Artificial intelligence-based cold chain transportation temperature control monitoring method and system

Abstract

The application discloses a cold chain transportation temperature control monitoring method and system based on artificial intelligence, comprising the following steps: firstly, acquiring attribute data of a transportation subject and state data of temperature control parameters; using an encoder pre-trained for different data classes to extract preliminary coding features; through a processing unit of each monitoring scene, scene correlation fusion is performed on the features to highlight scene-specific information; combining a time sequence correlation model to mine the time sequence conduction relationship between scenes; finally, fusing the scene correlation features and the time sequence correlation features, and outputting temperature control abnormality monitoring results of each scene. Through multi-source data fusion, scene-specific analysis and time sequence conduction mining, the application realizes accurate identification and forward-looking early warning of cold chain temperature control abnormalities.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a cold chain transportation temperature control monitoring method and system based on artificial intelligence. BACKGROUND

[0002] Cold chain transportation is a key link to ensure the quality of perishable goods such as fresh food and medicine, and the stability of its temperature control system directly affects the safety of goods. Traditional cold chain temperature control monitoring relies on single sensor threshold alarm, which has the following shortcomings: only focusing on the real-time value of single parameters such as temperature and current, not fully integrating transportation main body attributes and multi-source state data, resulting in one-sided feature extraction; the monitoring scene is isolated, and the time sequence conduction relationship between abnormal events is not mined, making it difficult to realize forward-looking early warning. SUMMARY

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

[0004] In a first aspect, the present application provides a cold chain transportation temperature control monitoring method based on artificial intelligence, comprising:

[0005] Obtaining main body attribute data corresponding to a transportation main body and parameter state data corresponding to a temperature control parameter;

[0006] Preliminary feature extraction is performed on the main body attribute data and the parameter state data according to a plurality of pre-training encoders, to obtain preliminary encoding features output by each pre-training encoder; each pre-training encoder is pre-trained for different data classes;

[0007] Scene correlation fusion is performed on the preliminary encoding features output by each pre-training encoder according to a scene processing unit corresponding to a plurality of monitoring scenes, to obtain scene correlation features output by each scene processing unit;

[0008] Time sequence correlation feature mining is performed on the preliminary encoding features output by each pre-training encoder according to a time sequence correlation model, to obtain time sequence correlation features output by the time sequence correlation model; the monitoring scene is used to represent a temperature control abnormal event occurring in the temperature control parameter of the transportation main body; there is a time sequence conduction relationship between a plurality of monitoring scenes;

[0009] Full-scene prediction calculation is performed according to the scene correlation features and the time sequence correlation features, to obtain temperature control abnormality monitoring results corresponding to each monitoring scene.

[0010] In a second aspect, the present application provides a server system comprising a server, wherein the server is used to execute the method of the first aspect.

[0011] Compared with the prior art, the application provides the following beneficial effects: by using the cold chain transportation temperature control monitoring method and system based on artificial intelligence, attribute data of a transportation subject and state data of temperature control parameters are obtained; an encoder pre-trained for different data types is used to extract preliminary coding features; scene correlation fusion of the features is performed by a processing unit of each monitoring scene to highlight scene-specific information; a time sequence correlation model is combined to mine the time sequence conduction relationship between scenes; and finally, the scene correlation features and the time sequence correlation features are fused to output temperature control abnormality monitoring results of each scene. Through multi-source data fusion, scene-specific analysis and time sequence conduction mining, the method realizes accurate identification and forward-looking early warning of cold chain temperature control abnormalities. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0013] Figure 1 The step flowchart of the cold chain transportation temperature control monitoring method based on artificial intelligence provided by the embodiments of the present application is shown in the figure.

[0014] Figure 2 The structural schematic block diagram of the computer device provided by the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0016] The specific embodiments of the present application will be described in detail in the following with reference to the drawings.

[0017] In order to solve the technical problems in the foregoing background art, Figure 1 The flowchart of the cold chain transportation temperature control monitoring method based on artificial intelligence provided by the embodiments of the present application is shown in the figure, and the cold chain transportation temperature control monitoring method based on artificial intelligence will be described in detail.

[0018] In step S201, the subject attribute data corresponding to the transportation subject and the parameter state data corresponding to the temperature control parameters are obtained.

[0019] Step S202, according to a plurality of pre-training encoders, the subject attribute data and the parameter state data are preliminarily feature extracted, and the preliminary encoding features output by each pre-training encoder are obtained; each pre-training encoder is pre-trained for different data classes;

[0020] Step S203, according to the scene processing unit corresponding to each of the plurality of monitoring scenes, the preliminary encoding features output by each of the pre-training encoders are scene-associatedly fused, and the scene-associated features output by each of the scene processing units are obtained;

[0021] Step S204, according to a time sequence correlation model, the preliminary encoding features output by each of the pre-training encoders are time sequence correlation feature mined, and the time sequence correlation features output by the time sequence correlation model are obtained; the monitoring scene is used to represent a temperature control abnormal event appearing in the temperature control parameter of the transport subject; there is a time sequence conduction relationship between a plurality of monitoring scenes;

[0022] Step S205, according to the scene-associated features and the time sequence correlation features, full-scene prediction calculation is performed, and the temperature control abnormal monitoring result corresponding to each of the monitoring scenes is obtained.

[0023] In the embodiment of the application, for example, the present embodiment takes a server of a certain fresh cold chain logistics enterprise as an execution subject, combines its actual cold chain transportation scene (such as a fresh strawberry transportation refrigerated vehicle), and specifically describes the specific implementation process of the cold chain transportation temperature control monitoring method based on artificial intelligence.

[0024] As the core of data processing, the server first needs to collect two types of key data from multiple heterogeneous data sources: attribute data of the transport subject and state data of the temperature control parameters. Take the refrigerated truck (transport subject) transporting fresh strawberries as an example: subject attribute data: the server obtains the inherent attribute information of the refrigerated truck through the enterprise resource planning (ERP) system, including vehicle type (such as a medium-sized van refrigerated truck), manufacturer (a well-known cold chain equipment brand), year of manufacture (2020), rated load (2 tons), refrigeration unit model (such as Cool King T-800), historical maintenance records (whether the compressor has been replaced in the past 12 months, whether the temperature sensor has been calibrated), etc. These data reflect the basic ability and reliability of the transport subject, for example, the refrigeration unit model is directly related to its refrigeration efficiency, and the historical maintenance records can indirectly reflect the current operating status of the equipment. Parameter state data: the server collects real-time dynamic data related to temperature control through the Internet of Things (IoT) device on the vehicle, including: temperature sensor data: 6 temperature sensors installed at different positions (such as the front, middle, and rear) of the truck compartment, collecting temperature values every 30 seconds (such as the current temperature at each measuring point is 2.3°C, 2.5°C, and 2.4°C); humidity sensor data: real-time humidity value in the truck compartment (such as 85%); refrigeration equipment operation data: compressor current (such as 15A), voltage (220V), operating mode (such as "strong cooling" or "energy saving"), cumulative running time (such as 5 hours of this transport has been running); environmental interference data: number of times the truck door is opened (such as 3 times during this transport), duration of each opening (such as the first opening for 2 minutes and the second for 3 minutes), external environment temperature (such as the current transport area temperature is 30°C). The server receives the real-time data stream of the IoT device through the MQTT protocol, and retrieves the historical maintenance records and other static attribute data from the enterprise database, and finally integrates the two types of data into a structured input data set for subsequent feature extraction.

[0025] To extract effective features from heterogeneous data, the server employs multiple pre-trained encoders for different data classes to perform preliminary feature extraction on the integrated subject attribute data and parameter state data. First, the server performs format unification and feature conversion on the original data through a feature mapping module: for subject attribute data (such as "refrigeration unit model" "historical maintenance record"), the feature mapping module converts it into numerical features: for example, "cool 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 text type maintenance notes (such as "replace the expansion valve on May 10, 2023"), a word embedding (Word2Vec) model is used to convert it into a low-dimensional vector. For parameter state data (such as temperature time series, compressor current), the feature mapping module arranges the time series data into time windows (such as the last 1 hour of temperature values, a total of 120 time points), and maps discrete state data (such as "strong cold mode") to one-hot encoding. Subsequently, the server integrates subject attribute features (such as vehicle reliability features) and parameter state features (such as temperature fluctuation features) through a multi-source integration module to avoid information loss in a single data dimension. For example, the feature vector corresponding to the vehicle "refrigeration unit model" (reflecting the theoretical refrigeration capacity) is concatenated with the feature vector corresponding to the "average temperature in the last 1 hour" (reflecting the actual refrigeration effect) to form a high-dimensional vector (such as 256 dimensions), forming multi-source integrated features. Next, the server inputs the multi-source integrated features into multiple pre-trained encoders. These encoders have been pre-trained in the training phase for different data classes (such as structured attribute data, time series sensor data) to learn general feature representations: structured attribute encoder: uses a multi-layer perceptron (MLP) structure, and uses a large amount of historical vehicle attribute data (such as 100,000 refrigerated trucks' model, maintenance record) to train in the pre-training phase, to learn "vehicle reliability" "refrigeration capacity" and other implicit features. For example, after inputting the "manufacturer" "year of manufacture" "maintenance record" of a vehicle, a 64-dimensional feature vector is output, where a certain dimension may represent the "probability of refrigeration efficiency decline due to equipment aging". Time series sensor encoder: uses a long short-term memory network (LSTM) or time series convolution network (TCN) structure, and uses millions of temperature, current, and other time series data to train in the pre-training phase, to learn "temperature fluctuation periodicity" "compressor abnormal current pattern" and other features. For example, after inputting the temperature time series (120 points) in the last 1 hour, a 64-dimensional feature vector is output, where a certain dimension may represent the "trend intensity of temperature exceeding the standard". Finally, each pre-trained encoder outputs preliminary encoded features for its data class (such as a 64-dimensional vector output by the structured attribute encoder, and a 64-dimensional vector output by the time series sensor encoder), which retains the core information of the original data and lays a foundation for subsequent scene fusion and time series analysis.

