Method and system for full traceability of medical cold chain logistics transport

Abstract

The application provides a medicine cold chain logistics transportation whole-process tracking method and system, and relates to the technical field of logistics transportation monitoring. First, the original sensor tracking data stream output by the internal sensor cluster of a transportation carrier in a continuous collection time window is acquired. Then, the transportation carrier operation trajectory line is generated according to the geographic positioning parameters, and the multi-dimensional environment state marker set is generated by combining the physical parameters at the same time of each geographic position coordinate point. Meanwhile, the logistics transfer node topology connection network containing the logistics transfer node unit and the directed transfer path edge is constructed. The operation trajectory line, the multi-dimensional environment state marker set and the topology connection network are input into the pre-constructed space-time state feature fusion tracking model to generate the whole-process state tracking feature tensor. The abnormal state type identification and the space-time distribution feature are obtained by performing the abnormal state mode matching on the whole-process state tracking feature tensor, and the abnormal early warning instruction set is generated. The application can comprehensively and accurately track the whole process of medicine cold chain transportation, intelligently identify the abnormality and timely generate the early warning.

Description

Technical Field

[0001] This invention relates to the field of logistics and transportation monitoring technology, and more specifically, to a method and system for tracking the entire process of pharmaceutical cold chain logistics transportation. Background Technology

[0002] Pharmaceutical products are extremely sensitive to environmental conditions such as temperature and humidity. Any abnormal fluctuations in these parameters can lead to product deterioration or failure, thereby affecting patients' health and safety. Therefore, accurate tracking and effective monitoring of the entire pharmaceutical cold chain logistics transportation process is a key requirement for the industry's development.

[0003] Currently, existing methods for tracking pharmaceutical cold chain logistics transportation have many limitations. Some methods rely solely on data from a single sensor, such as focusing only on temperature parameters, while ignoring other factors that may affect the quality of pharmaceutical products, such as humidity and vibration, thus failing to comprehensively reflect the environmental conditions during transportation. Furthermore, while some methods collect data on various environmental parameters, they simply record and display the data without in-depth analysis and fusion, making it difficult to uncover potential risks and anomalies hidden within the data.

[0004] In terms of transportation trajectory tracking, most existing technologies can only provide approximate location information of the transport vehicle, failing to accurately construct its operational trajectory or clearly display the various logistics transfer nodes it passes through during transportation and their interconnections. This makes it difficult to quickly pinpoint the specific location and stage of a problem when it occurs, hindering timely and effective countermeasures. Furthermore, existing tracking methods lack intelligent identification and early warning mechanisms for abnormal states. When abnormal environmental parameters or deviations from the transportation trajectory occur during transport, timely warnings are often not issued, leading to delayed problem resolution and increasing the quality risks of pharmaceutical products. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for end-to-end tracking of pharmaceutical cold chain logistics transportation, the method comprising:

[0006] The raw sensor tracking data stream is obtained from the sensor cluster deployed inside the transport vehicle during the cold chain logistics transportation of pharmaceuticals within a continuous acquisition time window. The raw sensor tracking data stream includes a temperature sensor parameter subsequence, a humidity sensor parameter subsequence, a vibration sensor parameter subsequence, and a geolocation parameter subsequence marked with acquisition timestamps.

[0007] The transport vehicle's trajectory is generated based on the movement order of the geographic location coordinates in the geographic location parameter subsequence, and a multi-dimensional environmental state marker set associated with each geographic location coordinate in the transport vehicle's trajectory is generated based on the physical parameters in the temperature sensor parameter subsequence, the humidity sensor parameter subsequence, and the vibration sensor parameter subsequence that have the same timestamp as each geographic location coordinate.

[0008] Construct a logistics transfer node topology connection network along the trajectory of the transport vehicle. The logistics transfer node topology connection network includes multiple logistics transfer node units and directed transfer path edges connecting the logistics transfer node units. The logistics transfer node units are obtained by matching the geographical coordinates of the transport vehicle's trajectory with a preset logistics transfer node database.

