related features, and time series-related features. Multi-branch prediction outputs anomaly probabilities based on the scene-related features (first prediction), time series-related features (second prediction), and fusion features (third prediction). Deviation calculation calculates the cross-entropy loss between the first prediction and the annotation (first anomaly deviation value), the loss of the second prediction (second anomaly deviation value), and the loss of the third prediction (third anomaly deviation value), with the total loss being a weighted sum of the three (e.g., 0.3:0.3:0.4). Backward propagation adjusts model parameters using the Adam optimizer until the total loss is less than a preset tolerance (e.g., 0.05), resulting in the target full-scenario temperature control perception model. Through this training, the model simultaneously optimizes scene-related analysis, time series conduction analysis, and fully fused prediction capabilities, ensuring accurate identification of temperature control anomalies in cold chain transportation in real-world applications.

[0030] In summary, this method achieves accurate monitoring of temperature control anomalies in cold chain transportation through multi-source data fusion, scenario-based feature association, time series transmission analysis and full-scenario prediction, providing core technical support for the intelligent management of cold chain logistics.

[0031] In an embodiment of the present invention, the preliminary feature extraction of the subject attribute data and the parameter state data based on multiple pre-trained encoders to obtain preliminary coding features output by each pre-trained encoder can be implemented through the following examples.

[0032] Performing feature transcoding operations on the subject attribute data and the parameter state data respectively according to the feature mapping module to obtain subject attribute features corresponding to the subject attribute data and parameter state features corresponding to the parameter state data;

[0033] Performing multi-source integration on the subject attribute features and the parameter state features according to a multi-source integration module to obtain a multi-source integrated feature;

[0034] The multi-source integrated features are input into a plurality of the pre-trained encoders for preliminary feature extraction to obtain preliminary encoding features output by each of the pre-trained encoders.

[0035] In the embodiment of the present invention, by way of example, taking the real-time temperature control monitoring task of "a refrigerated truck transporting fresh strawberries" processed by a server of a cold chain logistics enterprise, the specific implementation process of "preliminary feature extraction of subject attribute data and parameter state data based on multiple pre-trained encoders" is described in detail.

[0036] The server first needs to convert the heterogeneous raw data (subject attribute data, parameter state data) into structured features that can be processed by the model. This process is completed by the feature mapping module and is specifically divided into two parts: subject attribute data transcoding and parameter state data transcoding. (1) Subject attribute data transcoding: conversion from discrete description to numerical features. Subject attribute data is static or semi-static data that reflects the inherent attributes and historical status of the transportation subject (such as a refrigerated truck). Taking the refrigerated truck as an example, the subject attribute data retrieved by the server from the enterprise database include: basic attributes: vehicle type (medium-sized van refrigerated truck), manufacturer ("Ice Bear"), and year of manufacture (2020); equipment attributes: refrigeration unit model ("Thermo King T-800"), compressor type (scroll type); maintenance records: text descriptions (such as "expansion valve replaced on March 15, 2023; temperature sensor calibrated on June 20, 2023"). For this data, the feature mapping module performs the following transcoding operations: Structured attribute transcoding: Categorical data (such as "vehicle type", "manufacturer", and "refrigeration unit model") are mapped to numerical values ​​using a predefined coding table. For example, "medium-sized van refrigerated truck" corresponds to code 002, "Ice Bear" corresponds to code 005, and "Thermo King T-800" corresponds to code 010; numerical data (such as "year of manufacture") are directly normalized (for example, 2020 is converted to (2020-2018) / (2023-2018) = 0.4, assuming the data range is 2018-2023). Text maintenance record transcoding: The text description of the maintenance record is converted into a low-dimensional vector using a pre-trained word embedding model (such as Word2Vec). For example, the word vector corresponding to "replace expansion valve" captures semantic information such as "repair" and "key components", and ultimately the text of the entire maintenance record is converted into a 32-dimensional vector (through average pooling or attention-weighted aggregation of word vectors). After transcoding, the subject attribute data is converted into subject attribute features (such as a 128-dimensional vector), which contains implicit information such as vehicle basic capabilities and equipment reliability. (2) Parameter state data transcoding: conversion from time series signals to feature vectors. Parameter state data is dynamic data that reflects the real-time operating status of the temperature control system. Taking the refrigerated truck as an example, the parameter state data collected by the server through the on-board IoT device in real time include: temperature time series: temperature values ​​of 6 measuring points in the car (collected every 30 seconds, a total of 120 time points in the last hour, such as [2.1℃, 2.3℃, ..., 2.5℃]); equipment operation data: compressor current (such as [14.8A, 15.2A, ..., 15.0A], 120 time points), operation mode ("strong cooling" or "energy saving", discrete state); environmental interference data: door opening and closing times (the door has been opened 3 times during this transportation), and the duration of each door opening (such as [120s, 180s, 90s]).The feature mapping module performs the following transcoding operations on this data: Time series data windowing: Continuous time series data such as temperature and current are truncated into fixed time windows (e.g., 1 hour) to form a two-dimensional matrix (time step × measurement point / parameter). For example, a matrix of 6 temperature measurement points × 120 time steps (720-dimensional raw data) is normalized using a sliding window (subtracting the mean and dividing by the standard deviation) to eliminate dimensionality. Discrete state encoding: Discrete states such as "operating mode" are encoded using one-hot encoding (e.g., "strong cooling" corresponds to [1,0] and "energy saving" corresponds to [0,1]. Statistics such as "door opening count" and "door opening duration" are directly normalized (e.g., if the maximum number of door openings is 5, the current 3 is converted to 0.6). After transcoding, the parameter state data is converted into parameter state features (e.g., a 256-dimensional vector), which contains dynamic information such as temperature fluctuation patterns, device operating stability, and environmental interference intensity. The main attribute features and parameter state features reflect the "inherent capabilities" and "real-time status" of the transport entity, respectively. However, single-dimensional features may lose cross-information (such as "the vehicle has strong cooling capacity but abnormal current current"). Therefore, the server fuses the two types of features into a multi-source integrated feature through a multi-source integration module. The specific integration method is feature splicing + cross-dimensional interaction: Feature splicing: The main attribute features (128 dimensions) and parameter state features (256 dimensions) are directly spliced ​​by dimension to form an initial 384-dimensional fusion vector; Cross-dimensional interaction: The nonlinear association between the two types of features is learned through a fully connected layer (such as 384-dimensional input - 256-dimensional output). For example, if the dimension value corresponding to "refrigeration unit model" in the main attribute feature is high (indicating strong theoretical cooling capacity), while the dimension value corresponding to "compressor current" in the parameter state feature is abnormal (indicating excessive actual power consumption), the fully connected layer will highlight this contradiction through weight adjustment and output a 256-dimensional multi-source integrated feature, one of which may represent "the deviation between the theoretical capacity of the equipment and the actual operating status." Through multi-source integration, the server integrates the scattered subject attributes and parameter state information into a more comprehensive feature representation, providing richer input for the feature extraction of the subsequent encoder. The server inputs the multi-source integrated features into multiple pre-trained encoders. Each encoder is pre-trained on historical data for different data types (such as structured attributes and time series sensors) to extract more abstract implicit features. Take the two key encoders in this scenario as an example: (1) Structured attribute encoder: extracts the "subject reliability" feature. This encoder adopts a multi-layer perceptron (MLP) structure. In the pre-training stage, it uses the historical subject attribute data (such as model and maintenance records) of 100,000 refrigerated trucks to learn implicit features such as "vehicle reliability" and "equipment aging degree". After inputting the multi-source integrated features (256 dimensions), the encoder performs feature compression through three layers of fully connected layers (256-128-64) and outputs a 64-dimensional preliminary encoding feature.For example, if the word vector corresponding to the input feature "maintenance record" contains the semantic information of "frequent replacement of compressors", the encoder may reflect the probability of "low reliability of equipment components" (such as 0.8) in one dimension of the output feature. (2) Time series sensor encoder: extract "state trend" features. The encoder adopts a bidirectional LSTM structure and uses millions of time series data such as temperature and current for training in the pre-training stage to learn time series features such as "early signals of temperature exceeding the standard" and "current pattern of equipment failure". After inputting multi-source integrated features (256 dimensions, including time series window data of temperature and current), the encoder captures the time dependency through the LSTM layer (such as the trend of "the temperature rising by 0.1℃ every 5 minutes in the past 30 minutes") and outputs a 64-dimensional preliminary encoding feature through the pooling layer (such as taking the hidden state at the last moment). For example, if the time series data of "compressor current" in the input feature shows a "gradually rising" pattern, the encoder may reflect the trend intensity of "abnormal increase in compressor load" (such as 0.7) in one dimension of the output feature. The preliminary encoding features output by each pre-trained encoder (such as the 64-dimensional output of the structured attribute encoder and the 64-dimensional output of the time series sensor encoder) retain the core information of the original data: the former focuses on the inherent capabilities and potential risks of the transportation entity, and the latter focuses on the real-time status and changing trends of the temperature control system. These features will serve as the basic input for subsequent scene association fusion and time series analysis, ensuring that subsequent steps can analyze the possibility of temperature control anomalies from multiple dimensions and levels. In summary, the server converts heterogeneous data into unified features through the feature mapping module, fuses the subject and status information through the multi-source integration module, and then extracts implicit features through the pre-trained encoder, completing the conversion process from raw data to high-quality features, providing key support for subsequent scene analysis and time series mining.