[0026] The temperature control abnormal events (monitoring scenarios) that can occur in cold chain transportation are diverse and targeted, such as "continuous high temperature causing strawberry rot", "refrigeration equipment failure causing temperature rebound", "frequent door opening and closing causing temperature fluctuation", etc. These scenarios need to be analyzed in combination with different data characteristics, so the server configures a scenario processing unit for each monitoring scenario to associate and fuse the preliminary encoding features output by the pre-training encoder. Taking the "refrigeration equipment failure" scenario (the first monitoring scenario) and the "temperature continuously exceeding the upper limit" scenario (the second monitoring scenario) as examples, the two scenarios have a time sequence conduction relationship: equipment failure (the first scenario) can cause temperature to continuously exceed the limit (the second scenario). The fusion process of the scenario processing unit is as follows: First, the server dynamically allocates the weights (first routing weights) of each pre-training encoder according to the scene requirements through the first dynamic routing component corresponding to each monitoring scenario. For example, the "refrigeration equipment failure" scenario needs to focus on analyzing the compressor operation data (time sequence sensor data) and the equipment historical maintenance record (structured attribute data), so the dynamic routing component will allocate a higher weight (such as 0.7) to the time sequence sensor encoder and a lower weight (such as 0.3) to the structured attribute encoder; the "temperature continuously exceeding the upper limit" scenario needs to integrate the temperature time sequence trend (time sequence sensor data) and the vehicle refrigeration capacity (structured attribute data), so the dynamic routing component may allocate balanced weights (such as 0.5 and 0.5) to the two. Subsequently, the server weights and fuses the preliminary encoding features of each pre-training encoder according to the first routing weights to obtain the first scene fusion features of each scenario. For example, the fusion features of the "refrigeration equipment failure" scenario are: [0.7 x {time sequence sensor encoding features} + 0.3 x {structured attribute encoding features}] Finally, the scenario processing unit performs deep association fusion on the first scene fusion features. For example, for the "refrigeration equipment failure" scenario, the processing unit can include an attention mechanism submodule to further highlight key features such as "abnormal fluctuation of compressor current" and "historical compressor replacement frequency", outputting the scene association features (such as a 128-dimensional vector) of this scenario, which reflect information directly related to equipment failure.

[0027] Since there is a time conduction relationship between the monitoring scenes (such as equipment failure-temperature exceeding standard-goods spoilage), the server needs to capture the dependence of features in the time dimension through a time correlation model. The processing process of the time correlation model is as follows: first, the server determines the weight of each pre-training encoder in the time analysis (second routing weight) through the second dynamic routing component corresponding to each monitoring scene. Unlike the scene processing unit, the time correlation model pays more attention to the time continuity of the features, so the weight distribution may be biased towards the time sensor encoder. For example, in the conduction analysis of "equipment failure-temperature exceeding standard", the second routing weight may allocate 0.8 to the time sensor encoder and 0.2 to the structured attribute encoder to highlight the time dependence of time data such as temperature and current. Subsequently, the server weights and fuses the preliminary encoded features according to the second routing weight to obtain the second scene fusion features (such as 128-dimensional vectors) of each scene. Next, the time correlation model performs time correlation mining on the second scene fusion features. In this embodiment, the time correlation model includes a time convolution correlation module and a time nonlinear transformation module: the time convolution correlation module: uses a 1D convolution layer, and the sliding window size is set to 4 (i.e. analyzes the data of the last 4 time steps), and calculates the correlation of adjacent time step features through the convolution kernel. For example, the second scene fusion feature sequence (time steps t-3, t-2, t-1, t) of the "equipment failure" scene is input, and the convolution operation will extract local time patterns such as "t-1 to t time step current abnormal increase" and "t-2 to t-1 time step temperature slight rise", and output the time correlation weight (such as reflecting the importance of a time step to the subsequent scene) of each time step. The time nonlinear transformation module: uses a Gated Recurrent Unit (GRU) or an attention mechanism to aggregate the time correlation weight. For example, the attention mechanism is used to assign higher weights to key time steps (such as the initial time t0 of equipment failure), and finally output a global time correlation feature (such as a 128-dimensional vector) that integrates the time conduction information between scenes (such as "the temperature will continue to rise within 2 hours after the equipment failure occurs"). Taking "equipment failure-temperature exceeding standard" as an example, the time correlation model may output a feature indicating that when the equipment has current abnormalities (premonition of failure) at time t0, the temperature will start to rise at t0+1 hour (t1), and the temperature may exceed the threshold at t0+2 hours (t2) (the preservation range of strawberries is -2℃-5℃), thereby establishing a time conduction relationship between scenes.

[0028] The server finally needs to combine the scene correlation features (reflecting the key factors of the current scene) and the time correlation features (reflecting the time conduction between scenes) to output the temperature control anomaly probability (monitoring result) of each monitoring scene. The specific process is as follows: First, the server dynamically allocates the weights of the scene correlation features and the time correlation features (third routing weight) through the third dynamic routing component corresponding to each monitoring scene. For example: for the "equipment failure" scene (early warning scene), it depends more on the real-time features of the current device state (scene correlation features), so the scene correlation feature weight is set to 0.7, and the time correlation feature weight is set to 0.3; for the "temperature continues to exceed the upper limit" scene (subsequent derivative scene), it depends more on the time cumulative effect (such as the trend of temperature rising after failure), so the time correlation feature weight is set to 0.7, and the scene correlation feature weight is set to 0.3. Then, the server weights and fuses the two types of features according to the third routing weight to obtain the third scene fusion features (such as 256-dimensional vector) of each scene. Finally, the server maps the third scene fusion features through the full connection module to output the temperature control anomaly probability of each monitoring scene. For example: input the third fusion features of the "equipment failure" scene, the full connection layer (output dimension 1) outputs the anomaly probability (such as 0.85, indicating an 85% probability of equipment failure) through the sigmoid activation function; input the third fusion features of the "temperature continues to exceed the upper limit" scene, output the anomaly probability (such as 0.92, indicating a 92% probability of temperature exceeding the standard in the next 2 hours). The server finally compares these probability values with the preset threshold (such as 0.7), and if it exceeds the threshold, it triggers an alarm and notifies the monitoring personnel to take measures (such as checking the refrigeration equipment and adjusting the transportation route).

[0029] To ensure the accuracy of the above method, the server needs to train the model based on historical data. The training data includes: cold chain temperature control history training set: 100,000 transportation records in the past 3 years, each record contains sample subject attribute data (such as vehicle model, maintenance record) and sample parameter state data (such as temperature time series, compressor current); sample temperature control anomaly annotation: manually annotated abnormal events (such as "2022-05-10, a refrigerated truck caused temperature out of standard due to compressor failure"). During the training process, the server couples the initial pre-training encoder, scene processing unit and time series correlation model into an initial full-scene temperature control perception model, and optimizes the model parameters through multi-task learning: model forward propagation: input historical data, calculate preliminary encoding features, scene correlation features and time series correlation features in turn; multi-branch prediction: output abnormal probability based on scene correlation features (first prediction), time series correlation features (second prediction) and fused features (third prediction); bias calculation: calculate the cross-entropy loss of the first prediction and the annotation (first abnormal bias value), the loss of the second prediction (second abnormal bias value), and the loss of the third prediction (third abnormal bias value), and the total loss is the weighted sum of the three (such as 0.3:0.3:0.4); back propagation: adjust the model parameters through the Adam optimizer until the total loss is less than the preset tolerance (such as 0.05), and obtain the target full-scene temperature control perception model. Through the above training, the model can simultaneously optimize the capabilities of scene correlation analysis, time series conduction analysis and full-fusion prediction, ensuring accurate identification of temperature control abnormalities in cold chain transportation in actual application.

[0030] In summary, the method realizes precise monitoring of cold chain transportation temperature control abnormalities through multi-source data fusion, scene-based feature correlation, time series conduction analysis and full-scene prediction, and provides core technical support for intelligent management of cold chain logistics.

[0031] In the embodiment of the present application, the preliminary feature extraction of the subject attribute data and the parameter state data by the plurality of pre-training encoders to obtain the preliminary encoding features output by each pre-training encoder can be implemented by the following example.

[0032] According to the feature mapping module, the feature conversion operation is performed on the subject attribute data and the parameter state data respectively to obtain the subject attribute features corresponding to the subject attribute data and the parameter state features corresponding to the parameter state data;

[0033] According to the multi-source integration module, the subject attribute features and the parameter state features are integrated to obtain multi-source integrated features;

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

[0035] In the embodiment of the present application, the specific implementation process of "preliminary feature extraction of subject attribute data and parameter state data according to multiple pre-training encoders" is described in detail by taking the example of a server of a certain cold chain logistics enterprise processing the real-time temperature control monitoring task of "a certain refrigerated vehicle transporting fresh strawberries".