[0009] The transport vehicle's trajectory line, the multi-dimensional environmental state marker set, and the logistics transfer node topology connection network are input into a pre-constructed spatiotemporal state feature fusion tracking model for joint analysis and processing to generate a pharmaceutical cold chain logistics transportation full-process state tracking feature tensor.

[0010] Anomaly pattern matching processing is performed on the feature tensor of the entire state tracking of pharmaceutical cold chain logistics transportation to obtain the anomaly type identifier and spatiotemporal distribution characteristics of the anomaly state corresponding to the feature tensor of the entire state tracking of pharmaceutical cold chain logistics transportation. Anomaly warning instruction set for pharmaceutical cold chain logistics transportation is generated based on the anomaly type identifier and the spatiotemporal distribution characteristics of the anomaly state.

[0011] Furthermore, embodiments of the present invention also provide a pharmaceutical cold chain logistics transportation end-to-end tracking system, comprising:

[0012] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described pharmaceutical cold chain logistics transportation end-to-end tracking method by executing the machine-executable instructions.

[0013] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, the processor of the pharmaceutical cold chain logistics transportation end-to-end tracking system reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the pharmaceutical cold chain logistics transportation end-to-end tracking system to execute the above-described pharmaceutical cold chain logistics transportation end-to-end tracking method.

[0014] Based on the above, this study acquires raw sensor tracking data streams containing multiple parameters from the sensor clusters inside the transport vehicle, generates the transport vehicle's trajectory line based on geolocation parameters, and combines this with a multi-dimensional environmental state marker set associated with each geographical coordinate point. A logistics transfer node topology network is then constructed, demonstrating the various logistics transfer nodes the transport vehicle passes through and the directed transfer path relationships between them. The trajectory line, multi-dimensional environmental state marker set, and logistics transfer node topology network are input into a spatiotemporal state feature fusion tracking model for joint analysis, generating a full-process state tracking feature tensor for pharmaceutical cold chain logistics transportation. This achieves deep fusion and feature extraction of multiple data sources, uncovering hidden potential information and risk patterns. Abnormal state pattern matching processing of the state tracking feature tensor accurately identifies the types and spatiotemporal distribution characteristics of abnormal states during transportation and generates corresponding abnormal early warning instruction sets. This enables intelligent perception and timely early warning of abnormal situations in pharmaceutical cold chain logistics transportation, helping relevant personnel to quickly take measures to ensure the quality and safety of pharmaceutical products and improving the reliability and management level of pharmaceutical cold chain logistics transportation. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the execution flow of the pharmaceutical cold chain logistics transportation end-to-end tracking method provided in this embodiment of the invention.

[0016] Figure 2 This is a schematic diagram of exemplary hardware and software components of the pharmaceutical cold chain logistics transportation end-to-end tracking system provided in this embodiment of the invention. Detailed Implementation

[0017] Figure 1 This is a flowchart illustrating a method for tracking the entire process of pharmaceutical cold chain logistics transportation according to an embodiment of the present invention, which will be described in detail below.

[0018] Step S110: Obtain the raw sensor tracking data stream output by the sensor cluster deployed inside the transport vehicle during the pharmaceutical cold chain logistics transportation process within the continuous acquisition time window. The raw sensor tracking data stream includes temperature sensor parameter subsequence, humidity sensor parameter subsequence, vibration sensor parameter subsequence and geolocation parameter subsequence marked with acquisition timestamp.

[0019] In this embodiment, the transport vehicle is a fully equipped refrigerated truck for vaccine cold chain transport. The refrigerated truck houses an integrated sensor cluster containing multiple sensors of different types. Before the transport mission begins, the sensor cluster is initialized, and a unified continuous data acquisition time window is set, for example, the entire transport cycle from the mission start time T_start to the mission end time T_end.