[0037] In an embodiment of the present invention, the scene processing unit corresponding to multiple monitoring scenes performs scene association fusion on the preliminary coding features output by each pre-trained encoder to obtain the scene association features output by each scene processing unit, which can be implemented through the following examples.

[0038] Determining a first routing weight corresponding to each of the pre-trained encoders according to the first dynamic routing component corresponding to each of the monitoring scenarios, and performing weighted fusion on the preliminary encoding features output by each of the pre-trained encoders according to the first routing weight corresponding to each of the pre-trained encoders to obtain a first scene fusion feature corresponding to each of the monitoring scenarios;

[0039] The scene-related fusion is performed on each of the first scene fusion features according to the multiple scene processing units to obtain a scene-related feature output by each of the scene processing units.

[0040] In an embodiment of the present invention, for example, taking the temperature control monitoring task of "a refrigerated truck transporting fresh strawberries" processed by a server of a cold chain logistics enterprise as an example, combined with two typical monitoring scenarios of "refrigeration equipment failure" (first monitoring scenario) and "temperature continuously exceeding the standard" (second monitoring scenario), the specific implementation process of "scene association and fusion of preliminary coding features according to the scene processing units corresponding to multiple monitoring scenarios" is described in detail. The server needs to dynamically assign weights (first routing weights) of different pre-trained encoders to each monitoring scenario to highlight the core data dimensions of concern in the scenario. This process is completed by the first dynamic routing component corresponding to each scenario, which is essentially a small neural network (such as a fully connected layer + Softmax activation), with the input being the preliminary coding features of the pre-trained encoder and the output being the routing weights of each encoder (the sum is 1). Taking the "refrigeration equipment failure" scenario as an example, the core of this scenario is to identify whether the refrigeration unit (such as a compressor, expansion valve) has an abnormality, so it is necessary to focus on analyzing the real-time data of the equipment operation (such as the time series characteristics of the compressor current and voltage) and the historical reliability of the equipment (such as maintenance records, attribute characteristics of the unit model). At this point, the first dynamic routing component calculates the weights of each pre-trained encoder based on the input preliminary encoding features: For the structured attribute encoder (output feature A, 64 dimensions, reflecting inherent vehicle attributes): If its features include "frequent compressor replacements due to historical maintenance records" (corresponding to a dimension value of 0.8), the component will determine that the device is less reliable and need to reduce reliance on its "theoretical cooling capacity," assigning a lower weight (e.g., 0.3); for the time series sensor encoder (output feature B, 64 dimensions, reflecting real-time operating status): If its features include "compressor current fluctuation amplitude exceeds threshold" (corresponding to a dimension value of 0.9), the component will determine that real-time data is more critical for fault identification and assign a higher weight (e.g., 0.7). For the "persistent temperature exceeding the standard" scenario (determining whether the temperature will exceed the strawberry preservation range (-2°C to 5°C) for a long period of time), this scenario requires a comprehensive consideration of the vehicle's cooling capacity (structured attribute features) and temperature trends (time series sensor features). Therefore, the first dynamic routing component may assign more balanced weights (e.g., 0.5 for the structured attribute encoder and 0.5 for the time series sensor encoder). The server performs a weighted summation of the preliminary encoding features of each pre-trained encoder based on the first routing weight to obtain the first scene fusion feature for each scenario. Taking the "refrigeration equipment failure" scenario as an example, the fusion formula is: [{first scene fusion feature} = 0.3 × {feature A} + 0.7 × {feature B}], where feature A (structured attribute encoding feature) contains information about "historical reliability of the equipment" (for example, a dimension value of 0.2 indicates "good maintenance record"), and feature B (time series sensor encoding feature) contains information about "abnormal current fluctuation intensity" (for example, a dimension value of 0.8 indicates "severe current fluctuation").After weighting, the corresponding dimension value of the fusion feature is: [0.3×0.2+0.7×0.8=0.62], which further highlights the key fault signal of "current anomaly" while retaining the reference information of the historical status of the equipment. After obtaining the first scene fusion feature, the server further mines the implicit correlation features directly related to the scene through the scene processing unit corresponding to each monitoring scene. The scene processing unit usually contains modules such as attention mechanism, convolution layer or graph neural network, and its goal is to filter out the most critical information for the current scene from the fusion feature. (1) Correlation fusion of the "refrigeration equipment failure" scene. The core of the "refrigeration equipment failure" scene processing unit is to identify the early signals of equipment abnormality (such as compressor wear, refrigerant leakage), so its internal structure is designed as "attention mechanism + feature refinement": Attention mechanism: Calculate the importance of each dimension in the first scene fusion feature to "equipment failure". For example, dimensions such as "compressor current mean", "current standard deviation", and "historical compressor replacement times" in the feature will be focused on. The calculation of attention weights is based on pre-trained fault samples (e.g., in historical samples of failures caused by compressor wear, the value of the "current standard deviation" dimension is generally high), and finally an attention coefficient is assigned to each dimension (e.g., the coefficient of the "current standard deviation" dimension is 0.9, and the coefficient of the "historical replacement times" dimension is 0.8). Feature refinement: The first scene fusion feature is multiplied element-by-element by the attention weight to obtain the refined feature. For example, if the original value of the "current standard deviation" dimension is 0.7 (indicating large current fluctuations), it becomes 0.63 after multiplying it by the attention coefficient of 0.9. This value more directly reflects the "contribution of current anomaly to the failure." Finally, the "refrigeration equipment failure" scene processing unit outputs scene association features (e.g., a 128-dimensional vector), in which one dimension may clearly represent the "probability of current anomaly caused by compressor wear" (e.g., 0.85), and another dimension represents the "possibility of refrigerant leakage" (e.g., 0.12), highlighting the key information directly related to the equipment failure. (2) Association fusion of the "temperature continues to exceed the standard" scene. The core of the "temperature continues to exceed the standard" scenario processing unit is to determine whether the temperature will exceed the threshold for a long time due to insufficient equipment capacity or environmental interference (such as frequent door opening). Therefore, its internal structure is designed as "time series-attribute cross analysis + threshold prediction": Time series-attribute cross analysis: The "temperature rise rate" (time series feature) and "vehicle cooling capacity" (attribute feature) in the first scene fusion feature are cross-calculated. For example, if the "temperature rise rate" is 0.5°C / hour (feature dimension value 0.7) and the "vehicle cooling capacity" is 2°C / hour (feature dimension value 0.6), the cross-analysis result is "cooling capacity can only offset 50% of the temperature rise" (0.7 / 0.6≈1.17, exceeding 1 indicates insufficient cooling capacity). Threshold prediction: The cross-analysis result is mapped to the "temperature exceedance probability" dimension through the fully connected layer.For example, if the cross-analysis result is 1.17 (insufficient cooling capacity), combined with the "external ambient temperature 30°C" (feature dimension value 0.9) and "number of door openings and closings 3 times" (feature dimension value 0.6), the fully connected layer outputs the "probability of temperature exceeding the standard in the next two hours" (e.g., 0.92). Finally, the "sustained temperature exceeding the standard" scenario processing unit outputs scenario-related features (e.g., a 128-dimensional vector), one dimension of which explicitly represents the "risk of temperature exceeding the standard due to insufficient cooling capacity" (e.g., 0.92), and another dimension represents the "intensity of the impact of environmental interference (door opening) on ​​temperature" (e.g., 0.75), providing targeted features for subsequent time series conduction analysis. Through this process, the server generates scenario-related features for each monitoring scenario, specifically tailored to the specific needs of that scenario. For example, in the "refrigeration equipment failure" scenario, the weights of dimensions such as "abnormal current" and "history of poor maintenance" are significantly increased, directly pointing to the root cause of the equipment failure. In the "sustained temperature exceeding the standard" scenario, the weights of dimensions such as "matching of cooling capacity and temperature rise rate" and "intensity of environmental interference" are highlighted, clearly identifying the driving factors of temperature exceeding the standard. These features not only retain the multi-source information of the original data, but also transform common features into high-value features specific to the scenario through scenario-based weight allocation and deep correlation analysis, laying a key foundation for subsequent full-scenario prediction based on time series correlation features.

[0041] In an embodiment of the present invention, performing time series correlation feature mining on the preliminary coding features output by each pre-trained encoder according to the time series correlation model to obtain the time series correlation features output by the time series correlation model includes:

[0042] Determining a second routing weight corresponding to each of the pre-trained encoders according to the second dynamic routing component corresponding to each of the monitoring scenarios, and performing weighted fusion on the preliminary encoding features output by each of the pre-trained encoders according to the second routing weight corresponding to each of the pre-trained encoders to obtain a second scene fusion feature corresponding to each of the monitoring scenarios;

[0043] Performing time series correlation feature mining on each of the second scene fusion features according to the time series correlation model to obtain the time series correlation features output by the time series correlation model.