[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, which is divided into two parts: subject attribute data conversion and parameter state data conversion.(1) Subject attribute data conversion: from discrete description to numerical features. Subject attribute data is static or semi-static data that reflects the inherent properties and historical state of the transport subject (such as a refrigerated truck). For example, the server retrieves the subject attribute data from the enterprise database, including: basic attributes: vehicle type (medium van refrigerated truck), manufacturer ("Ice Bear"), year of manufacture (2020); equipment attributes: refrigeration unit model ("Cold King T-800"), compressor type (scroll); maintenance records: text description (e.g., "Replace expansion valve on March 15, 2023; Calibrate temperature sensor on June 20, 2023"). For these data, the feature mapping module performs the following conversion operations: structured attribute conversion: for category data (e.g., "vehicle type", "manufacturer", "refrigeration unit model"), use a pre-defined encoding table to map to numerical values. For example, "medium van refrigerated truck" corresponds to encoding 002, "Ice Bear" corresponds to encoding 005, and "Cold King T-800" corresponds to encoding 010; for numerical data (e.g., "year of manufacture"), directly normalize (e.g., 2020 is converted to (2020-2018) / (2023-2018) = 0.4, assuming the data range is 2018-2023). Text maintenance record conversion: for the text description of the maintenance record, use a pre-trained word embedding model (e.g., Word2Vec) to convert it to a low-dimensional vector. For example, the word vector corresponding to "replace expansion valve" captures semantic information such as "repair" and "critical component", and finally converts the text of the entire maintenance record to a 32-dimensional vector (by averaging pooling or attention weighted aggregation of word vectors). After conversion, the subject attribute data is converted into subject attribute features (e.g., 128-dimensional vectors), which contain implicit information such as vehicle basic capabilities and equipment reliability.(2) Parameter state data conversion: from time series signal to feature vector. Parameter state data is dynamic data that reflects the real-time running state of the temperature control system. For example, the server collects parameter state data from the vehicle-mounted IoT device in real time, including: temperature time series: temperature values of 6 measuring points in the truck (collected every 30 seconds, 120 time points in the last 1 hour, e.g., [2.1℃, 2.3℃,..., 2.5℃]); equipment running data: compressor current (e.g., [14.8A, 15.2A,..., 15.0A], 120 time points), running mode ("strong cooling" or "energy saving", discrete state); environmental interference data: number of times the door is opened (3 times this time), duration of each opening (e.g., [120s, 180s, 90s]).For these data, the feature mapping module performs the following conversion operations: time series data windowing: continuous time series data such as temperature, current, etc. are intercepted according to fixed time windows (such as 1 hour) to form a two-dimensional matrix (time step x measuring point / parameter). For example, a 6-temperature-measuring-point x 120-time-step matrix (720-dimensional original data) is normalized by sliding window (subtracting the mean and dividing by the standard deviation) to eliminate the dimension effect. Discrete state encoding: for discrete states such as "operation mode", One-Hot encoding is used, such as "strong cooling" corresponding to [1, 0] and "energy saving" corresponding to [0, 1]; for "number of door opening and closing" and "door opening duration", direct normalization processing is performed (such as the maximum number of door opening being 5 times, and the current 3 times being converted to 0.6). After the conversion, the parameter state data is converted into parameter state features (such as 256-dimensional vectors), which contain temperature fluctuation patterns, device operation stability, and environmental disturbance intensity. The subject attribute features and the parameter state features reflect the "inherent ability" and "real-time state" of the transport subject respectively, but single-dimensional features may lose cross information (such as "vehicle refrigeration capacity is strong but current is abnormal"). Therefore, the server integrates the two types of features into multi-source integrated features through a multi-source integration module. The specific integration method is feature splicing + cross-dimensional interaction: feature splicing: the subject attribute features (128-dimensional) and the parameter state features (256-dimensional) are directly spliced according to the dimensions to form an initial fusion vector of 384 dimensions; cross-dimensional interaction: a fully connected layer (such as 384-dimensional input - 256-dimensional output) is used to learn the non-linear correlation between the two types of features. For example, if the "refrigeration unit model" in the subject attribute features has a high dimension value (indicating strong theoretical refrigeration capacity), and the "compressor current" in the parameter state features has an abnormal dimension value (indicating high actual power consumption), the fully connected layer will highlight this contradiction by adjusting the weights, and output a 256-dimensional multi-source integrated feature, where a dimension may represent the "deviation between theoretical capacity and actual running state of the device". Through multi-source integration, the server integrates the scattered subject attributes and parameter state information into a more comprehensive feature representation, providing more abundant input for the subsequent encoder feature extraction. The server inputs the multi-source integrated features into multiple pre-trained encoders, each of which is pre-trained on historical data for different data types (such as structured attributes and time series sensors) to extract more abstract implicit features. Taking two key encoders in this scenario as an example: (1) structured attribute encoder: extracts "subject reliability" features. This encoder uses a multi-layer perceptron (MLP) structure and is trained using 100,000 historical subject attribute data (such as model, maintenance record) of refrigerated vehicles in the pre-training stage to learn "vehicle reliability", "device aging degree", and other implicit features. After inputting the multi-source integrated features (256-dimensional), the encoder compresses the features through 3 fully connected layers (256-128-64) to output 64-dimensional preliminary encoded features.For example, if the word vector corresponding to the "maintenance record" in the input feature contains semantic information of "frequent replacement of the compressor", the encoder can reflect the probability (such as 0.8) of "low reliability of the equipment component" in a certain dimension of the output feature.(2) Time series sensor encoder: extract "state trend" features. This encoder uses a bidirectional LSTM structure, and is trained using one million temperature, current and other time series data in the pre-training stage to learn time series features such as "early signals of temperature exceeding the standard" and "current patterns of equipment failure". After inputting the multi-source integrated features (256 dimensions, including temperature and current time window data), the encoder captures the time dependence (such as the trend of "temperature rising by 0.1℃ every 5 minutes in the past 30 minutes") through the LSTM layer, and outputs 64-dimensional preliminary encoding features through the pooling layer (such as taking the hidden state at the last time). For example, if the time series data of "compressor current" in the input feature shows a "gradual increase" pattern, the encoder can reflect the trend intensity (such as 0.7) of "abnormal increase of compressor load" in a certain dimension of the output feature. The preliminary encoding features output by each pre-trained encoder (such as 64 dimensions output by the structured attribute encoder and 64 dimensions output by the time series sensor encoder) retain the core information of the original data: the former focuses on the inherent ability and potential risks of the transportation subject, and the latter focuses on the real-time state and change trend of the temperature control system. These features will serve as the basis input for subsequent scene correlation fusion and time series analysis, ensuring that the subsequent steps can analyze the possibility of temperature control abnormalities from multiple dimensions and multiple levels. In summary, the server converts heterogeneous data into unified features through the feature mapping module, integrates subject and state information through the multi-source integration module, and extracts implicit features through the pre-trained encoder, completing the conversion from raw data to high-quality features and providing key support for subsequent scene analysis and time series mining.

[0037] In the embodiments of the present application, each of the pre-trained encoders outputs preliminary encoding features, and each of the scene processing units performs scene correlation fusion on the preliminary encoding features output by each of the pre-trained encoders to obtain scene correlation features output by each of the scene processing units. The implementation can be performed through the following examples.

[0038] According to the first dynamic routing component corresponding to each of the monitoring scenes, a first routing weight corresponding to each of the pre-trained encoders is determined, and the preliminary encoding features output by each of the pre-trained encoders are weighted and fused 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 scenes.

[0039] According to the first dynamic routing component corresponding to each of the monitoring scenes, a first routing weight corresponding to each of the pre-trained encoders is determined, and the preliminary encoding features output by each of the pre-trained encoders are weighted and fused 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 scenes.

[0040] In the embodiments of the present application, exemplary, taking the temperature control monitoring task of "a certain refrigerated vehicle transporting fresh strawberries" handled by a certain cold chain logistics enterprise server as an example, combined with two typical monitoring scenarios of "refrigeration equipment failure" (first monitoring scenario) and "temperature persistently exceeding the standard" (second monitoring scenario), the specific implementation process of "scene association 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 allocate weights (first routing weight) of different pre-trained encoders for each monitoring scenario to highlight the core data dimensions that the scenario pays attention to. This process is completed by the first dynamic routing component corresponding to each scene, which is essentially a small neural network (such as a fully connected layer + Softmax activation), the input is the preliminary coding features of the pre-trained encoder, and the output is the routing weight (sum = 1) of each encoder. Taking the "refrigeration equipment failure" scenario as an example, the core of this scenario is to identify whether the refrigeration unit (such as compressor, expansion valve) is abnormal, so it is necessary to analyze the real-time data of the equipment operation (such as time series features of compressor current, voltage) and the historical reliability of the equipment (such as maintenance records, attribute features of unit model). At this time, the first dynamic routing component will calculate the weight of each pre-trained encoder according to the input preliminary coding features: structured attribute encoder (output feature A, 64 dimensions, reflecting the inherent attributes of the vehicle): if it contains "frequent replacement of compressor in historical maintenance records" (corresponding to a dimension value of 0.8), the component will judge that the equipment reliability is low, and the dependence on "theoretical refrigeration capacity" should be reduced, and a lower weight (such as 0.3) is allocated; time series sensor encoder (output feature B, 64 dimensions, reflecting the real-time running state): if it contains "compressor current fluctuation amplitude exceeds threshold" (corresponding to a dimension value of 0.9), the component will judge that real-time data is more critical to fault identification, and a higher weight (such as 0.7) is allocated. For the "temperature persistently exceeding the standard" scenario (which needs to judge whether the temperature will be out of the strawberry preservation range [-2℃~5℃] for a long time), the scenario needs to integrate the refrigeration capacity of the vehicle (structured attribute features) and the trend of temperature change (time series sensor features), so the first dynamic routing component may allocate more balanced weights (such as structured attribute encoder 0.5, time series sensor encoder 0.5). The server weights and sums the preliminary coding features of each pre-trained encoder according to the first routing weight to obtain the first scene fusion features of each scene. Taking the "refrigeration equipment failure" scenario as an example, the fusion formula is: [{first scene fusion features} = 0.3 × {feature A} + 0.7 × {feature B}], where feature A (structured attribute coding feature) contains "equipment historical reliability" information (such as a dimension value of 0.2 indicating "good maintenance record"), and feature B (time series sensor coding feature) contains "current abnormal fluctuation intensity" information (such as a dimension value of 0.8 indicating "severe current fluctuation").After weighting, the corresponding dimension value of the fused feature is: [0.3*0.2+0.7*0.8=0.62], which highlights the key signal of "current anomaly" and retains the reference information of the historical state of the equipment. After obtaining the first scene fusion feature, the server further mines the implicit associated features related to the scene through the scene processing unit corresponding to each monitoring scene. The scene processing unit usually includes attention mechanism, convolution layer or graph neural network modules, and its goal is to filter out the most critical information for the current scene from the fusion feature. (1) Associated fusion of "refrigeration equipment failure" scene. The core of the "refrigeration equipment failure" scene processing unit is to identify early signals of equipment anomalies (such as compressor wear and tear, refrigerant leakage), so its internal structure is designed as "attention mechanism + feature refinement": attention mechanism: calculates 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", "historical compressor replacement frequency" in the feature will be focused on. The calculation of attention weight is based on pre-trained fault samples (such as samples in history that caused failure due to compressor wear and tear, the "current standard deviation" dimension value is generally higher), and finally an attention coefficient is assigned to each dimension (such as "current standard deviation" dimension coefficient 0.9, "historical replacement frequency" dimension coefficient 0.8). Feature refinement: multiply the first scene fusion feature and the attention weight element by element to get the refined feature. For example, if the original value of the "current standard deviation" dimension is 0.7 (indicating that the current fluctuation is large), multiplied by the attention coefficient 0.9, it becomes 0.63, which more directly reflects the "contribution of current anomaly to failure". Finally, the "refrigeration equipment failure" scene processing unit outputs the scene associated feature (such as a 128-dimensional vector), where a certain dimension may explicitly represent the "probability of compressor failure due to wear and tear" (such as 0.85), and another dimension represents the "probability of refrigerant leakage" (such as 0.12), highlighting the key information directly related to equipment failure. (2) Associated fusion of "temperature continuously exceeds threshold" scene. The core of the "temperature continuously exceeds threshold" scene 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), so its internal structure is designed as "time-series-attribute cross analysis + threshold prediction": time-series-attribute cross analysis: cross-calculate "temperature rise rate" (time-series feature) and "vehicle refrigeration capacity" (attribute feature) in the first scene fusion feature. For example, if the "temperature rise rate" is 0.5°C / hour (feature dimension value 0.7), and the "vehicle refrigeration capacity" is 2°C / hour (feature dimension value 0.6), the cross-analysis result is "refrigeration capacity can only offset 50% of the temperature rise" (0.7 / 0.6≈1.17, more than 1 indicates insufficient refrigeration capacity). Threshold prediction: map the cross-analysis result to the "temperature exceeding threshold probability" dimension through a fully connected layer.For example, if the cross-analysis result is 1.17 (insufficient refrigeration capacity), combined with "ambient temperature 30°C" (feature dimension value 0.9) and "door opening times 3 times" (feature dimension value 0.6), the full connection layer outputs the "probability of temperature exceeding the standard in the next 2 hours" (such as 0.92). Finally, the "temperature continues to exceed the standard" scene processing unit outputs the scene association features (such as a 128-dimensional vector), where a certain dimension explicitly indicates the "risk of temperature continuing to exceed the standard due to insufficient refrigeration capacity" (such as 0.92), and another dimension indicates the "environmental interference (opening the door) on the temperature" (such as 0.75), providing targeted features for subsequent time series conduction analysis. Through the above process, the server generates scene association features for each monitoring scene that focus on the needs of the scene. For example, in the "refrigeration equipment failure" scene, the weights of dimensions such as "current anomaly" and "poor historical maintenance" are significantly increased, directly pointing to the root cause of equipment failure; in the "temperature continues to exceed the standard" scene, the dimensions such as "matching degree of refrigeration capacity and temperature rise rate" and "environmental interference intensity" are highlighted, which clearly identifies the driving factors of temperature exceeding the standard. These features not only retain the multi-source information of the original data, but also through the scene-based weight allocation and deep association analysis, convert general features into high-value features specific to the scene, laying a key foundation for subsequent full-scene prediction combined with time series association features.