[0020] The sensor cluster continuously outputs raw sensor tracking data streams at a preset acquisition frequency, such as once every second. Temperature sensors acquire the temperature value inside the refrigerated truck compartment once per second, adding a timestamp to each temperature sample to generate a temperature sensing parameter subsequence. Humidity sensors acquire the relative humidity value inside the compartment at the same frequency, adding a timestamp to each humidity sample to generate a humidity sensing parameter subsequence. Vibration sensors, installed inside the refrigerated truck chassis or cargo box, sense mechanical vibrations during transportation, acquiring vibration amplitude values ​​once per second and adding timestamps to generate a vibration sensing parameter subsequence. The GPS module acquires the refrigerated truck's current geographic coordinates, including longitude and latitude, once per second, adding a timestamp to each geographic coordinate to generate a geolocation parameter subsequence. These four subsequences together constitute the raw sensor tracking data stream, with the number of parameters in each subsequence equal to the total number of seconds from T_start to T_end plus one.

[0021] Step S120: Generate the transport vehicle's running trajectory line according to the movement order of the geographic location coordinates in the geographic location parameter subsequence, and generate a multi-dimensional environmental state marker set associated with each geographic location coordinate point in the transport vehicle's running trajectory line according to the physical parameters in the temperature sensing parameter subsequence, humidity sensing parameter subsequence, and vibration sensing parameter subsequence that have the same timestamp as each geographic location coordinate point.

[0022] Step S121: Extract multiple geographic location coordinate points arranged in chronological order of collection time from the geographic location parameter subsequence, and perform redundant point removal processing on the geographic location coordinate points to obtain an effective geographic location coordinate point sequence for constructing the operating trajectory line of the transport vehicle. The redundant point removal processing is achieved by calculating the displacement distance parameter between adjacent geographic location coordinate points and filtering the geographic location coordinate points whose displacement distance parameter is less than a preset static discrimination threshold.

[0023] A geolocation parameter subsequence is extracted from the raw sensor tracking data stream. This subsequence contains one geolocation coordinate point per second from time T_start to time T_end, forming an initial list of coordinate points with a length equal to the total number of seconds plus one. This list is iterated through, and for each pair of adjacent coordinate points, such as the i-th geolocation coordinate point P_i and the (i+1)-th geolocation coordinate point P_{i+1}, the Euclidean distance between these two points is calculated as the displacement distance parameter D_i. This displacement distance parameter is calculated as follows: the longitude and latitude values ​​of P_i and P_{i+1} are converted to their actual distance projections on the Earth's surface, obtaining the distance difference in the longitude and latitude directions. Then, the square root of the sum of the squares of these two differences is calculated using the Pythagorean theorem to obtain the straight-line distance D_i. A stationary discrimination threshold D_static is set, which can be adjusted according to actual needs, for example, to 5 meters. If the calculated displacement distance parameter D_i is less than the stationary discrimination threshold of 5 meters, the vehicle is considered to be stationary during that time period, and the next point P_{i+1} is considered redundant and is removed from the initial list. If the displacement distance parameter D_i is greater than or equal to 5 meters, the point P_{i+1} is retained. The process continues with the next point pair until the entire list has been traversed. After traversing and removing points from the entire initial list, a sequence containing only valid location points during the vehicle's actual movement is obtained, denoted as sequence Q. The length of this valid geographic location coordinate sequence is M, which is much smaller than the length of the original sequence.

[0024] Step S122: Based on the acquisition timestamp corresponding to each valid geographic location coordinate point in the valid geographic location coordinate point sequence, calculate the time interval parameter and displacement direction angle parameter between adjacent valid geographic location coordinate points; normalize the time interval parameter and displacement direction angle parameter respectively, and generate a motion state change description vector between adjacent valid geographic location coordinate points based on the normalized time interval parameter and displacement direction angle parameter. The motion state change description vector includes a standardized motion duration component and a standardized motion direction change component.