[0044] In the embodiment of the present invention, for example, taking the temperature control monitoring task of "a refrigerated truck transporting fresh strawberries" processed by a server of a cold chain logistics enterprise as an example, combined with "refrigeration equipment failure" (the first monitoring scenario) and "temperature continues to exceed the standard" (the second monitoring scenario, which has a time series transmission relationship with the first scenario), the specific implementation process of "mining time series correlation features of preliminary coding features based on the time series correlation model" is described in detail. The server needs to dynamically assign the weight of the pre-trained encoder (the second routing weight) to each monitoring scenario to highlight the key data of the scenario in the time dimension. This process is completed by the second dynamic routing component corresponding to each scenario. Its essence is a small neural network (such as a fully connected layer + Softmax) trained based on historical time series patterns. The input is the preliminary coding feature sequence of the pre-trained encoder (time step t-2, t-1, t), and the output is the weight of each encoder in the time series analysis (the sum is 1). (1) Routing weight calculation driven by time series characteristics. Take the time series transmission scenario of "refrigeration equipment failure - persistent temperature exceeding the standard" as an example: In the "refrigeration equipment failure" scenario (the first monitoring scenario), the key is to capture early signals of equipment anomalies (such as small fluctuations in compressor current). These signals primarily exist in time series sensor data (e.g., current and temperature time series). Therefore, the second dynamic routing component assigns a higher weight to the time series sensor encoder. For example, if the "current standard deviation" feature of the input time series sensor encoder suddenly increases at time t (the dimension value increases from 0.3 to 0.7), the component determines that real-time time series data is more critical for fault identification and assigns a weight of 0.8 to the time series sensor encoder and a weight of 0.2 to the structured attribute encoder (based solely on the historical reliability of the equipment). In the "persistent temperature exceeding the standard" scenario (the second monitoring scenario), the key is to analyze the cumulative upward trend in temperature after the equipment failure. This requires both the temporal continuity of the equipment status (e.g., persistent compressor anomalies after the failure) and the temporal variation of the temperature (e.g., a 0.5°C increase per hour). Therefore, the second dynamic routing component assigns a balanced weight. For example, if the initial encoding features of the structured attribute encoder (reflecting the cooling capacity of the equipment) and the initial encoding features of the time series sensor encoder (reflecting the temperature rise rate) are input, and the dimension value of "cooling capacity" is 0.6 (medium) and the dimension value of "temperature rise rate" is 0.8 (fast), the component assigns a weight of 0.4 to the structured attribute encoder and a weight of 0.6 to the time series sensor encoder. (2) Weighted fusion obtains the second scene fusion feature. The server performs weighted summation on the initial encoding features (time series) of each pre-trained encoder according to the second routing weight to obtain the second scene fusion feature sequence (time step t-2, t-1, t) for each scene. Taking the "refrigeration equipment failure" scene as an example, the fusion formula is: [{second scene fusion feature} t =0.2×{Structured attribute encoding feature} t+0.8×{Timing sensor encoding characteristics} t]. Among them, the value of a certain dimension of the structured attribute encoding feature at time t is 0.4 (indicating "the equipment has a good history of maintenance"), and the value of a certain dimension of the time series sensor encoding feature at time t is 0.7 (indicating "severe current fluctuations"). After weighting, the corresponding dimension value of the fusion feature is: [0.2×0.4+0.8×0.7=0.64]. This value highlights the real-time signal of the fault "current abnormality" while retaining the reference information of the historical status of the equipment, forming a fusion feature of the time series (such as [t-2:0.5,t-1:0.6,t:0.64]). The server inputs the second scene fusion feature sequence of each scene into the time series association model, which contains a time series convolution association module and a time series nonlinear transformation module. The goal is to capture the time conduction relationship between scenes (such as how the temperature gradually exceeds the standard over time after the equipment failure occurs). (1) Time series convolution association module: calculates the correlation between time steps. This module uses a 1D convolution layer (convolution kernel size 3, stride 1) to extract local temporal patterns from the fused feature sequence of the second scenario (time window of length 3: t-2, t-1, t). For example, if the fused feature sequence of the "refrigeration equipment failure" scenario is input [0.5, 0.6, 0.64], the convolution operation will calculate the difference between adjacent time steps (such as an increase of 0.04 from t-1 to t) and the pattern (such as a continuous rise), and output the temporal correlation weight of each time step (indicating the degree of influence of the time step on the subsequent scenario). Specifically, the weight of the convolution kernel is learned through historical fault data during the training phase: if the "current abnormality at time t" is often accompanied by "a slight increase in temperature at time t+1" in the history, the convolution kernel will assign a higher weight to the feature at time t. For example, the output temporal correlation weight may be [t-2: 0.2, t-1: 0.3, t: 0.5], indicating that the feature at time t has the greatest impact on the subsequent temperature exceeding the standard. (2) Temporal nonlinear transformation module: aggregates temporal information. This module uses a gated recurrent unit (GRU) or attention mechanism to globally aggregate timing-related weights and fusion feature sequences, and output timing-related features (reflecting the temporal conduction rules between scenes).For example, consider the scenario of "refrigeration equipment failure—sustained temperature exceeding the specified value." Given the fused feature sequence [0.5, 0.6, 0.64] and time-series correlation weights [0.2, 0.3, 0.5] for the "refrigeration equipment failure" scenario, the GRU unit processes the data time-step by time, recording the trend of "the current anomaly gradually intensifies from t-2 to t." Simultaneously, given the fused feature sequence for the "sustained temperature exceeding the specified value" scenario (assuming it's [t-2: 0.3, t-1: 0.4, t: 0.5], reflecting a gradual increase in temperature), the GRU unit learns the correlation pattern of "temperature increasing over time after equipment failure." Ultimately, through the GRU's hidden state output (e.g., a 128-dimensional vector), one dimension of the time-series correlation features explicitly indicates "the temperature will begin to rise one hour after the equipment failure" (e.g., a confidence level of 0.8), while another dimension indicates "the probability of the temperature exceeding the specified value two hours after the failure" (e.g., 0.9). Through this process, the server-generated time-series correlation features clearly define the temporal transmission relationship between scenarios. For example, if the value of a dimension in the temporal correlation feature of the "refrigeration equipment failure" scenario is 0.8 (indicating the "probability of the current equipment failure"), and the value of the corresponding dimension in the temporal correlation feature of the "temperature continues to exceed the standard" scenario is 0.9 (indicating "temperature exceeds the standard 2 hours after the failure"), the server can trigger an early warning 1 hour in advance, prompting monitoring personnel to check the equipment or adjust the transportation plan; if the temporal correlation feature shows that the transmission time between "equipment failure" and "temperature exceeding the standard" is shortened from the historical 2 hours to 1 hour (for example, a dimension value changes from 0.7 to 0.9), the server can identify the new pattern of "equipment aging leading to accelerated failure impact" and optimize subsequent anomaly prediction strategies. In summary, the temporal correlation model transforms discrete time point features into continuous features that reflect the law of scenario transmission through dynamic routing weighting and temporal pattern mining, providing key support in the time dimension for full-scenario prediction and significantly improving the foresight and accuracy of cold chain transportation temperature control anomaly monitoring.

[0045] In an embodiment of the present invention, multiple monitoring scenarios include a first monitoring scenario and a second monitoring scenario, and the first monitoring scenario is located before the second monitoring scenario in a sequence with a timing conduction relationship; the timing correlation feature mining is performed on each fusion feature of the second scenario according to the timing correlation model to obtain the timing correlation feature output by the timing correlation model, which can be implemented through the following examples.

[0046] performing time series correlation feature mining on the second scene fusion feature corresponding to the first monitoring scene according to the time series correlation model to obtain the time series correlation feature corresponding to the first monitoring scene, and performing time series correlation feature mining on the second scene fusion feature corresponding to the first monitoring scene and the second scene fusion feature corresponding to the second monitoring scene according to the time series correlation model to obtain the time series correlation feature corresponding to the second monitoring scene;

[0047] The time series correlation features corresponding to the first monitoring scenario and the time series correlation features corresponding to the second monitoring scenario are determined as the time series correlation features output by the time series correlation model.