[0041] In the embodiment of the application, the time series association feature mining is performed on the preliminary encoding features output by each pre-training encoder according to the time series association model to obtain the time series association features output by the time series association model, which comprises:

[0042] The second routing weight value corresponding to each pre-training encoder is determined according to the second dynamic routing component corresponding to each monitoring scene, and the preliminary encoding features output by each pre-training encoder are weighted and fused according to the second routing weight value corresponding to each pre-training encoder to obtain the second scene fusion features corresponding to each monitoring scene;

[0043] The time series association feature mining is performed on each second scene fusion feature according to the time series association model to obtain the time series association features output by the time series association model.

[0044] In this embodiment of the invention, for example, taking the temperature control monitoring task of "a refrigerated truck transporting fresh strawberries" processed by the server of a cold chain logistics enterprise as an example, and combining "refrigeration equipment failure" (first monitoring scenario) and "temperature continuously exceeding the standard" (second monitoring scenario, which has a temporal transmission relationship with the first scenario), the specific implementation process of "mining temporal correlation features based on the temporal correlation model for the preliminary encoded features" is explained in detail. The server needs to dynamically allocate the weights (second routing weights) of the pre-trained encoder for 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, which is essentially a small neural network (such as a fully connected layer + Softmax) trained based on historical temporal patterns. The input is the preliminary encoded feature sequence of the pre-trained encoder (time step t-2, t-1, t), and the output is the weights of each encoder in the temporal analysis (summing up to 1). (1) Calculation of routing weights driven by temporal characteristics. Taking the time-series propagation of "refrigeration equipment failure - continuous temperature exceeding the standard" as an example: - "Refrigeration equipment failure" scenario (first monitoring scenario): The key is to capture early signals of equipment abnormality (such as small fluctuations in compressor current). These signals mainly exist in time-series sensor data (such as time series of current and temperature). Therefore, the second dynamic routing component will assign higher weights to the time-series sensor encoder. For example, if the initial encoded feature sequence of the input time-series sensor encoder (reflecting the current fluctuation from t-2 to t) suddenly increases at t (dimensionality 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 (only referring to the historical reliability of the equipment). "Continuous temperature exceeding the standard" scenario (second monitoring scenario): The key is to analyze the cumulative upward trend of temperature after equipment failure, which requires simultaneous reliance on the temporal continuity of equipment status (such as continuous abnormality of the compressor after failure) and the temporal change of temperature (such as an increase of 0.5℃ per hour). Therefore, the second dynamic routing component will distribute the weights evenly. For example, input the preliminary coding features of the structured attribute encoder (reflecting the equipment's cooling capacity) and the preliminary coding features of the time-series sensor encoder (reflecting the temperature rise rate). If the "cooling capacity" dimension value is 0.6 (medium) and the "temperature rise rate" dimension value is 0.8 (fast), the component is assigned a weight of 0.4 for the structured attribute encoder and a weight of 0.6 for the time-series sensor encoder. (2) Weighted fusion to obtain the second scene fusion features. The server performs weighted summation on the preliminary coding features (time series) of each pre-trained encoder according to the second routing weight to obtain the second scene fusion feature sequence (time steps t-2, t-1, t) for each scene. Taking the "cooling equipment failure" scene as an example, the fusion formula is: [{second scene fusion features} t = 0.2 × {Structured Attribute Encoding Features} t+ 0.8 x {timing sensor encoding feature} t]. Among them, the value of a dimension of the structured attribute encoding feature at time t is 0.4 (indicating that the "device history maintenance is good"), and the value of a dimension of the time sequence sensor encoding feature at time t is 0.7 (indicating that the "current fluctuation is severe"). 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 "current anomaly" fault, while retaining the reference information of the device history state, forming a time sequence fusion feature (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 sequence correlation model, which includes a time sequence convolution correlation module and a time sequence nonlinear transformation module, and the target is to capture the time conduction relationship between scenes (such as how the temperature gradually exceeds the standard over time after the device fault occurs).(1) Time sequence convolution correlation module: calculate the correlation between time steps. This module uses a 1D convolution layer (convolution kernel size 3, step 1) to perform local time sequence pattern extraction on the second scene fusion feature sequence (time window length 3: t-2, t-1, t). For example, the input fusion feature sequence [0.5, 0.6, 0.64] of the "refrigeration device fault" scene, the convolution operation will calculate the difference (such as the increase of 0.04 from t-1 to t) and the pattern (such as continuous rise) between adjacent time steps, and output the time sequence correlation weight (indicating the influence degree of this time step on the subsequent scene) of each time step. Specifically, the weights of the convolution kernel are learned in the training phase through historical fault data: if "current anomaly at time t" is often accompanied by "temperature slightly rising at time t+1" in history, the convolution kernel will assign higher weights to the features at time t. For example, the output time sequence 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 influence on the subsequent temperature exceeding the standard.(2) Time sequence nonlinear transformation module: aggregate time sequence information. This module uses a Gated Recurrent Unit (GRU) or an attention mechanism to globally aggregate the time sequence correlation weight and the fusion feature sequence, and outputs the time sequence correlation feature (reflecting the time conduction law between scenes).Taking "refrigeration equipment failure-temperature continues to exceed" as an example: input the fusion feature sequence [0.5, 0.6, 0.64] of the "refrigeration equipment failure" scene and the time sequence correlation weight [0.2, 0.3, 0.5], the GRU unit will process the data step by step, and record the trend of "current abnormality gradually intensifies from t-2 to t"; at the same time, input the fusion feature sequence (assuming [t-2:0.3, t-1:0.4, t:0.5] of the "temperature continues to exceed" scene, reflecting the temperature gradually rising), the GRU unit will learn the correlation mode of "temperature rising over time after equipment failure"; finally, through the hidden state output (such as a 128-dimensional vector) of the GRU, a certain dimension of the time sequence correlation feature clearly indicates that "1 hour after the equipment failure, the temperature will start to rise" (such as a confidence of 0.8), and another dimension indicates the probability of "temperature exceeding 2 hours after the failure" (such as 0.9). Through the above process, the time sequence correlation feature generated by the server can clearly indicate the time transmission relationship between scenes. For example: if a certain dimension value of the time sequence correlation feature of the "refrigeration equipment failure" scene is 0.8 (indicating the probability of "current equipment failure"), and the corresponding dimension value of the time sequence correlation feature of the "temperature continues to exceed" scene is 0.9 (indicating "temperature exceeding 2 hours after the failure"), the server can trigger an early warning 1 hour in advance, prompting the monitoring personnel to check the equipment or adjust the transportation plan; if the time sequence correlation feature shows that the transmission time of "equipment failure" and "temperature exceeding" is shortened from 2 hours to 1 hour (such as a certain dimension value changes from 0.7 to 0.9), the server can identify a new mode of "equipment aging leading to accelerated failure impact", and optimize the subsequent abnormal prediction strategy. In summary, the time sequence correlation model converts discrete time point features into continuous features reflecting the scene transmission law through dynamic routing weighting and time sequence pattern mining, providing key support for the time dimension of full-scene prediction, and significantly improving the forward-looking and accuracy of cold chain transportation temperature control abnormality monitoring.

[0045] In the embodiment of the application, the plurality of monitoring scenes include a first monitoring scene and a second monitoring scene, and the first monitoring scene is located before the second monitoring scene in a sequence with a time sequence transmission relationship; the time sequence correlation feature mining of each second scene fusion feature according to the time sequence correlation model to obtain the time sequence correlation feature output by the time sequence correlation model can be implemented by the following examples.

[0046] According to the time sequence correlation model, the second scene fusion feature corresponding to the first monitoring scene is subjected to time sequence correlation feature mining to obtain the time sequence correlation feature corresponding to the first monitoring scene, and the second scene fusion feature corresponding to the first monitoring scene and the second scene fusion feature corresponding to the second monitoring scene are subjected to time sequence correlation feature mining according to the time sequence correlation model to obtain the time sequence correlation feature corresponding to the second monitoring scene;