[0025] For each pair of adjacent points in the sequence of valid geographic coordinate points Q, such as the j-th valid geographic coordinate point Q_j and the (j+1)-th valid geographic coordinate point Q_{j+1}, the acquisition times T_j and T_{j+1} are extracted from the timestamps attached to these two points, and the difference between T_{j+1} and T_j is calculated to obtain the time interval parameter Δt_j. Next, the displacement direction from point Q_j to point Q_{j+1} is calculated. The displacement direction is referenced to true north at 0 degrees, and is obtained by calculating the clockwise angle between the line connecting the two points and true north. This angle ranges from 0 degrees to 360 degrees and is used as the displacement direction angle parameter θ_j. To eliminate the influence of dimensions and make subsequent processing more stable, the time interval parameter and the displacement direction angle parameter are normalized. For the time interval parameter Δt_j, it is divided by the maximum time interval value Δt_max that occurs throughout the entire transportation process to obtain the normalized motion duration component α_j, which ranges from 0 to 1. α_j is equal to Δt_j divided by Δt_max. For the displacement direction angle parameter θ_j, normalization is performed using the sine and cosine values ​​of the angle, calculating sin(θ_j) and cos(θ_j). Both values ​​range from -1 to 1, together forming the normalized motion direction change component. The normalized motion duration component α_j, the normalized motion direction sine component sin(θ_j), and the normalized motion direction cosine component cos(θ_j) are combined to form a 3-dimensional motion state change description vector V_j, where V_j equals (α_j, sin(θ_j), cos(θ_j)). This motion state change description vector is used to characterize the duration and direction change features of the motion from point Q_j to point Q_{j+1}.

[0026] Step S123: Associate and combine the effective geographic location coordinate point sequence with the motion state change description vector between the adjacent effective geographic location coordinate points to generate an initial trajectory segment set with temporal continuity. The initial trajectory segment set consists of consecutive effective geographic location coordinate points in the effective geographic location coordinate point sequence and their corresponding motion state change description vectors.

[0027] Each point in the sequence of valid geographic coordinates Q is associated with a motion state change description vector connecting that point to the next point. Specifically, for point Q_j, it is associated with a motion state change description vector V_j from itself to Q_{j+1}. For the last point Q_M, there is no subsequent motion vector, so it only contains the point coordinates. A series of data units are generated, each containing a point coordinate and a motion vector (except for the last point). These units are arranged in the original time order, forming an initial trajectory segment set with temporal coherence. This set describes the details of the vehicle's motion state changes on different road segments, and each element in the set corresponds to a trajectory segment.

[0028] Step S124: Perform trajectory segment splicing processing on the initial trajectory segment set. Based on the matching degree of the geographical coordinates of the endpoints of adjacent trajectory segments in the initial trajectory segment set and the continuity parameter calculated based on the motion state change description vector, connect the adjacent trajectory segments that meet the preset splicing conditions to generate the transport vehicle running trajectory line.

[0029] Since redundant points have been removed, the segments in the initial trajectory segment set are already continuous. However, in certain special cases, such as data interruptions caused by GPS signal loss, truly discontinuous segments may occur. It is necessary to ensure the final integrity of the trajectory. The initial trajectory segment set is traversed, checking the connection between the endpoints of adjacent segments. For the last point Q_k of the previous segment and the first point Q_{k+1} of the next segment, the matching degree of their geographic coordinates is calculated, i.e., the displacement distance parameter D_k between them. Simultaneously, the similarity between the last motion vector V_k of the previous segment and the first motion vector V_{k+1} of the next segment is compared, and the cosine similarity between the two vectors is calculated as the continuity parameter C_k. The cosine similarity is equal to the dot product of V_k and V_{k+1} divided by the modulus of V_k multiplied by the modulus of V_k. If the displacement distance parameter D_k is less than a preset splicing threshold, such as 10 meters, and the continuity parameter C_k is higher than a preset similarity threshold, such as 0.95, then the two segments are considered continuous in space and motion trend, and the latter segment is directly connected to the former segment. If D_k is greater than or equal to 10 meters or C_k is less than or equal to 0.95, then a data interruption is identified, and this point needs to be marked as a breakpoint. After checking and splicing all adjacent segments, the final transport vehicle trajectory line is generated, which is a complete sequence composed of valid point coordinates and their associated motion vectors.