[0048] In an embodiment of the present invention, for example, taking the temperature control monitoring task of "a refrigerated truck transporting fresh strawberries" processed by a server of a cold chain logistics enterprise as an example, "refrigeration equipment failure" is set as the first monitoring scenario (S1), and "temperature continues to exceed the standard" is set as the second monitoring scenario (S2), and S1 precedes S2 in the timing conduction sequence (that is, equipment failure may cause temperature to exceed the standard). The following details how the server mines the timing correlation features of S1 and S2 through the timing correlation model. The server has completed the previous steps: collecting key data of S1 (such as compressor current timing, equipment maintenance records) and key data of S2 (such as temperature timing, external ambient temperature); after processing by the pre-trained encoder, the preliminary coding feature sequence corresponding to S1 is obtained (F1 t-2 ,F1 t-1 ,F1 t ) and the preliminary coding feature sequence corresponding to S2 (F2 t-2 ,F2 t-1 ,F2 t ); assign routing weights to S1 and S2 through the second dynamic routing component (e.g., S1’s temporal sensor encoder weight is 0.8, and its structural attribute encoder weight is 0.2; S2’s temporal sensor encoder weight is 0.6, and its structural attribute encoder weight is 0.4), and obtain the second scene fusion feature sequence (C1) of S1. t-2 ,C1 t-1 ,C1 t ) and the second scene fusion feature sequence of S2 (C2 t-2 ,C2 t-1 ,C2 t). For example, the C1 sequence of S1 may be: [C1 = [0.5, 0.6, 0.64]] (corresponding to the fusion feature values ​​at time t-2, t-1, and t respectively, the higher the value, the greater the risk of equipment abnormality); the C2 sequence of S2 may be: [C2 = [0.3, 0.4, 0.5]] (corresponding to the fusion feature values ​​at time t-2, t-1, and t respectively, the higher the value, the greater the risk of temperature exceeding the standard). The server first performs temporal correlation feature mining on the second scene fusion feature sequence (C1) of S1, with the goal of identifying the time evolution law of S1 itself (such as the time pattern of equipment failure from initiation to manifestation). (1) The temporal convolutional correlation module processes the C1 sequence of S1. The temporal convolutional correlation module uses a 1D convolution layer (convolution kernel size 3, step size 1), and the sliding window covers the three time steps of C1 (t-2, t-1, t). The weights of the convolution kernel are learned during the training phase through historical equipment failure data. For example, if the "current anomaly at time t-1" is often followed by "fault confirmation at time t", the convolution kernel will strengthen the association between the time steps t-1 and t. In the specific calculation, the C1 sequence [0.5, 0.6, 0.64] is input, and the convolution operation calculates the difference between adjacent time steps (such as the increase of 0.04 from t-1 to t) and the continuous pattern (such as the continuous upward trend of "0.5-0.6-0.64"), and outputs the time-related weight of each time step (W1 = [0.2, 0.3, 0.5]), indicating that the feature at time t has the greatest impact on the development of S1 itself (weight 0.5). (2) The temporal nonlinear transformation module aggregates the temporal information of S1. The temporal nonlinear transformation module uses GRU units to process the C1 sequence and W1 weights time by time step and record the time evolution of S1. For example: at time t-2: C1 = 0.5 (equipment anomaly germinates), GRU hidden state h-2 = tanh (W h *h-3+W x *0.5)(initial h-3=0); at time t-1: C1=0.6 (abnormal aggravation), h-1=tanh(W h *h-2+W x *0.6); at time t: C1=0.64 (abnormally significant), h t =tanh(W h *h-1+W x *0.64). Finally, GRU outputs the temporal correlation feature of S1 (H1=h t), for example, one dimension in the 128-dimensional vector clearly represents the "confirmation probability of equipment failure at time t" (such as 0.85), and the other dimension represents the "time span from the initiation of the failure to its manifestation" (such as 2 hours). The server further conducts joint mining on the C1 sequence of S1 and the C2 sequence of S2, with the goal of capturing the temporal transmission relationship from S1 to S2 (such as how the temperature exceeds the standard over time after the equipment failure occurs). (1) The temporal convolution association module processes the joint sequence. The server splices the fused feature sequences of S1 and S2 into a joint sequence (C1 t-2 ,C2 t-2 ,C1 t-1 ,C2 t-1 ,C1 t ,C2 t ), input convolution kernel size 6 1D convolution layer (covering the entire time window). The weight of the convolution kernel is learned by historical transmission data during the training phase. For example, if the history shows that "equipment failure is confirmed at time t", "temperature begins to rise at t+1 hour" and "exceeds the standard at t+2 hour", then the convolution kernel will strengthen C1 t With C2 t+1 、C2 t+2 Taking the current data as an example, the joint sequence is [0.5, 0.3, 0.6, 0.4, 0.64, 0.5], and the convolution operation calculates the temporal correlation across scenes (such as C1 t =0.64 and C2 t+1 The potential relationship between C1 and C2 is obtained by outputting the temporal correlation weight across scenes (W2 = [0.1, 0.1, 0.2, 0.2, 0.3, 0.1]), which represents the potential relationship between C1 and C2. t (Equipment failure confirmation) has the greatest impact on the subsequent S2 (weight 0.3). (2) The temporal nonlinear transformation module aggregates cross-scenario temporal information, and the GRU unit is expanded into a bidirectional structure to process the time series of S1 and S2 at the same time and learn the conduction mode of both. For example: Forward transmission: from t-2 to t, record the forward conduction of "equipment failure aggravation - temperature rises slowly"; Backward transmission: from t to t-2, verify the reverse association of "whether the temperature rise is caused by equipment failure"; the final hidden state h t' Integrate the bidirectional information and output the temporal correlation feature of S2 (H2 = h t'For example, one dimension of the 128-dimensional vector represents "the temperature will begin to rise one hour after the device fails" (e.g., a confidence level of 0.8), while another dimension represents "the probability of the temperature exceeding the standard two hours after the failure" (e.g., 0.92). The server combines the temporal correlation features of S1 (H1) and S2 (H2) as the output of the temporal correlation model. For example, H1 includes "the probability of confirming the device failure at time t is 0.85" and "the time span of the failure development is 2 hours"; H2 includes "the probability of the temperature starting to rise one hour after the failure is 0.8" and "the probability of the temperature exceeding the standard two hours after the failure is 0.92." These features directly reflect the temporal transmission patterns between scenarios and provide a key basis for the time dimension of subsequent full-scenario predictions. For example, if H1 indicates "the probability of confirming the device failure is 0.85" and H2 indicates "the probability of the temperature exceeding the standard two hours after the failure is 0.92," the server can trigger an early warning one hour in advance, prompting monitoring personnel to inspect the equipment or adjust the cooling parameters before t+1 hour to prevent the strawberries from spoiling due to excessive temperatures. Through the above process, the server can not only identify the temporal evolution of a single scenario (such as how an equipment failure develops over time), but also capture the transmission relationship across scenarios (such as how an equipment failure causes a temperature exceeding the standard). This mining of time-series correlation features has upgraded cold chain temperature control monitoring from "post-event alarm" to "pre-event warning", significantly improving the reliability of cold chain transportation and the quality of goods preservation. For example, a refrigerated truck once had a slow increase in current due to hidden wear of the compressor (S1 budding). The server identified the "fault development time span of 2 hours" through H1 and predicted "temperature exceeding the standard 2 hours after the fault" through H2, notifying the driver 1.5 hours in advance to carry out maintenance and prevent the spoilage of 50,000 yuan worth of strawberries. In summary, the time series correlation model achieves accurate perception of the entire cycle of "occurrence-transmission-evolution" of temperature control anomalies by jointly mining the time series of multiple scenarios, providing core technical support for the intelligent management of cold chain transportation.

[0049] In an embodiment of the present invention, the temporal association model includes at least one temporal association branch, each of which includes a temporal convolutional association module and a temporal nonlinear transformation module; performing temporal association feature mining on each second scene fusion feature according to the temporal association model to obtain the temporal association features output by the temporal association model includes:

[0050] Inputting each second scene fusion feature into the temporal convolution association module to perform temporal correlation calculation to obtain a temporal correlation weight corresponding to each second scene fusion feature;

[0051] The temporal correlation weight corresponding to each second scene fusion feature is input into the temporal nonlinear transformation module to perform sliding window feature aggregation to obtain the temporal correlation feature output by the temporal correlation model.

[0052] In an embodiment of the present invention, for example, taking a cold chain logistics enterprise server processing the temperature control monitoring task of "a refrigerated truck transporting fresh strawberries" as an example, combined with the second scene fusion feature sequence of the two monitoring scenes of "refrigeration equipment failure" (S1) and "temperature continuously exceeding the standard" (S2), the workflow of the time series association branch (including the time series convolution association module and the time series nonlinear transformation module) in the time series association model is described in detail. The server has completed: collecting key data of S1 (such as compressor current time series, equipment maintenance records) and key data of S2 (such as temperature time series, external ambient temperature); processing through the pre-trained encoder to obtain the preliminary coding features of S1 and S2; and weighted fusion through the second dynamic routing component to obtain the second scene fusion feature sequence of S1 (C1=[c1 t-2 ,c1 t-1 ,c1 t ]=[0.5,0.6,0.64])(the higher the value, the greater the risk of equipment abnormality), the second scene fusion feature sequence of S2 (C2=[c2 t-2 ,c2 t-1 ,c2 t]=[0.3,0.4,0.5])(the higher the value, the greater the risk of temperature exceeding the standard). The server inputs the second scene fusion feature sequences of S1 and S2 into the temporal convolution association module respectively. The core of this module is to capture the local correlation between time steps through 1D convolution operation. Taking the sequence (C1) of S1 as an example, the processing process of the module is as follows: (1) Design and input processing of the convolution kernel. The temporal convolution association module adopts a 1D convolution layer with a convolution kernel size of 3 (covering 3 consecutive time steps: t-2, t-1, t) and a step size of 1 to ensure that the local pattern of each time step is fully extracted. The weight of the convolution kernel is learned through historical equipment failure data during the training phase. For example, if the "abnormal current fluctuation at time t-1" is often followed by "fault confirmation at time t" in the history, the convolution kernel will strengthen the correlation between time steps t-1 to t. (2) Calculate the time series correlation. Input S1 (C1 = [0.5, 0.6, 0.64]). Perform a sliding dot product operation on the convolution kernel and the input sequence to calculate the local correlation of each time step. Specifically: Window 1 (t-2, t-1, t): The convolution kernel weights are ([w_1, w_2, w_3]) (obtained through training, for example, ([0.2, 0.3, 0.5])). Calculate (0.5 × 0.2 + 0.6 × 0.3 + 0.64 × 0.5 = 0.62); this value reflects the comprehensive correlation of the time steps within the window. Finally, the time series correlation weight (W1 = [0.2, 0.3, 0.5]) for each time step is output (the sum of the weights is 1, and the weight at time t is the highest, indicating that it has the greatest impact on the device failure). Similarly, when processing S2's (C2=[0.3,0.4,0.5]), the convolution kernel weights are adjusted according to the historical temperature exceedance data (for example, more attention is paid to the pattern of continuous temperature rise), and the time series related weights (W2=[0.1,0.2,0.7]) are output (the weight is the highest at time t, indicating that the current temperature state has the greatest impact on subsequent exceedances). The server inputs the time series related weights (W1) and (W2) into the time series nonlinear transformation module with the corresponding second scene fusion feature sequences (C1) and (C2). The module adopts a sliding window + GRU structure, and the goal is to aggregate the correlation of local time steps into global time series correlation features. The sliding window size is set to 3 (the same as the convolution kernel size) to ensure that the complete time evolution cycle is covered. For example, when processing S1's (C1) and (W1), the window contains ((c1 t-2 ,w1 t-2 ),(c1 t-1 ,w1 t-1 ),(c1 t ,w1 t)), that is, ((0.5, 0.2), (0.6, 0.3), (0.64, 0.5)). The GRU unit processes the data in the window step by time, and records the time dependency by updating the gate and resetting the gate: At time t-2: input ((0.5, 0.2)), GRU hidden state (h t-2 =tanh(W h ·h t-3 +W x ·(0.5×0.2)))(initial(h t-3 =0)); t-1 time: input ((0.6,0.3)), update gate (z t =sigma(W z ·h t-2 +U z (0.6×0.3))), reset gate (r t =sigma(W r ·h t-2 +U r (0.6×0.3))), new candidate state (h') t =tanh(W·(r t ⊙h t-2 )+U·(0.6×0.3))), the final hidden state (h t-1 =(1-z t )⊙h t-2 +z t ⊙h' t ); at time t: input ((0.64,0.5)), repeat the above steps to get the final hidden state (h t ). The final hidden state of GRU (h t) is the time series correlation feature of S1 (e.g., a 128-dimensional vector). For example, one dimension of this feature represents the "rate of equipment failure development from t-2 to t" (e.g., 0.7, indicating a 35% increase in failure risk per hour), and another dimension represents the "probability of failure confirmation at time t" (e.g., 0.85). Similarly, after processing (C2) and (W2) of S2, the GRU outputs the time series correlation feature of S2 (e.g., a 128-dimensional vector), one of which represents the "rate of temperature increase from t-2 to t" (e.g., 0.6, indicating a 0.3°C increase per hour), and another dimension represents the "probability of temperature exceeding the standard 2 hours after time t" (e.g., 0.92). Through the collaborative work of the time series convolutional association module and the time series nonlinear transformation module, the time series association features generated by the server can accurately reflect the temporal evolution of temperature control anomalies: For S1 (equipment failure), the time series association features clearly define the "rate at which the failure risk increases over time" and the "current confirmation probability", helping the server predict its development trend at the incipient stage of the failure (t-2) (for example, "the failure probability will increase from 0.5 to 0.85 in the next hour"); for S2 (temperature exceeding the standard), the time series association features are combined with the conduction relationship of S1 (for example, "after the equipment failure is confirmed, the temperature rises by 0.3°C per hour") to predict the time point of the standard exceeding the standard in advance (for example, "the temperature will reach 6°C in t+2 hours, exceeding the shelf life of strawberries"). For example, during the transportation of a refrigerated truck, the server discovered through the time-series correlation branch that the time-series correlation feature of S1 indicated that "the risk of equipment failure reached 0.85 at time t and increased by 35% per hour," and the time-series correlation feature of S2 indicated that "the temperature would exceed the standard two hours after the failure (with a probability of 0.92)." The server immediately triggered an alert, prompting the driver to check the compressor before t+1 hours, ultimately avoiding the temperature exceeding the standard and the spoilage of strawberries caused by the equipment failure. In summary, the time-series correlation branch, through "local correlation calculation + global feature aggregation," transforms discrete time point data into continuous features that reflect the evolution of events, providing core technical support for the forward-looking monitoring of temperature control anomalies in cold chain transportation.