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

[0048] In the embodiments of the present application, an exemplary cold chain logistics enterprise server processes the temperature control monitoring task of "a refrigerated vehicle transporting fresh strawberries" as an example, sets "refrigeration equipment failure" as the first monitoring scene (S1) and "temperature persistently exceeding the standard" as the second monitoring scene (S2), and S1 precedes S2 in the time sequence conduction sequence (i.e., equipment failure may cause temperature to exceed the standard). The following details how the server mines the time sequence correlation features of S1 and S2 through the time sequence correlation model. The server has completed the following through the previous steps: collecting key data of S1 (such as compressor current time sequence, equipment maintenance records) and key data of S2 (such as temperature time sequence, external environment temperature); obtaining the preliminary encoding feature sequence (F1 t-2 ,F1 t-1 ,F1 t ) corresponding to S1 and the preliminary encoding feature sequence (F2 t-2 ,F2 t-1 ,F2 t ) corresponding to S2 through the pre-trained encoder; assigning routing weights to S1 and S2 through the second dynamic routing component (such as S1's time sequence sensor encoder weight 0.8, structured attribute encoder 0.2; S2's time sequence sensor encoder weight 0.6, structured attribute encoder 0.4), obtaining the second scene fusion feature sequence (C1 t-2 ,C1 t-1 ,C1 t ) of S1 and the second scene fusion feature sequence (C2 t-2 ,C2 t-1 ,C2 t). For example, the C1 sequence of S1 can be: [C1=[0.5, 0.6, 0.64]] (corresponding to the fusion feature values at t-2, t-1, t time, the higher the value, the greater the risk of device anomaly); the C2 sequence of S2 can be: [C2=[0.3, 0.4, 0.5]] (corresponding to the fusion feature values at t-2, t-1, t time, the higher the value, the greater the risk of temperature exceeding standard). The server first performs time sequence correlation feature mining on the second scene fusion feature sequence (C1) of S1, and the goal is to identify the time evolution law of S1 itself (such as the time pattern of device failure from inception to manifestation). (1) The time sequence convolution correlation module processes the C1 sequence of S1, and the time sequence convolution correlation module adopts a 1D convolution layer (convolution kernel size 3, step 1), and the sliding window covers 3 time steps (t-2, t-1, t) of C1. The weights of the convolution kernel are learned in the training stage through historical device failure data, for example, if “t-1 time current anomaly” is often accompanied by “t time fault confirmation” in the history, then the convolution kernel will strengthen the association of t-1 to t time steps. In specific calculation, input the C1 sequence [0.5, 0.6, 0.64], 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 mode (such as the continuous rising trend of “0.5-0.6-0.64”), and outputs the time sequence correlation weight (W1=[0.2, 0.3, 0.5]) of each time step, indicating that the feature at t time has the greatest influence on the development of S1 itself (weight 0.5). (2) The time sequence nonlinear transformation module aggregates the time sequence information of S1, and the time sequence nonlinear transformation module adopts a GRU unit, which processes the C1 sequence and the W1 weight step by step, and records the time evolution of S1. For example: t-2 time: C1=0.5 (device anomaly inception), GRU hidden state h-2=tanh(W h *h-3+W x *0.5) (initial h-3=0); t-1 time: C1=0.6 (abnormal aggravation), h-1=tanh(W h *h-2+W x *0.6); t time: C1=0.64 (abnormality is significant), h t =tanh(W h *h-1+W x *0.64). Finally, the GRU outputs the time sequence correlation feature of S1 (H1=h t), for example, one dimension of the 128-dimensional vector explicitly represents the "confirmation probability of device failure at time t" (such as 0.85), and another dimension represents the "time span from the inception to the appearance of the failure" (such as 2 hours). The server further jointly mines the C1 sequence of S1 and the C2 sequence of S2, aiming to capture the timing conduction relationship from S1 to S2 (such as how the temperature exceeds the standard over time after the occurrence of device failure). (1) The timing convolution association module processes the joint sequence, and the server splices the fusion 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 ), and inputs a 1D convolution layer with a kernel size of 6 (covering the entire time window). The weights of the convolution kernel are learned in the training phase through historical conduction data, for example, if "the temperature starts to rise at t+1 hour" and "exceeds the standard at t+2 hour" after "the device failure is confirmed at time t", the convolution kernel will strengthen the association of C1 t and 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], the convolution operation calculates the time correlation across scenarios (such as the potential relationship between C1 t = 0.64 and C2 t+1 ), and outputs the timing correlation weight across scenarios (W2 = [0.1, 0.1, 0.2, 0.2, 0.3, 0.1]), indicating that C1 t (device failure confirmation) has the greatest impact on the subsequent S2 (weight 0.3). (2) The timing nonlinear transformation module aggregates the cross-scenario timing information, and the GRU unit is expanded into a bidirectional structure, simultaneously processing the time sequences of S1 and S2, and learning the conduction patterns of both. For example: forward transmission: from t-2 to t, recording the positive conduction of "device failure aggravation-temperature slow rise"; backward transmission: from t to t-2, verifying the reverse association of "whether the temperature rise is caused by device failure"; and the final hidden state h t' integrates the bidirectional information and outputs the timing association feature of S2 (H2 = h t'), such as a dimension in a 128-dimensional vector representing "1 hour after the device fails, the temperature will start to rise" (e.g., with a confidence of 0.8), and another dimension representing "2 hours after the failure, the probability of temperature exceeding the standard" (e.g., 0.92). The server combines the time sequence association features (H1) of S1 and the time sequence association features (H2) of S2 as the output of the time sequence association model. For example: H1 contains "the probability of confirming the device failure at time t is 0.85" "the time span of failure development is 2 hours"; H2 contains "the probability of temperature starting to rise 1 hour after the failure is 0.8" "the probability of temperature exceeding the standard 2 hours after the failure is 0.92". These features directly reflect the time sequence conduction law between scenes, providing key basis for subsequent full-scene prediction in the time dimension. For example, if H1 shows "the probability of confirming the device failure is 0.85" and H2 shows "the probability of temperature exceeding the standard 2 hours after the failure is 0.92", the server can trigger an early warning 1 hour in advance, prompting the monitoring personnel to check the equipment or adjust the refrigeration parameters 1 hour before t+1 hour, to avoid the strawberries from being rotten due to temperature exceeding the standard. Through the above process, the server not only can identify the time evolution of a single scene (such as how the device failure develops over time), but also can capture the conduction relationship across scenes (such as how the device failure leads to temperature exceeding the standard). The mining of such time sequence association features enables the cold chain temperature monitoring to upgrade from "post-alarm" to "pre-warning", significantly improving the reliability of cold chain transportation and the quality of preserved goods. For example, a refrigerated truck once caused slow rise of current due to hidden wear of the compressor (S1 sprouts), the server identifies "the time span of failure development is 2 hours" through H1, and predicts "temperature exceeding the standard 2 hours after the failure" through H2, and notifies the driver to repair 1.5 hours in advance, avoiding the 50,000 yuan worth of strawberries from being rotten. In summary, the time sequence association model realizes accurate perception of the whole cycle of temperature control abnormal events "occurrence-conduction-evolution" by jointly mining the time sequences of multiple scenes, and provides core technical support for intelligent management of cold chain transportation.

[0049] In the embodiment of the application, the time sequence association model comprises at least one time sequence association branch, each time sequence association branch comprises a time sequence convolution association module and a time sequence nonlinear transformation module; the time sequence association feature mining of each second scene fusion feature according to the time sequence association model comprises:

[0050] inputting each second scene fusion feature into the time sequence convolution association module to perform time sequence correlation calculation, to obtain the time sequence correlation weight corresponding to each second scene fusion feature;

[0051] inputting the time sequence correlation weight corresponding to each second scene fusion feature into the time sequence nonlinear transformation module to perform sliding window feature aggregation, to obtain the time sequence association feature output by the time sequence association model.

[0052] In the embodiments of the present application, an exemplary cold chain logistics enterprise server processes the temperature control monitoring task of "a refrigerated vehicle transporting fresh strawberries" as an example, and in combination with the second scene fusion feature sequence of the two monitoring scenes of "refrigeration equipment failure" (S1) and "temperature persistently exceeding the standard" (S2), the working process of the time sequence association branch (including the time sequence convolution association module and the time sequence nonlinear transformation module) in the time sequence association model is described in detail. The server has completed: collecting key data of S1 (such as compressor current time sequence, equipment maintenance record) and key data of S2 (such as temperature time sequence, external environment temperature); after processing by the pre-training encoder, the preliminary encoding features of S1 and S2 are obtained; through the second dynamic routing component, 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) and 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 standard). The server inputs the second scenario fusion feature sequence of S1 and S2 into the time convolution correlation module, respectively. The core of the 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 of convolution kernel and input processing. The time convolution correlation module adopts a 1D convolution layer, and the convolution kernel size is 3 (covering 3 consecutive time steps: t-2, t-1, t) and the step is 1, ensuring that the local pattern of each time step is fully extracted. The weights of the convolution kernel are learned in the training stage through historical device failure data. For example, if “abnormal fluctuation of current at t-1 time” is often accompanied by “failure confirmation at t time” in the history, the convolution kernel will strengthen the correlation of time steps from t-1 to t. (2) Time correlation calculation. The input (C1 = [0.5, 0.6, 0.64]) of S1 is subjected to sliding dot product operation with the convolution kernel to calculate the local correlation of each time step. Specifically: Window 1 (t-2, t-1, t): the convolution kernel weight is ([w_1, w_2, w_3]) (learned in training, for example, ([0.2, 0.3, 0.5])), and (0.5 x 0.2 + 0.6 x 0.3 + 0.64 x 0.5 = 0.62) is calculated. This value reflects the comprehensive correlation of the time steps in the window, and the final output is the time correlation weight (W1 = [0.2, 0.3, 0.5]) of each time step (the weights sum to 1, and the weight of t time is the highest, indicating that it has the greatest impact on device failure). Similarly, when processing (C2 = [0.3, 0.4, 0.5]) of S2, the convolution kernel weight is adjusted according to the historical temperature exceeding standard data (for example, more attention is paid to the pattern of continuous temperature rise), and the output time correlation weight (W2 = [0.1, 0.2, 0.7]) (the weight of t time is the highest, indicating that the current temperature state has the greatest impact on subsequent exceeding). The server inputs the time correlation weights (W1), (W2) and the corresponding second scenario fusion feature sequences (C1), (C2) into the time nonlinear transformation module, respectively. The module adopts a sliding window + GRU structure, and the goal is to aggregate the correlation of local time steps into global time correlation features. The sliding window size is set to 3 (consistent with the convolution kernel size), ensuring that the complete time evolution period is covered. For example, when processing (C1) and (W1) of S1, the window contains ((c1 t-2 ,w1 t-2 ),(c1 t-1 ,w1 t-1 ),(c1 t ,w1 t)), i.e., ((0.5, 0.2), (0.6, 0.3), (0.64, 0.5)). GRU units process data within a window step-by-step, recording temporal dependencies through update and reset gates: at time t-2: input ((0.5, 0.2)), GRU hidden state (h t-2 = tanh(W h · h t-3 + U x · (0.5 x 0.2)) (initial (h t-3 = 0)); at time t-1: input ((0.6, 0.3)), update gate (z t = sigma(W z · h t-2 + U z · (0.6 x 0.3))), reset gate (r t = sigma(W r · h t-2 + U r · (0.6 x 0.3))), new candidate state (h' t = tanh(W · (r t ⊙ h t-2 ) + U · (0.6 x 0.3))), 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 (h t) is the time sequence correlation feature (such as a 128-dimensional vector) of S1. For example, a certain dimension in the feature represents the "development rate of device failure from t-2 to t" (such as 0.7, indicating that the failure risk increases by 35% per hour), and another dimension represents the "probability of failure confirmation at t" (such as 0.85). Similarly, after processing (C2) and (W2) of S2, the time sequence correlation feature (such as a 128-dimensional vector) of the GRU output of S2 is obtained, wherein a certain dimension represents the "rise rate of temperature from t-2 to t" (such as 0.6, indicating an increase of 0.3°C per hour), and another dimension represents the "probability of temperature exceeding the standard 2 hours after t" (such as 0.92). Through the cooperative work of the time sequence convolution correlation module and the time sequence nonlinear transformation module, the time sequence correlation feature generated by the server can accurately reflect the time evolution law of the temperature control abnormal event: for S1 (device failure), the time sequence correlation feature clearly shows the "rate of increasing failure risk over time" and the "current confirmation probability", helping the server to predict the development trend (such as "the failure probability will rise from 0.5 to 0.85 in the next 1 hour") at the failure germination stage (t-2); for S2 (temperature exceeding the standard), the time sequence correlation feature combines the conduction relationship of S1 (such as "after the device failure is confirmed, the temperature rises by 0.3°C per hour"), and predicts the exceeding time point in advance (such as "the temperature will reach 6°C at t+2 hours, exceeding the strawberry preservation range"). For example, during the transportation process of a refrigerated vehicle, the server discovers through the time sequence correlation branch that the time sequence correlation feature of S1 shows that the "device failure risk reaches 0.85 at t and increases by 35% per hour", and the time sequence correlation feature of S2 shows that "the temperature will exceed the standard 2 hours after the failure (with a probability of 0.92)"; the server immediately triggers an alarm to prompt the driver to check the compressor before t+1 hour, and ultimately avoids the temperature exceeding the standard and the strawberry rotting caused by device failure. In summary, the time sequence correlation branch converts discrete time point data into continuous features reflecting the evolution law of the event through "local correlation calculation + global feature aggregation", providing core technical support for the forward-looking monitoring of temperature control abnormalities in cold chain transportation.

[0053] In the embodiments of the present application, the full-scene prediction calculation according to the scene correlation feature and the time sequence correlation feature to obtain the temperature control abnormality monitoring result corresponding to each monitoring scene can be implemented through the following examples.

[0054] According to the third dynamic routing component corresponding to each monitoring scene, the routing weight value corresponding to the scene correlation feature and the routing weight value corresponding to the time sequence correlation feature are determined, and the scene correlation feature and the time sequence correlation feature are weighted and fused according to the routing weight value corresponding to the scene correlation feature and the routing weight value corresponding to the time sequence correlation feature, to obtain the third scene fusion feature corresponding to each monitoring scene.

[0055] According to the full connection module, the third scene fusion features corresponding to each monitoring scene are subjected to result dimension mapping, so as to obtain a temperature control abnormality monitoring result corresponding to each monitoring scene.

[0056] In the embodiments of the present application, an exemplary cold chain logistics enterprise server processes the temperature control monitoring task of "a refrigerated vehicle transporting fresh strawberries" as an example, combined with two monitoring scenarios of "refrigeration equipment failure" (S1) and "temperature continues to exceed the standard" (S2), and the process of how the server outputs the temperature control abnormal monitoring result by dynamically fusing the scene correlation features and the time sequence correlation features is described in detail. The server has completed the preliminary processing: scene correlation features: through the scene processing unit, the scene correlation features of S1 (refrigeration equipment failure) focus on real-time state information such as "current compressor current fluctuation intensity" and "poor historical maintenance record of the equipment"; the scene correlation features of S2 (temperature continues to exceed the standard) focus on current contradictions such as "the gap between the current temperature and the strawberry preservation threshold" and "the matching degree of refrigeration capacity and temperature rise rate". Time sequence correlation features: through the time sequence correlation model, the time sequence correlation features of S1 reflect the "trend of increasing equipment failure risk over time" (such as "the failure probability increased from 30% to 80% in the past 2 hours"); the time sequence correlation features of S2 reflect the "time law of temperature rise after equipment failure" (such as "temperature began to exceed the standard 1 hour after the failure occurred"). The server configures a third dynamic routing component for each monitoring scenario, the core of which is to dynamically judge the contribution proportion of "current state" (scene correlation features) and "time trend" (time sequence correlation features) to abnormal prediction according to the scene requirements. (1) Weight distribution of S1 (refrigeration equipment failure), the core of S1 is to identify whether the equipment "currently has an abnormality", so it depends more on scene correlation features (reflecting the real-time running state of the equipment). The server analyzes the scene correlation features of S1 and finds that "the current compressor current fluctuation amplitude exceeds 2 times the normal range" (strong real-time signal), and the time sequence correlation features show that "the failure probability increased from 50% to 80% in the past 1 hour" (time trend support). At this time, the dynamic routing component judges that "the current current abnormality" is direct evidence of failure, and assigns a higher weight (such as 0.7) to the scene correlation features and a lower weight (such as 0.3) to the time sequence correlation features, ensuring that the prediction result is more consistent with the real-time state of the equipment. (2) Weight distribution of S2 (temperature continues to exceed the standard), the core of S2 is to predict whether the temperature "will exceed the standard due to time accumulation", so it depends more on time sequence correlation features (reflecting the rising law of temperature after failure). The server analyzes the scene correlation features of S2 and finds that "the current temperature is close to the preservation upper limit (4.5°C)" (current state critical), and the time sequence correlation features show that "after the equipment failure, the temperature rises by 0.5°C per hour" (time trend is clear). At this time, the dynamic routing component judges that "the trend of temperature rising over time" is the key driver of exceeding the standard, and assigns a higher weight (such as 0.6) to the time sequence correlation features and a lower weight (such as 0.4) to the scene correlation features, ensuring that the prediction result pays more attention to the future evolution direction. The server weights and fuses the scene correlation features and the time sequence correlation features according to the dynamically allocated weights to generate the third scene fusion features of each scene.This process is like "scoring different information sources" and combines the "current state" and "time trend" information according to their importance to form a more comprehensive feature representation. In the scenario-related features of S1, the "compressor current abnormal intensity" dimension value is 0.8 (indicating strong fluctuations); in the time-related features, the "fault probability increasing rate over time" dimension value is 0.7 (indicating a 35% increase per hour). After fusion according to weights 0.7 (scenario) and 0.3 (time), the fusion value of this dimension is: [0.8×0.7+0.7×0.3=0.77], which not only retains the strong signal of "current current abnormality" (0.8×0.7=0.56), but also combines the trend of "increasing fault risk over time" (0.7×0.3=0.21), comprehensively reflecting the high possibility of device failure. In the scenario-related features of S2, the "current temperature gap from threshold" dimension value is 0.6 (indicating 4.5℃, only 0.5℃ away from the upper limit of 5℃); in the time-related features, the "temperature increase rate per hour" dimension value is 0.9 (indicating an increase of 0.5℃ per hour). After fusion according to weights 0.4 (scenario) and 0.6 (time), the fusion value of this dimension is: [0.6×0.4+0.9×0.6=0.78], which not only reflects the critical state of "current temperature close to the upper limit" (0.6×0.4=0.24), but also strengthens the risk of "rapid temperature rise over time" (0.9×0.6=0.54), comprehensively reflecting the high risk of temperature exceeding the standard. The server inputs the third scenario fusion features of each scenario into the fully connected module, maps the high-dimensional abstract features to probability values between 0 and 1, and outputs the monitoring results of the temperature control anomaly. This process is like a "translator" that converts the complex features learned by the model into "risk probability" that humans can understand. The third scenario fusion features of S1 contain 128-dimensional information such as "current abnormal intensity" and "fault trend rate". The fully connected module learns historical fault data (such as "when these features appear, there is an 85% probability of failure") to compress the 128-dimensional features into a 1-dimensional probability value. For example, the server outputs the failure probability of S1 as 0.88 (88%), indicating that "the current device has a very high probability of failure". The third scenario fusion features of S2 contain 128-dimensional information such as "current temperature threshold" and "temperature increase rate". The fully connected module learns historical over-standard data (such as "when these features appear, there is a 90% probability of temperature exceeding the standard") to output the over-standard probability of S2 as 0.92 (92%), indicating that "the temperature is likely to exceed the strawberry preservation range within the next 2 hours".The server compares the output probability value with a preset threshold (such as 0.7) to trigger a hierarchical early warning: if the failure probability of S1 is 0.88 (> 0.7), the server pushes a "first-level equipment failure early warning" to the monitoring platform, with key evidence (such as "the compressor current fluctuates violently, with a failure probability of 88%"), prompting immediate shutdown for maintenance; if the exceeding probability of S2 is 0.92 (> 0.7), the server pushes a "first-level temperature exceeding warning", with trend prediction (such as "at the current rate, the temperature will reach 6℃ in 2 hours"), suggesting switching to a refrigeration mode or shortening the transportation time. For example, when a refrigerated truck is transporting strawberries, the server calculates through full-scene prediction that the failure probability of S1 is 0.88 and the exceeding probability of S2 is 0.92; the monitoring personnel immediately contact the driver and stop the truck 30 minutes in advance for inspection, find that the compressor current is abnormal due to wear and tear, and replace the parts in time, so that the temperature returns to normal, avoiding the corruption of 50,000 yuan of strawberries. The "on-demand allocation of weights" of the third dynamic routing component ensures the differentiated needs of different scenes for "current state" and "time trend" (such as failure identification for current and exceeding prediction for trend); the "feature translation" of the full connection module converts the "internal language" of the model into an operable "risk probability". The synergy of the two makes the cold chain temperature control monitoring from "data analysis" to "precise decision-making", and truly realizes "early detection and early intervention" of abnormal events. In summary, the full-scene prediction calculation dynamically integrates multi-dimensional features and outputs quantitative results, which is the "last mile" of cold chain temperature control monitoring and provides direct technical support for intelligent management of cold chain logistics.

[0057] In the embodiments of the present application, the following implementation modes are further provided.

[0058] Obtain a plurality of initial pre-training encoders, a plurality of initial scene processing units corresponding to the monitoring scenes, an initial time sequence correlation model, a cold chain temperature control historical training set, and sample temperature control abnormality annotations corresponding to the cold chain temperature control historical training set;

[0059] Couple the plurality of initial pre-training encoders, the plurality of initial scene processing units corresponding to the monitoring scenes, and the initial time sequence correlation model into an initial full-scene temperature control perception model;

[0060] Train the initial full-scene temperature control perception model according to the cold chain temperature control historical training set and the sample temperature control abnormality annotations corresponding to the cold chain temperature control historical training set to obtain a target full-scene temperature control perception model; the target full-scene temperature control perception model includes a plurality of pre-training encoders, a plurality of scene processing units corresponding to the monitoring scenes, and a time sequence correlation model.