[0053] In an embodiment of the present invention, the full-scenario prediction calculation is performed based on the scenario association features and the time series association features to obtain the temperature control anomaly monitoring result corresponding to each monitoring scenario, which can be implemented through the following examples.

[0054] Determine, according to the third dynamic routing component corresponding to each of the monitoring scenarios, the routing weight corresponding to the scenario-related feature and the routing weight corresponding to the timing-related feature, and perform weighted fusion on the scenario-related feature and the timing-related feature based on the routing weight corresponding to the scenario-related feature and the routing weight corresponding to the timing-related feature to obtain a third scenario fusion feature corresponding to each of the monitoring scenarios;

[0055] According to the fully connected module, the result dimension mapping is performed on the third scene fusion feature corresponding to each monitoring scene to obtain the temperature control abnormality monitoring result corresponding to each monitoring scene.

[0056] In this embodiment of the present invention, for example, a server at a cold chain logistics company processes a temperature control monitoring task for a refrigerated truck transporting fresh strawberries. Combining two monitoring scenarios, namely, "refrigeration equipment failure" (S1) and "sustained temperature exceeding the specified value" (S2), the server details how it dynamically integrates scenario-related features with time-series-related features to ultimately output temperature control anomaly monitoring results. The server has completed preliminary processing: Scenario-related features: Through the scenario processing unit, the scenario-related features of S1 (refrigeration equipment failure) focus on real-time status information such as "current compressor current fluctuation intensity" and "historical equipment maintenance failure records." The scenario-related features of S2 (sustained temperature exceeding the specified value) focus on current conflicts such as "the difference between the current temperature and the strawberry preservation threshold" and "the matching degree between refrigeration capacity and the temperature rise rate." Time-series-related features: Through the time-series-related model, the time-series-related features of S1 reflect the "trend of increasing equipment failure risk over time" (e.g., "failure probability increased from 30% to 80% in the past two hours"). The time-series-related features of S2 reflect the "temporal pattern of temperature rise after equipment failure" (e.g., "temperature began exceeding the specified value one hour after the failure"). The server configures a third dynamic routing component for each monitoring scenario. Its core is to dynamically determine the contribution ratio of "current status" (scenario-related features) and "time trend" (time series-related features) to abnormality prediction based on scenario requirements. (1) Weight allocation of S1 (refrigeration equipment failure). The core of S1 is to identify whether the equipment is "currently abnormal", so it relies more on scenario-related features (reflecting the real-time operating status of the equipment). The server analyzes the scenario-related features of S1 and finds that "the current current fluctuation amplitude of the compressor exceeds the normal range by 2 times" (strong real-time signal). At the same time, the time series-related features show that "the probability of failure has increased from 50% to 80% in the past hour" (time trend support). At this time, the dynamic routing component determines that "current current abnormality" is direct evidence of the failure, assigns a higher weight (such as 0.7) to the scenario-related features and a lower weight (such as 0.3) to the time series-related features, ensuring that the prediction results are more in line with the real-time status of the equipment. (2) Weight allocation of S2 (temperature continues to exceed the standard). The core of S2 is to predict whether the temperature "exceeds the standard due to accumulation over time", so it relies more on time series-related features (reflecting the temperature rise pattern after the failure). The server analyzes S2's scene-related features and discovers that "the current temperature is approaching the upper limit of freshness preservation (4.5°C)" (the current state is critical), while the time-series-related features show that "after the equipment failure, the temperature rises by 0.5°C per hour" (the time trend is clear). At this point, the dynamic routing component determines that "the trend of temperature rising over time" is the key driver of the exceeding standard, and assigns a higher weight (such as 0.6) to the time-series-related features and a lower weight (such as 0.4) to the scene-related features to ensure that the prediction results focus more on the future evolution direction. Based on the dynamically assigned weights, the server weightedly fuses the scene-related features and the time-series-related features to generate a third scene fusion feature for each scene.This process is like "scoring different information sources," combining "current state" and "time trend" information based on importance to form a more comprehensive feature representation. In S1's scenario-related features, the dimension value of "compressor current anomaly intensity" is 0.8 (indicating severe fluctuations); in the time-series-related features, the dimension value of "failure probability increasing rate over time" is 0.7 (indicating a 35% increase per hour). After fusing with weights of 0.7 (scenario) and 0.3 (time series), the fused value of this dimension is: [0.8 × 0.7 + 0.7 × 0.3 = 0.77]. This value retains the strong signal of "current anomaly" (0.8 × 0.7 = 0.56) and incorporates the trend of "failure risk increasing over time" (0.7 × 0.3 = 0.21), comprehensively reflecting the high probability of equipment failure. In S2's scenario-related features, the dimension value for "Current Temperature Difference Between Threshold and Current Temperature" is 0.6 (indicating a 4.5°C temperature, only 0.5°C away from the upper limit of 5°C). In the time-series-related features, the dimension value for "Hourly Temperature Rise Rate" is 0.9 (indicating a 0.5°C rise per hour). After fusing the weights 0.4 (scenario) and 0.6 (time series), the fused value for this dimension is [0.6 × 0.4 + 0.9 × 0.6 = 0.78]. This value reflects both the critical state of "current temperature approaching the upper limit" (0.6 × 0.4 = 0.24) and the risk of "temperature rising rapidly over time" (0.9 × 0.6 = 0.54), comprehensively reflecting the high risk of temperature exceeding the limit. The server inputs the third-scenario fused features of each scenario into a fully connected module, mapping the high-dimensional abstract features into 0-1 probability values, which serve as monitoring results for temperature control anomalies. This process acts as a "translator," translating the complex features learned by the model into human-understandable "risk probabilities." The third-scenario fusion feature of S1 contains 128-dimensional information such as "abnormal current intensity" and "fault trend rate". The fully connected module compresses the 128-dimensional features into a one-dimensional probability value by learning historical fault data (such as "when these feature combinations appear, there is an 85% probability of a fault"). For example, the server outputs that the failure probability of S1 is 0.88 (88%), which means that "the current device has a very high probability of failure." The third-scenario fusion feature of S2 contains 128-dimensional information such as "current temperature critical value" and "temperature rise rate". The fully connected module learns historical exceeding standard data (such as "when these feature combinations appear, there is a 90% probability that the temperature exceeds the standard") and outputs the exceeding standard probability of S2 as 0.92 (92%), which means that "the temperature is very likely to exceed the strawberry preservation range in the next 2 hours."The server compares the output probability value with a preset threshold (such as 0.7) and triggers a graded warning. If the failure probability of S1 is 0.88 (>0.7), the server sends a "Level 1 Equipment Failure Warning" to the monitoring platform, along with key evidence (such as "Compressor current fluctuates violently, failure probability 88%), prompting the vehicle to immediately stop for maintenance. If the probability of S2 exceeding the standard is 0.92 (>0.7), the server sends a "Level 1 Temperature Exceedance Warning" with a trend prediction (such as "At the current rate, the temperature will reach 6°C in 2 hours"), suggesting switching to refrigeration mode or shortening transportation time. For example, when a refrigerated truck was transporting strawberries, the server, through full-scenario prediction calculations, discovered that the failure probability of S1 was 0.88 and the probability of S2 exceeding the standard was 0.92. The monitoring personnel immediately contacted the driver and stopped the truck 30 minutes in advance for inspection. They discovered that the compressor current was abnormal due to wear. After timely replacement of the component, the temperature returned to normal, preventing the spoilage of 50,000 yuan worth of strawberries. The "on-demand weight allocation" of the third dynamic routing component ensures the differentiated needs of different scenarios for "current status" and "time trend" (for example, fault identification focuses on the current state, while over-standard prediction focuses on the trend); the "feature translation" of the fully connected module converts the "internal language" of the model into an actionable "risk probability". The collaboration between the two enables cold chain temperature control monitoring to move from "data analysis" to "precise decision-making", truly realizing the "early detection and early intervention" of abnormal events. In summary, full-scenario predictive calculation is the "last mile" of cold chain transportation temperature control monitoring by dynamically integrating multi-dimensional features and outputting quantitative results, providing direct technical support for the intelligent management of cold chain logistics.

[0057] In the embodiments of the present invention, the following implementation modes are also provided.

[0058] Obtain multiple initial pre-trained encoders, multiple initial scene processing units corresponding to the monitoring scenes, an initial time series association model, a cold chain temperature control historical training set, and sample temperature control anomaly annotations corresponding to the cold chain temperature control historical training set;

[0059] Coupling the multiple initial pre-trained encoders, the multiple initial scene processing units corresponding to the monitoring scenes, and the initial time series association model into an initial full-scene temperature control perception model;

[0060] According to the cold chain temperature control history training set and the sample temperature control anomaly annotations corresponding to the cold chain temperature control history training set, the initial full-scene temperature control perception model is trained to obtain a target full-scene temperature control perception model; the target full-scene temperature control perception model includes multiple pre-trained encoders, multiple scene processing units corresponding to the monitoring scenes and the time series association model.

[0061] In the embodiment of the present invention, for example, a server of a cold chain logistics enterprise is used as the execution subject, and the complete process of "initial full-scene temperature control perception model training" is described in detail in combination with the historical data of its "refrigerated truck transporting fresh strawberries" scene. The server first needs to prepare the initial model components and historical data required for training: (1) Acquisition of initial model components, initial pre-trained encoder: The server retrieves the encoder pre-trained for cold chain data from the model library, for example: a structured attribute encoder pre-trained based on 100,000 refrigerated truck attribute data (such as model, maintenance record) (the initial parameters are the training results on general attribute data); a time series sensor encoder pre-trained based on 2 million sensor time series data (such as temperature, current) (the initial parameters learn general patterns such as "temperature fluctuation" and "current anomaly"). Initial scene processing unit: an initial processing unit configured for monitoring scenes such as "refrigeration equipment failure" and "temperature continuously exceeding the standard" (such as a network containing a fully connected layer and an attention mechanism, with parameters randomly initialized). Initial time series association model: an initial model containing a time series convolution module and a GRU unit (parameters are randomly initialized and not optimized for scene conduction relationships). (2) Cold chain temperature control historical training set and sample annotation. The server extracts 100,000 cold chain transportation records from the enterprise database in the past three years to form a cold chain temperature control historical training set. Each record contains: main attribute data: such as refrigerated truck A (produced in 2020, Thermo King T-800 refrigeration unit, the expansion valve was replaced in March 2023), refrigerated truck B (produced in 2018, Carrier V-1000 unit, no maintenance record in the past year), etc.; parameter status data: such as the temperature of refrigerated truck A when transporting strawberries on May 10, 2023 The system can record the temperature of refrigerated truck A and the temperature of refrigerated truck B. ...The server couples the initial pre-trained encoder, scene processing unit, and time series association model according to the data flow into an end-to-end initial full-scene temperature control perception model, whose structure is as follows: Input layer: Receives subject attribute data (such as vehicle model, maintenance record) and parameter state data (such as temperature time series, current time series); Feature extraction layer: Extracts subject attribute features and time series state features respectively through structured attribute encoder and time series sensor encoder; Scene fusion layer: The initial scene processing unit of each monitoring scene receives the encoder output and generates scene association features through dynamic routing and attention mechanism; Time series mining layer: The initial time series association model receives the encoder output and generates time series association features through convolution and GRU units; Prediction layer: Fusion of scene association features and time series association features to output the temperature control abnormality probability of each scene. For example, after inputting historical data for refrigerated truck A, the model process follows: [{subject attribute data}rightarrow{structured attribute encoder}rightarrow{scenario processing unit (equipment failure)}], [{temperature time series data}rightarrow{time series sensor encoder}rightarrow{time series association model (equipment failure - temperature exceeding the limit)}], ultimately outputting the "equipment failure probability" and "temperature exceeding the limit probability" through the prediction layer. Based on the historical training set and sample annotations, the server optimizes model parameters through a "forward propagation-loss calculation-backward propagation" cycle until the model converges. Take a historical record of refrigerated truck B (transported on August 15, 2022, and eventually the temperature exceeded the standard due to expansion valve failure) as an example: input the main attribute data ("produced in 2018, Carrier V-1000 unit, no maintenance record in the past year") and parameter status data ("the temperature was 2.1℃~2.5℃ one hour before transportation, the compressor current was 15A~17A, and the current rose to 18A~20A and the temperature rose to 3.0℃~4.5℃ after one hour"); the structured attribute encoder extracts "equipment aging (produced in 2018) )" and "maintenance failure (no maintenance in the past year)" features; the timing sensor encoder extracts the "abnormal increase in current in the later period (15A-20A)" and "accelerated temperature increase in the later period (2.5℃-4.5℃)" features; the scenario processing unit (equipment failure) generates the scenario-related features of "maintenance failure + abnormal current"; the timing association model generates the timing association features of "abnormal current - temperature rise"; the prediction layer outputs "equipment failure probability 0.6" and "temperature exceeding the standard probability 0.5" (the initial model has random parameters and the prediction is inaccurate). The server calculates the deviation (loss value) between the model prediction result and the sample annotation.For example, for a sample record labeled "Probability of device failure 1.0" and "Probability of temperature exceeding the specified value 1.0," the cross-entropy loss between the model's predicted value (0.6, 0.5) and the annotation is: [{loss} = -[1.0×log(0.6)+(1-1.0)×log(1-0.6)]-[1.0×log(0.5)+(1-1.0)×log(1-0.5)]approx0.92] (the larger the loss, the less accurate the model). The server uses the Adam optimizer to backpropagate the gradient of the loss function and adjust the parameters of each model layer (such as the encoder weights, the attention coefficient of the scene processing unit, and the GRU weights of the time series model). For example, if the weight of the "missing maintenance" feature in equipment failure prediction is too low (resulting in a prediction probability of 0.6 < 1.0), the weight of the "maintenance record" dimension in the structured attribute encoder is increased. If the temporal association of "current abnormality-temperature rise" is not captured (resulting in a temperature exceedance probability of 0.5 < 1.0), the convolution kernel weights of the temporal convolution module are adjusted to strengthen the temporal association between current and temperature. The server repeats the above steps (with the batch size set to 64, i.e., training with 64 historical records each time), gradually reducing the loss value. After approximately 500 rounds of training, the average loss of the model on the test set (20,000 historical records not involved in training) dropped below 0.1 (with a deviation tolerance set to 0.15). At this point, the model converged and became the target full-scenario temperature control perception model. The target model has significantly improved performance in practical applications. For the "equipment failure" scenario, the model can identify early signs of compressor wear (such as a slow current increase of 0.5A / hour) two hours in advance, increasing prediction accuracy from an initial 60% to 90%. For the "temperature exceeding standard" scenario, the model can incorporate the temporal transmission of equipment failures (such as "temperature begins to rise one hour after the failure") to provide a 1.5-hour advance warning, increasing prediction accuracy from 55% to 88%. For example, in March 2024, when a refrigerated truck was transporting strawberries, the target model analysis revealed: structured attribute features indicated "equipment manufactured in 2019 and not maintained for nearly a year"; time-series sensor features indicated "compressor current increased from 14A to 15.5A (slowly increasing) in the past hour"; scenario-related features indicated "maintenance loss + abnormal current, high risk of equipment failure"; and time-series correlation features indicated "if the current continues to rise, the temperature will exceed the standard in 1.5 hours." The server triggered an early warning one hour in advance, and the driver, upon inspection, discovered compressor bearing wear and promptly replaced the component, preventing excessive temperatures and spoiling the strawberries. In summary, through historical data training, the initial model was upgraded from a "general feature extractor" to a "scenario-based and time-series precise perception model", providing core technical support for real-time monitoring of temperature control anomalies in cold chain transportation.