[0061] In the embodiments of the present application, an exemplary cold chain logistics enterprise server is taken as the execution subject, combined with its historical data of the "fresh strawberry transportation by refrigerated truck" scene, and the complete process of "initial full-scene temperature control perception model training" is described in detail. The server first needs to prepare the initial model components and historical data required for training: (1) obtaining the initial model components, the initial pre-trained encoder: the server retrieves the pre-trained encoder for cold chain data from the model library, for example: a structured attribute encoder based on 100,000 pieces of refrigerated truck attribute data (such as model, maintenance record) pre-trained (the initial parameters are the training results on general attribute data); a time series sensor encoder based on 2 million pieces of sensor time series data (such as temperature, current) pre-trained (the initial parameters learn general patterns such as "temperature fluctuation" and "current anomaly"). Initial scene processing unit: initial processing unit configured for monitoring scenes such as "refrigeration equipment failure" and "temperature persistently exceeding the standard" (such as a network containing a fully connected layer and an attention mechanism, with randomly initialized parameters). Initial time series association model: an initial model containing a time series convolution module and a GRU unit (with randomly initialized parameters, not optimized for scene conduction relationship). (2) Cold chain temperature control historical training set and sample annotation: the server extracts 100,000 pieces of cold chain transportation records from the enterprise database in the past 3 years to form a cold chain temperature control historical training set, each record containing: subject attribute data: such as refrigerated truck A (produced in 2020, cold king T-800 refrigeration unit, replaced the expansion valve in March 2023), refrigerated truck B (produced in 2018, Carrier V-1000 unit, no maintenance record for nearly 1 year) and the like; parameter state data: such as the temperature time series of refrigerated truck A when transporting strawberries on May 10, 2023 (recorded every 30 seconds, a total of 200 time points, temperature range 1.2℃-3.5℃), compressor current (14A-16A), and vehicle door opening and closing record (3 times, each time 1-2 minutes) and the like; sample temperature control anomaly annotation: abnormal events marked by artificial or historical system, such as "on May 10, 2023, refrigerated truck A caused temperature to rise to 6℃ (exceeding the standard) due to compressor wear during transportation" and "on August 15, 2022, refrigerated truck B failed to maintain due to expansion valve failure, and the temperature continued to rise during transportation".The server couples the initial pre-training encoder, scene processing unit, and time series correlation model into an initial full-scene temperature control perception model according to the data flow, and the structure is as follows: the input layer receives the subject attribute data (such as vehicle model, maintenance record) and parameter state data (such as temperature time series, current time series); the feature extraction layer extracts the subject attribute features and time series state features through the structured attribute encoder and the time series sensor encoder; the scene fusion layer receives the encoder output of each initial scene processing unit in the monitoring scene, and generates scene correlation features through dynamic routing and attention mechanism; the time series mining layer receives the encoder output of the initial time series correlation model, and generates time series correlation features through convolution and GRU unit; the prediction layer fuses the scene correlation features and the time series correlation features, and outputs the temperature control anomaly probability of each scene. For example, after the historical data of refrigerated truck A is input, the model flow is: [{subject attribute data}rightarrow{structured attribute encoder}rightarrow{scene processing unit (device failure)}], [{temperature time series data}rightarrow{time series sensor encoder}rightarrow{time series correlation model (device failure-temperature exceeding standard)}], and finally the prediction layer outputs the "device failure probability" and "temperature exceeding standard probability". Based on the historical training set and sample labeling, the server optimizes the model parameters through "forward propagation-loss calculation-back propagation" cycle until the model converges. Taking a historical record of refrigerated truck B (transportation on August 15, 2022, and finally caused temperature exceeding standard due to expansion valve failure) as an example: input the subject attribute data ("produced in 2018, Kelv V-1000 unit, no maintenance record in the past year") and parameter state data ("1 hour before transportation, temperature 2.1℃-2.5℃, compressor current 15A-17A, 1 hour later, current rises to 18A-20A, temperature rises to 3.0℃-4.5℃"); the structured attribute encoder extracts the "device aging (produced in 2018)" and "maintenance absence (no maintenance in the past year)" features; the time series sensor encoder extracts the "current late abnormal rise (15A-20A)" and "temperature late accelerated rise (2.5℃-4.5℃)" features; the scene processing unit (device failure) generates the scene correlation features of "maintenance absence + current anomaly"; the time series correlation model generates the time series correlation features of "current anomaly-temperature rise"; the prediction layer outputs "device failure probability 0.6" and "temperature exceeding standard probability 0.5" (the initial model is not accurate due to random parameters). The server calculates the deviation (loss value) between the model prediction result and the sample labeling.For example, the sample record is labeled as "Device Failure Probability 1.0" and "Temperature Exceedance Probability 1.0". The cross-entropy loss between the model's predicted value (0.6, 0.5) and the label 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)] approx 0.92] (the larger the loss, the less accurate the model). The server adjusts the parameters of each layer of the model (such as the weights of the encoder, the attention coefficients of the scene processing unit, and the GRU weights of the time series model) by backpropagating the gradient along the loss function using the Adam optimizer. For example, if the "maintenance missing" feature has too low a weight in the device failure prediction (resulting in a predicted probability of 0.6 < 1.0), the weight of the "maintenance record" dimension in the structured attribute encoder is increased. If the time series association of "current anomaly - temperature rise" is not captured (resulting in a temperature exceedance probability of 0.5 < 1.0), the convolution kernel weights of the time series convolution module are adjusted to strengthen the temporal association between current and temperature. The server repeats the above steps (batch size set to 64, i.e., 64 historical records are used for training each time) to gradually reduce the loss value. After about 500 rounds of training, the average loss of the model on the test set (20,000 historical records not involved in training) is reduced to below 0.1 (with a bias tolerance of 0.15), at which point the model converges and becomes the target full-scenario temperature control perception model. The target model performs significantly better in practical applications: for the "device failure" scenario, the model can identify early signs of compressor wear (such as a slow rise in current of 0.5 A / hour) 2 hours in advance, and the prediction accuracy is improved from 60% to 90%; for the "temperature exceedance" scenario, the model can combine the time series conduction of device failure (such as "1 hour after failure, temperature starts to rise"), and give an early warning 1.5 hours in advance, with the prediction accuracy improved from 55% to 88%. For example, when a refrigerated truck is transporting strawberries in March 2024, the model analyzes the data using the target model and finds that the structured attribute features show "device produced in 2019, no maintenance for nearly 1 year"; the time series sensor features show "compressor current has risen from 14 A to 15.5 A (slowly rising) in the past 1 hour"; the scene association features indicate "maintenance missing + current anomaly, high risk of device failure"; the time series association features indicate "if the current continues to rise, the temperature will exceed the standard in 1.5 hours"; the server triggers an early warning 1 hour in advance, and the driver checks and finds that the compressor bearing is worn out, and replaces the parts in time, preventing temperature exceedance and strawberry spoilage. In summary, through historical data training, the initial model is upgraded from a "general feature extractor" to a "scenario-based and time-series-based precision perception model", providing core technical support for real-time monitoring of temperature control abnormalities in cold chain transportation.

[0062] In the embodiment of the present application, the cold chain temperature control history training set includes sample main body attribute data corresponding to the historical transportation main body instance and sample parameter state data corresponding to the historical temperature control parameter record; the initial full-scene temperature control perception model is trained according to the cold chain temperature control history training set and the sample temperature control abnormality annotation corresponding to the cold chain temperature control history training set, and a target full-scene temperature control perception model is obtained. The implementation can be performed through the following examples.

[0063] The sample main body attribute data and the sample parameter state data are preliminarily feature extracted according to the plurality of initial pre-training encoders, and preliminary coding features output by each initial pre-training encoder are obtained.

[0064] The preliminary coding features output by each initial pre-training encoder are scene-relatedly fused according to the plurality of initial scene processing units, and sample scene-related features output by each initial scene processing unit are obtained, and the preliminary coding features output by each initial pre-training encoder are time-series-related feature mined according to the initial time-series-related model, and sample time-series-related features output by the initial time-series-related model are obtained.

[0065] Each monitoring scene is predicted according to the sample scene-related features corresponding to the monitoring scene, and a first temperature control abnormality monitoring result corresponding to each monitoring scene is obtained, and each monitoring scene is predicted according to the sample time-series-related features corresponding to the monitoring scene, and a second temperature control abnormality monitoring result corresponding to each monitoring scene is obtained.

[0066] Each monitoring scene is predicted according to the sample scene-related features and the sample time-series-related features, and a third temperature control abnormality monitoring result corresponding to each monitoring scene is obtained.

[0067] A first abnormality deviation value is obtained by performing deviation calculation according to the first temperature control abnormality monitoring result and the sample temperature control abnormality annotation, a second abnormality deviation value is obtained by performing deviation calculation according to the second temperature control abnormality monitoring result and the sample temperature control abnormality annotation, and a third abnormality deviation value is obtained by performing deviation calculation according to the third temperature control abnormality monitoring result and the sample temperature control abnormality annotation.

[0068] A model deviation value is determined according to the first abnormality deviation value, the second abnormality deviation value and the third abnormality 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 a deviation tolerance, and a target full-scene temperature control perception model is obtained.

[0070] In the embodiments of the present application, an exemplary cold chain logistics enterprise server is taken as the execution subject, combined with the historical data of its "fresh strawberry transportation by refrigerated truck" scene, and the complete process of "training an initial full-scene temperature control perception model based on 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, and each record contains: sample subject attribute data such as refrigerated truck ID (e.g. "LC-001"), vehicle model ("Ice Bear medium-sized refrigerated truck"), refrigeration unit model ("Cold King T-800"), year of manufacture (2020), and maintenance records in the past year ("replace the expansion valve in March 2023; calibrate the temperature sensor in June 2023"), etc.; sample parameter state data such as the temperature time series of 6 measuring points in a certain transportation task (collecting every 30 seconds, a total of 200 time points, temperature range 1.2℃-6.8℃), compressor current time series (14A-22A), and vehicle door opening and closing records (3 times, each time 1-5 minutes), etc.; sample temperature control abnormality annotation: abnormal events marked by artificial or historical system, such as "LC-001's temperature rose to 6℃ (over standard) after 2 hours of transportation on May 10, 2023 due to compressor wear" and "LC-002's temperature continued to rise during transportation on August 15, 2022 due to expansion valve failure without timely maintenance", etc. The server inputs the sample subject attribute data and sample parameter state data into the initial pre-training encoder (including structured attribute encoder and time series sensor encoder) to extract preliminary encoding features. Taking a record of refrigerated truck LC-002 as an example (transportation on August 15, 2022, and finally temperature over standard due to expansion valve failure), its subject attribute data is: "vehicle model: Ice Bear medium-sized; refrigeration unit: Carrier V-1000; year of manufacture: 2018; maintenance records in the past year: none". The structured attribute encoder (initial parameters are the pre-training results of general attribute data) converts these data into 128-dimensional preliminary encoding features, including "equipment aging" feature (dimension value 0.7) for "year of manufacture 2018" and "maintenance missing" feature (dimension value 0.8) for "no maintenance records in the past year". The parameter state data of LC-002 is: temperature 2.1℃-2.5℃ (stable) and compressor current 15A-17A (normal) 1 hour before transportation; current rises to 18A-20A (abnormal) and temperature rises to 3.0℃-4.5℃ (close to over standard) 1 hour later. The time series sensor encoder (initial parameters are the pre-training results of general time series data) converts these time series data into 128-dimensional preliminary encoding features, including "equipment load abnormality" feature (dimension value 0.6) for "current late rise" and "temperature control failure precursor" feature (dimension value 0.5) for "temperature late rise". The server performs scene association fusion and time series association mining on the preliminary encoding features through the initial scene processing unit and the initial time series association model, respectively.Taking the initial scene processing unit of the "refrigeration equipment failure" scenario as an example, it receives the preliminary encoded features of the structured attribute encoder and the time sequence sensor encoder ("equipment aging 0.7", "maintenance absence 0.8", "equipment load anomaly 0.6", "temperature control failure precursor 0.5"), generates sample scene association features (128 dimensions) of the scenario through a dynamic routing component (initial weights are random) and an attention mechanism (initial parameters are random). In the initial stage, the model may not accurately capture the association between "maintenance absence" and "equipment load anomaly", so the value of the "failure risk" dimension in the feature is only 0.4 (low). The initial time sequence association model receives the time sequence of the preliminary encoded features (such as the feature changes in the previous 1 hour and the next 1 hour), and mines the time sequence conduction relationship of "equipment load anomaly-temperature rise" through a time sequence convolution module (initial convolution kernel is random) and a GRU unit (initial weights are random). In the initial stage, the model may not capture the rule that "the temperature starts to rise 1 hour after the current rises", so the value of the "failure-over-standard conduction probability" dimension in the time sequence association feature is only 0.3 (low). The server predicts the association features through three branches and calculates the deviation from the sample label. (1) The first branch: only use scene association features to predict (the first temperature control anomaly monitoring result), the initial scene processing unit of the "refrigeration equipment failure" scenario outputs the sample scene association features ("failure risk 0.4"), and predicts the failure probability to be 0.4 through the full connection layer (the first monitoring result). The sample label of this record is "equipment failure probability 1.0", and the deviation (the first abnormal deviation value) is large (such as cross-entropy loss 0.92).(2) The second branch: only use time sequence association features to predict (the second temperature control anomaly monitoring result), the initial time sequence association model outputs the sample time sequence association features ("failure-over-standard conduction probability 0.3"), and predicts the temperature over-standard probability to be 0.3 through the full connection layer (the second monitoring result). The sample label is "temperature over-standard probability 1.0", and the deviation (the second abnormal deviation value) is also large (such as cross-entropy loss 1.2).(3) The third branch: fusion of scene and time sequence features to predict (the third temperature control anomaly monitoring result), the server weights and fuses the scene association features ("failure risk 0.4") and the time sequence association features ("conduction probability 0.3") (the initial weights are random, such as 0.5 each) to generate the third scene fusion features ("comprehensive risk 0.35"), and predict the failure probability to be 0.35 and the over-standard probability to be 0.32 (the third monitoring result). The deviation (the third abnormal deviation value) from the sample label is further increased (such as cross-entropy loss 1.3). The server weights and sums the first, second, and third abnormal deviation values (such as weights 0.3:0.3:0.4), to obtain the model total deviation value (such as 1.1), which represents the overall gap between the current prediction of the model and the actual label.The server adjusts the model parameters by the Adam optimizer based on the model deviation value: adjust the structured attribute encoder: if the "maintenance absence" feature has too low weight in fault prediction (leading to the scene correlation feature "fault risk 0.4" being too low), increase the encoding weight of this dimension, so that the feature value corresponding to "no maintenance record in the past 1 year" is increased from 0.8 to 0.9; adjust the scene processing unit: if the dynamic routing component does not pay attention to the correlation between "maintenance absence" and "abnormal device load", optimize the attention mechanism parameters, so that the attention weight of the "maintenance absence" dimension is increased from 0.3 to 0.6; adjust the time series correlation model: if the time series convolution module does not capture the rule of "temperature rise after 1 hour of current rise", optimize the convolution kernel weight, so that the correlation weight of "current rise" and "temperature rise after 1 hour" is increased from 0.2 to 0.7; adjust the fully connected layer: if the fused prediction probability is low, optimize the weight of the fully connected layer, so that the prediction probability corresponding to "comprehensive risk 0.35" is increased from 0.35 to 0.6. After 500 rounds of iterative training (64 records per round), the model deviation value gradually decreases to 0.1 (the deviation tolerance is set to 0.15), at which point the model converges and becomes the target full-scene temperature control perception model. The target model performs significantly better on the test set (20,000 historical records not involved in training): for the "device fault" scene, the first prediction accuracy using only the scene correlation feature is increased from 40% to 85%; for the "temperature exceeding standard" scene, the second prediction accuracy using only the time series correlation feature is increased from 35% to 80%; the third prediction accuracy after fusion is as high as 90%, which can identify the complete abnormal chain of "maintenance absence - current anomaly - temperature exceeding standard" 2 hours in advance. For example, in March 2024, when a refrigerated truck was transporting strawberries, the target model analyzed that the structured attribute features showed "no maintenance in the past 1 year (dimension value 0.9)" and "device aging (dimension value 0.8)"; the time series sensor features showed "compressor current increased from 14A to 15.5A in the past 1 hour (dimension value 0.7)"; the scene correlation feature prompted "device fault risk 0.85"; the time series correlation feature prompted "temperature will exceed standard 1.5 hours after the fault (probability 0.9)"; the server triggered an early warning 1 hour in advance, and the driver found that the compressor bearing was worn out after checking, and replaced the parts in time, avoiding the strawberries from rotting. In summary, through multi-branch prediction and deviation optimization, the initial model is upgraded from "random prediction" to "accurate perception", providing a core technical guarantee for real-time monitoring of cold chain transportation temperature control anomalies.