[0062] In an embodiment of the present invention, the cold chain temperature control history training set includes sample subject attribute data corresponding to historical transportation subject instances and sample parameter status data corresponding to historical temperature control parameter records; the initial full-scene temperature control perception model is trained based on the cold chain temperature control history training set and the sample temperature control anomaly annotations corresponding to the cold chain temperature control history training set to obtain the target full-scene temperature control perception model, which can be implemented through the following examples.

[0063] Performing preliminary feature extraction on the sample subject attribute data and the sample parameter state data according to the plurality of the initial pre-trained encoders to obtain preliminary coding features output by each of the initial pre-trained encoders;

[0064] Performing scene association fusion on the preliminary coding features output by each of the initial pre-trained encoders according to the multiple initial scene processing units to obtain sample scene association features output by each of the initial scene processing units, and performing time series association feature mining on the preliminary coding features output by each of the initial pre-trained encoders according to the initial time series association model to obtain sample time series association features output by the initial time series association model;

[0065] Predicting based on the sample scene association features corresponding to each of the monitoring scenes to obtain a first temperature control abnormality monitoring result corresponding to each of the monitoring scenes, and predicting based on the sample time series association features corresponding to each of the monitoring scenes to obtain a second temperature control abnormality monitoring result corresponding to each of the monitoring scenes;

[0066] Performing full-scenario prediction calculations based on the sample scene association features and the sample time series association features to obtain a third temperature control anomaly monitoring result corresponding to each monitoring scene;

[0067] performing a deviation calculation based on the first temperature control abnormality monitoring result and the sample temperature control abnormality label to obtain a first abnormal deviation value, performing a deviation calculation based on the second temperature control abnormality monitoring result and the sample temperature control abnormality label to obtain a second abnormal deviation value, and performing a deviation calculation based on the third temperature control abnormality monitoring result and the sample temperature control abnormality label to obtain a third abnormal deviation value;

[0068] determining a model deviation value according to the first abnormal deviation value, the second abnormal deviation value, and the third abnormal deviation value;

[0069] The initial full-scene temperature control perception model is trained according to the model deviation value until the model deviation value does not exceed the deviation tolerance, thereby obtaining a target full-scene temperature control perception model.

[0070] In an embodiment of the present invention, exemplarily, a cold chain logistics enterprise server is used as the execution subject, and combined with its historical data of the "refrigerated truck transporting fresh strawberries" scenario, the complete process of "training the initial full-scene temperature control perception model based on the cold chain temperature control historical training set" is described in detail. The server selects 100,000 cold chain transportation records from 2020 to 2023 as the cold chain temperature control historical training set. Each record contains: sample subject attribute data: such as refrigerated truck ID (such as "LC-001"), vehicle model ("Ice Bear medium-sized refrigerated truck"), refrigeration unit model ("Leng Wang T-800"), year of manufacture (2020), maintenance records in the past year ("expansion valve replaced in March 2023; temperature sensor calibrated in June 2023"), etc.; sample parameter status data: such as the temperature of 6 measuring points in the carriage in a certain transportation task. Time series (collected every 30 seconds, a total of 200 time points, temperature range 1.2℃~6.8℃), compressor current time series (14A~22A), door switch records (3 times, each time 1~5 minutes), etc.; sample temperature control anomaly annotation: abnormal events marked by manual or historical systems, such as "On May 10, 2023, LC-001 caused the temperature to rise to 6℃ (exceeding the standard) 2 hours after transportation due to compressor wear" and "On August 15, 2022, LC-002 failed to maintain the expansion valve in time, and the temperature continued to rise during transportation". The server inputs the sample subject attribute data and sample parameter status data into the initial pre-trained encoder (including the structured attribute encoder and the time series sensor encoder) to extract preliminary encoding features. For example, consider a record for refrigerated truck LC-002 (transported on August 15, 2022, ultimately exceeding the specified temperature limit due to an expansion valve failure). Its primary attribute data includes: "Vehicle Model: Ice Bear Medium; Refrigeration Unit: Carrier V-1000; Year of Manufacture: 2018; Maintenance Record in Recent Year: None." A structured attribute encoder (initial parameters based on pre-training on general attribute data) converts this data into 128-dimensional preliminary encoding features, where "Year of Manufacture: 2018" is encoded as the "Equipment Aging" feature (dimension value 0.7), and "No Maintenance Record in Recent Year" is encoded as the "Maintenance Missing" feature (dimension value 0.8). LC-002's parameter status data includes: One hour before transport, the temperature was 2.1°C to 2.5°C (stable), and the compressor current was 15A to 17A (normal); one hour later, the current rose to 18A to 20A (abnormal), and the temperature rose to 3.0°C to 4.5°C (close to exceeding the specified limit). The time series sensor encoder (initial parameters are pre-trained using general time series data) converts this time series data into 128-dimensional preliminary encoding features. A "late rise in current" is encoded as an "abnormal device load" feature (dimension value 0.6), and a "late rise in temperature" is encoded as a "precursor to temperature control failure" feature (dimension value 0.5). The server uses the initial scene processing unit and initial time series association model to perform scene association fusion and time series association mining on the preliminary encoding features.Taking the initial scenario processing unit for the "refrigeration equipment failure" scenario as an example, it receives preliminary encoded features from the structured attribute encoder and the time series sensor encoder ("equipment aging 0.7," "maintenance loss 0.8," "abnormal equipment load 0.6," and "temperature control failure precursor 0.5"). Through the dynamic routing component (with random initial weights) and the attention mechanism (with random initial parameters), it generates sample scenario association features (128 dimensions) for this scenario. Initially, the model may not accurately capture the association between "lack of maintenance" and "abnormal equipment load," resulting in a low value of 0.4 for the "failure risk" dimension in the features. The initial time series association model receives the time series of preliminary encoded features (e.g., feature changes between the previous hour and the next hour) and uses a time series convolution module (with random initial convolution kernels) and a GRU unit (with random initial weights) to mine the temporal transmission relationship between "abnormal equipment load and temperature rise." Initially, the model may not capture the pattern that "temperature begins to rise one hour after current increases," resulting in a low value of 0.3 for the "failure-exceeding-standard transmission probability" dimension in the time series association features. The server predicts the associated features through three branches and calculates the deviation from the sample label. (1) First branch: Only the scene-related features are used for prediction (the first temperature control anomaly monitoring result). The initial scene processing unit of the "refrigeration equipment failure" scene outputs the sample scene-related features ("failure risk 0.4"), and the failure probability is predicted to be 0.4 through the fully connected layer (the first monitoring result). However, the sample label of this record is "equipment failure probability 1.0", and the deviation (the first abnormal deviation value) between the two is large (such as the cross entropy loss of 0.92). (2) Second branch: Only the time series associated features are used for prediction (the second temperature control anomaly monitoring result). The initial time series associated model outputs the sample time series associated features ("failure-exceeding standard conduction probability 0.3"), and the temperature exceeding standard probability is predicted to be 0.3 through the fully connected layer (the second monitoring result). The sample label is "temperature exceeding standard probability 1.0", and the deviation (the second abnormal deviation value) is also large (such as the cross entropy loss of 1.2). (3) The third branch: Fusion scene and time series feature prediction (the third temperature control anomaly monitoring result). The server weights and fuses the scene-related features ("failure risk 0.4") and the time series-related features ("conduction probability 0.3") (with random initial weights, such as 0.5 each), to generate the third scene fusion feature ("comprehensive risk 0.35"), and predicts a failure probability of 0.35 and an exceeding probability of 0.32 (the third monitoring result). The deviation from the sample annotation (the third anomaly deviation value) further increases (such as a cross entropy loss of 1.3). The server weights and sums the first, second, and third anomaly deviation values ​​(such as weights 0.3:0.3:0.4) to obtain the total model deviation value (such as 1.1), which represents the overall gap between the current model prediction and the actual annotation.The server adjusts the model parameters based on the model deviation value through backpropagation of the Adam optimizer: Adjust the structured attribute encoder: If the weight of the "missing maintenance" feature in fault prediction is too low (resulting in a low scenario-related feature "fault risk 0.4"), increase the encoding weight of this dimension to increase the feature value corresponding to "no maintenance record in the past year" from 0.8 to 0.9; Adjust the scenario processing unit: If the dynamic routing component does not pay attention to the association between "missing maintenance" and "abnormal equipment load", optimize the attention mechanism parameters to increase the attention weight of the "missing maintenance" dimension from 0.3 to 0.6; Adjust the time series association model: If the time series convolution module does not capture the pattern of "temperature rise one hour after current rise", optimize the convolution kernel weight to increase the association weight between "current rise" and "temperature rise one hour after" from 0.2 to 0.7; Adjust the fully connected layer: If the fused prediction probability is too low, optimize the weight of the fully connected layer to increase the prediction probability corresponding to "comprehensive risk 0.35" from 0.35 to 0.6. After 500 rounds of iterative training (64 records per round), the model's deviation gradually decreased to 0.1 (with a deviation tolerance of 0.15), at which point the model converged and became the target full-scenario temperature control perception model. The target model's performance improved significantly in the test set (20,000 historical records not included in training): for the "equipment failure" scenario, the accuracy of the first prediction using only scenario-related features increased from 40% to 85%. For the "temperature exceeding the standard" scenario, the accuracy of the second prediction using only time-series related features increased from 35% to 80%. The fused third prediction achieved an accuracy of 90%, enabling the identification of the complete anomaly chain of "maintenance failure - current anomaly - temperature exceeding the standard" two hours in advance. For example, in March 2024, when a refrigerated truck was transporting strawberries, target model analysis revealed the following: structured attribute features indicated "no maintenance in the past year (dimension value 0.9)" and "equipment aging (dimension value 0.8)"; time series sensor features indicated "compressor current increased from 14A to 15.5A in the past hour (dimension value 0.7)"; scenario-related features indicated "equipment failure risk 0.85"; and time series-related features indicated "temperature will exceed the standard 1.5 hours after the failure (probability 0.9)"; the server triggered an early warning one hour in advance, and the driver discovered worn compressor bearings upon inspection, replacing the parts promptly to prevent the strawberries from spoiling. In summary, through multi-branch prediction and deviation optimization, the initial model was upgraded from "random prediction" to "precise perception," providing core technical support for real-time monitoring of temperature control anomalies in cold chain transportation.

[0071] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned artificial intelligence-based cold chain transportation temperature control monitoring method and system. Figure 2 As shown, Figure 21. A block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To achieve data transmission or interaction, the memory 111, the processor 112, and the communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Based on the above teachings, numerous modifications and variations are possible. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. A cold chain transportation temperature control monitoring method based on artificial intelligence, characterized in that: include: Obtain the subject attribute data corresponding to the transport subject and the parameter status data corresponding to the temperature control parameter; Performing preliminary feature extraction on the subject attribute data and the parameter state data according to a plurality of pre-trained encoders to obtain preliminary coding features output by each of the pre-trained encoders; Each of the pre-trained encoders is pre-trained for different data types; Performing scene association fusion on the preliminary coding features output by each of the pre-trained encoders according to the scene processing units corresponding to the multiple monitoring scenes to obtain scene association features output by each of the scene processing units; Performing time series correlation feature mining on the preliminary coding features output by each pre-trained encoder according to a time series correlation model to obtain time series correlation features output by the time series correlation model; the monitoring scenario is used to characterize temperature control abnormal events occurring in the temperature control parameters of the transport entity; and a time series conduction relationship exists between the plurality of monitoring scenarios; A full-scene prediction calculation is performed based on the scene association features and the time series association features to obtain the temperature control anomaly monitoring results corresponding to each monitoring scene.

2. The method according to claim 1, characterized in that The performing preliminary feature extraction on the subject attribute data and the parameter state data according to the plurality of pre-trained encoders to obtain preliminary coding features output by each of the pre-trained encoders includes: Performing feature transcoding operations on the subject attribute data and the parameter state data respectively according to the feature mapping module to obtain subject attribute features corresponding to the subject attribute data and parameter state features corresponding to the parameter state data; Performing multi-source integration on the subject attribute features and the parameter state features according to a multi-source integration module to obtain a multi-source integrated feature; The multi-source integrated features are input into a plurality of the pre-trained encoders for preliminary feature extraction to obtain preliminary encoding features output by each of the pre-trained encoders.

3. The method according to claim 1 or 2, characterized in that The scene processing units corresponding to the multiple monitoring scenes perform scene association fusion on the preliminary coding features output by each of the pre-trained encoders to obtain scene association features output by each of the scene processing units, including: Determining a first routing weight corresponding to each of the pre-trained encoders according to the first dynamic routing component corresponding to each of the monitoring scenarios, and performing weighted fusion on the preliminary encoding features output by each of the pre-trained encoders according to the first routing weight corresponding to each of the pre-trained encoders to obtain a first scene fusion feature corresponding to each of the monitoring scenarios; The scene-related fusion is performed on each of the first scene fusion features according to the multiple scene processing units to obtain a scene-related feature output by each of the scene processing units.

4. The method according to claim 1 or 2, characterized in that The performing time series correlation feature mining on the preliminary coding features output by each of the pre-trained encoders according to the time series correlation model to obtain the time series correlation features output by the time series correlation model includes: Determining a second routing weight corresponding to each of the pre-trained encoders according to the second dynamic routing component corresponding to each of the monitoring scenarios, and performing weighted fusion on the preliminary encoding features output by each of the pre-trained encoders according to the second routing weight corresponding to each of the pre-trained encoders to obtain a second scene fusion feature corresponding to each of the monitoring scenarios; Performing time series correlation feature mining on each of the second scene fusion features according to the time series correlation model to obtain the time series correlation features output by the time series correlation model.

5. The method according to claim 4, characterized in that The multiple monitoring scenarios include a first monitoring scenario and a second monitoring scenario, and the first monitoring scenario is located before the second monitoring scenario in a sequence having a temporal conduction relationship; The performing time series correlation feature mining on each second scene fusion feature according to the time series correlation model to obtain the time series correlation feature output by the time series correlation model includes: performing time series correlation feature mining on the second scene fusion feature corresponding to the first monitoring scene according to the time series correlation model to obtain the time series correlation feature corresponding to the first monitoring scene, and performing time series correlation feature mining on the second scene fusion feature corresponding to the first monitoring scene and the second scene fusion feature corresponding to the second monitoring scene according to the time series correlation model to obtain the time series correlation feature corresponding to the second monitoring scene; The time series correlation features corresponding to the first monitoring scenario and the time series correlation features corresponding to the second monitoring scenario are determined as the time series correlation features output by the time series correlation model.

6. The method according to claim 4, characterized in that The temporal association model includes at least one temporal association branch, each of which includes a temporal convolution association module and a temporal nonlinear transformation module; performing temporal association feature mining on each second scene fusion feature according to the temporal association model to obtain the temporal association features output by the temporal association model includes: Inputting each second scene fusion feature into the temporal convolution association module to perform temporal correlation calculation to obtain a temporal correlation weight corresponding to each second scene fusion feature; The temporal correlation weight corresponding to each second scene fusion feature is input into the temporal nonlinear transformation module to perform sliding window feature aggregation to obtain the temporal correlation feature output by the temporal correlation model.

7. The method according to claim 1, characterized in that The full-scenario prediction calculation is performed based on the scenario association features and the time series association features to obtain the temperature control abnormality monitoring result corresponding to each monitoring scenario, including: Determine, according to the third dynamic routing component corresponding to each of the monitoring scenarios, the routing weight corresponding to the scenario-related feature and the routing weight corresponding to the timing-related feature, and perform weighted fusion on the scenario-related feature and the timing-related feature based on the routing weight corresponding to the scenario-related feature and the routing weight corresponding to the timing-related feature to obtain a third scenario fusion feature corresponding to each of the monitoring scenarios; According to the fully connected module, the result dimension mapping is performed on the third scene fusion feature corresponding to each monitoring scene to obtain the temperature control abnormality monitoring result corresponding to each monitoring scene.

8. The method according to claim 1, characterized in that The method comprises: Obtain multiple initial pre-trained encoders, multiple initial scene processing units corresponding to the monitoring scenes, an initial time series association model, a cold chain temperature control historical training set, and sample temperature control anomaly annotations corresponding to the cold chain temperature control historical training set; Coupling the multiple initial pre-trained encoders, the multiple initial scene processing units corresponding to the monitoring scenes, and the initial time series association model into an initial full-scene temperature control perception model; According to the cold chain temperature control history training set and the sample temperature control anomaly annotations corresponding to the cold chain temperature control history training set, the initial full-scene temperature control perception model is trained to obtain a target full-scene temperature control perception model; the target full-scene temperature control perception model includes multiple pre-trained encoders, multiple scene processing units corresponding to the monitoring scenes and the time series association model.

9. The method according to claim 8, characterized in that The cold chain temperature control history training set includes sample subject attribute data corresponding to historical transport subject instances and sample parameter status data corresponding to historical temperature control parameter records; the initial full-scene temperature control perception model is trained based on the cold chain temperature control history training set and the sample temperature control anomaly annotations corresponding to the cold chain temperature control history training set to obtain a target full-scene temperature control perception model, including: Performing preliminary feature extraction on the sample subject attribute data and the sample parameter state data according to the plurality of the initial pre-trained encoders to obtain preliminary coding features output by each of the initial pre-trained encoders; Performing scene association fusion on the preliminary coding features output by each of the initial pre-trained encoders according to the multiple initial scene processing units to obtain sample scene association features output by each of the initial scene processing units, and performing time series association feature mining on the preliminary coding features output by each of the initial pre-trained encoders according to the initial time series association model to obtain sample time series association features output by the initial time series association model; Predicting based on the sample scene association features corresponding to each of the monitoring scenes to obtain a first temperature control abnormality monitoring result corresponding to each of the monitoring scenes, and predicting based on the sample time series association features corresponding to each of the monitoring scenes to obtain a second temperature control abnormality monitoring result corresponding to each of the monitoring scenes; Performing full-scenario prediction calculations based on the sample scene association features and the sample time series association features to obtain a third temperature control anomaly monitoring result corresponding to each monitoring scene; performing a deviation calculation based on the first temperature control abnormality monitoring result and the sample temperature control abnormality label to obtain a first abnormal deviation value, performing a deviation calculation based on the second temperature control abnormality monitoring result and the sample temperature control abnormality label to obtain a second abnormal deviation value, and performing a deviation calculation based on the third temperature control abnormality monitoring result and the sample temperature control abnormality label to obtain a third abnormal deviation value; determining a model deviation value according to the first abnormal deviation value, the second abnormal deviation value, and the third abnormal deviation value; The initial full-scene temperature control perception model is trained according to the model deviation value until the model deviation value does not exceed the deviation tolerance, thereby obtaining a target full-scene temperature control perception model.

10. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 9.

Images