[0071] The embodiment of the present application provides a computer device 100, which comprises a processor and a non-volatile memory storing computer instructions, and when the computer instructions are executed by the processor, the computer device 100 executes the aforementioned cold chain transportation temperature control monitoring method and system based on artificial intelligence. As shown in Figure 2 Figure 2 ​A structural block diagram of the computer device 100 is provided for the embodiments of the present application. The computer device 100 comprises a memory 111, a processor 112 and a communication unit 113. For realizing the transmission or interaction of data, the memory 111, the processor 112 and the communication unit 113 are electrically connected with each other directly or indirectly. For example, the electrical connection between these elements can be realized through one or more communication buses or signal lines. For the purpose of illustration, 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 disclosure to the precise forms disclosed. Many modifications and variations are possible in light of the above teachings. The embodiments are chosen and described in order to best explain the principles of the disclosure and its practical application to thereby enable others skilled in the art to best utilize the disclosure, and various embodiments with various modifications as are suited to the particular use contemplated.

Claims

1. An artificial intelligence-based cold chain transportation temperature control monitoring method, characterized in that, The method comprises: acquiring subject attribute data corresponding to a transportation subject and parameter state data corresponding to a temperature control parameter; preliminary feature extraction is performed on the subject attribute data and the parameter state data by using a plurality of pre-trained encoders to obtain preliminary encoding features output by each pre-trained encoder; each pre-trained encoder is pre-trained for different data categories; scene correlation fusion is performed on the preliminary encoding features output by each pre-trained encoder by using a plurality of scene processing units corresponding to monitoring scenes to obtain scene correlation features output by each scene processing unit, including: determining a first routing weight corresponding to each pre-trained encoder according to a first dynamic routing component corresponding to each monitoring scene, and performing weighted fusion on the preliminary encoding features output by each pre-trained encoder according to the first routing weight corresponding to each pre-trained encoder to obtain a first scene fusion feature corresponding to each monitoring scene; scene correlation fusion is performed on each first scene fusion feature by using a plurality of scene processing units to obtain scene correlation features output by each scene processing unit; time sequence correlation feature mining is performed on the preliminary encoding features output by each pre-trained encoder by using a time sequence correlation model to obtain time sequence correlation features output by the time sequence correlation model, including: determining a second routing weight corresponding to each pre-trained encoder according to a second dynamic routing component corresponding to each monitoring scene, and performing weighted fusion on the preliminary encoding features output by each pre-trained encoder according to the second routing weight corresponding to each pre-trained encoder to obtain a second scene fusion feature corresponding to each monitoring scene; time sequence correlation feature mining is performed on each second scene fusion feature by using the time sequence correlation model to obtain time sequence correlation features output by the time sequence correlation model; the monitoring scene is used to represent a temperature control abnormal event occurring in the temperature control parameter of the transportation subject; there is a time sequence conduction relationship between a plurality of monitoring scenes; full-scene prediction calculation is performed according to the scene correlation features and the time sequence correlation features to obtain a temperature control abnormality monitoring result corresponding to each monitoring scene.

2. The method of claim 1, wherein, The preliminary feature extraction performed on the subject attribute data and the parameter state data by using a plurality of pre-trained encoders to obtain preliminary encoding features output by each pre-trained encoder comprises: feature conversion operations are respectively performed on the subject attribute data and the parameter state data by using a feature mapping module to obtain subject attribute features corresponding to the subject attribute data and parameter state features corresponding to the parameter state data; multi-source integration is performed on the subject attribute features and the parameter state features by using a multi-source integration module to obtain multi-source integration features; the multi-source integration features are input into a plurality of pre-trained encoders for preliminary feature extraction to obtain preliminary encoding features output by each pre-trained encoder.

3. The method of claim 1, wherein, The plurality of monitoring scenes include a first monitoring scene and a second monitoring scene, and the first monitoring scene is located before the second monitoring scene in a sequence in which there is a time sequence conduction relationship. The time sequence correlation feature mining according to the time sequence correlation model on each second scene fusion feature comprises: The time sequence correlation feature mining according to the time sequence correlation model on the second scene fusion feature corresponding to the first monitoring scene and the second scene fusion feature corresponding to the second monitoring scene comprises: The time sequence correlation feature mining according to the time sequence correlation model on the second scene fusion feature corresponding to the first monitoring scene and the second scene fusion feature corresponding to the second monitoring scene comprises:

4. The method of claim 1, wherein, The time sequence correlation feature mining according to the time sequence correlation model on the second scene fusion feature corresponding to the first monitoring scene and the second scene fusion feature corresponding to the second monitoring scene comprises: The time sequence correlation feature mining according to the time sequence correlation model on the second scene fusion feature corresponding to the first monitoring scene and the second scene fusion feature corresponding to the second monitoring scene comprises: The time sequence correlation feature mining according to the time sequence correlation model on the second scene fusion feature corresponding to the first monitoring scene and the second scene fusion feature corresponding to the second monitoring scene comprises:

5. The method of claim 1, wherein, The method comprises: Obtaining a plurality of initial pre-training encoders, a plurality of initial scene processing units corresponding to the monitoring scenes, an initial time sequence correlation model, a cold chain temperature control historical training set and sample temperature control abnormality labels corresponding to the cold chain temperature control historical training set; Coupling the plurality of initial pre-training encoders, the plurality of initial scene processing units corresponding to the monitoring scenes and the initial time sequence correlation model into an initial full-scene temperature control perception model; 6. The method of claim 1, wherein, Coupling the plurality of initial pre-training encoders, the plurality of initial scene processing units corresponding to the monitoring scenes and the initial time sequence correlation 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 abnormality annotation 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 comprises a plurality of pre-training encoders, a plurality of scene processing units corresponding to the monitoring scenes, and the time sequence correlation model.

7. The method of claim 6, wherein, The cold chain temperature control history training set comprises sample subject attribute data corresponding to a historical transportation subject instance and sample parameter state data corresponding to a historical temperature control parameter record; according to the cold chain temperature control history training set and the sample temperature control abnormality annotation 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, which comprises: According to a plurality of initial pre-training encoders, the sample subject attribute data and the sample parameter state data are preliminarily feature-extracted to obtain preliminary encoding features output by each initial pre-training encoder; According to a plurality of initial scene processing units, the preliminary encoding features output by each initial pre-training encoder are scene-correlation-fused to obtain sample scene-correlation features output by each initial scene processing unit, and according to the initial time sequence correlation model, the preliminary encoding features output by each initial pre-training encoder are time sequence correlation feature-mined to obtain sample time sequence correlation features output by the initial time sequence correlation model; According to the sample scene-correlation features corresponding to each monitoring scene, a first temperature control abnormality monitoring result corresponding to each monitoring scene is obtained, and according to the sample time sequence correlation features corresponding to each monitoring scene, a second temperature control abnormality monitoring result corresponding to each monitoring scene is obtained; According to the sample scene-correlation features and the sample time sequence correlation features, full-scene prediction calculation is performed to obtain a third temperature control abnormality monitoring result corresponding to each monitoring scene; According to the first temperature control abnormality monitoring result and the sample temperature control abnormality annotation, a first abnormality deviation value is obtained, according to the second temperature control abnormality monitoring result and the sample temperature control abnormality annotation, a second abnormality deviation value is obtained, and according to the third temperature control abnormality monitoring result and the sample temperature control abnormality annotation, a third abnormality deviation value is obtained; According to the first abnormality deviation value, the second abnormality deviation value, and the third abnormality deviation value, a model deviation value is determined; According to the model deviation value, the initial full-scene temperature control perception model is trained until the model deviation value does not exceed a deviation tolerance, to obtain a target full-scene temperature control perception model.

8. A server system, characterized by The system comprises a server configured to perform the method of any one of claims 1-7.

Images