[0030] Step S125: Perform logistics node proximity marking processing on each trajectory point in the operation trajectory line of the transport vehicle, calculate the spatial distance parameter between each trajectory point and all logistics transfer nodes in the preset logistics transfer node database, and mark the trajectory points whose spatial distance parameter is less than the preset proximity distance threshold as logistics node proximity points. At the same time, record the identifier of the logistics transfer node closest to the trajectory point as the node association identifier.

[0031] For each trajectory point in the transport vehicle's trajectory line, such as trajectory point Q_j, its geographical coordinates are obtained. Simultaneously, information on all logistics transfer nodes is read from a pre-set logistics transfer node database. Each node contains a node identifier and its geographical coordinates. The logistics transfer node database is pre-built and includes all possible warehouses, transit stations, distribution centers, etc., along the route, such as nodes A, B, C, up to H. For trajectory point Q_j, the spatial distance parameter between it and each logistics transfer node, such as node A, is calculated sequentially. The calculation method is the same as the displacement distance parameter in step S121, i.e., calculating the Euclidean distance between the two points. Distance values ​​d(Q_j, A), d(Q_j, B), up to d(Q_j, H) are obtained. A proximity distance threshold is set, for example, 500 meters. All calculated distance values ​​are iterated. If any distance value is less than 500 meters, for example, d(Q_j, C) equals 300 meters, then trajectory point Q_j is marked as a logistics node proximity point. Simultaneously, the minimum value is found among all distance values ​​less than 500 meters. The logistics transfer node corresponding to this minimum value is the node closest to trajectory point Q_j, and its identifier, such as node C, is recorded as the node association identifier for trajectory point Q_j. If all distance values ​​are greater than or equal to 500 meters, trajectory point Q_j is not marked as a nearby point.

[0032] Step S126: Based on the distribution density and distribution order of the logistics node proximity points on the transport vehicle's operating trajectory, determine the sequence of logistics transfer nodes traversed by the transport vehicle's operating trajectory. The sequence of logistics transfer nodes is composed of logistics transfer node identifiers arranged sequentially according to the travel direction of the transport vehicle's operating trajectory.

[0033] The entire transport vehicle's trajectory is traversed to identify all trajectory points marked as neighboring points to logistics nodes. Based on the order in which these neighboring points appear on the trajectory, a preliminary node visit list is formed. However, since vehicles may linger near the same node for extended periods, multiple consecutive neighboring points may correspond to the same node, requiring deduplication. For example, trajectory points Q_5, Q_6, Q_7, and Q_8 are all marked as neighboring points, and their node association identifier is node C. Therefore, when determining the sequence of nodes passed through, only node C is retained, representing that the vehicle passed through node C. Following the direction of travel on the trajectory, each different node identifier is processed sequentially, ultimately generating an ordered sequence of logistics transfer nodes, for example, starting from node A, then passing through node C, then node E, and finally arriving at node H. This sequence represents the order of critical logistics nodes in the vehicle's actual route.

[0034] Step S127: Generate a multi-dimensional environmental state marker set associated with each geographical coordinate point in the transport vehicle's trajectory line based on the physical parameters in the temperature sensing parameter subsequence, humidity sensing parameter subsequence, and vibration sensing parameter subsequence that have the same timestamp as each geographical coordinate point.

[0035] Step S1271: Extract the instantaneous temperature sampling value that is the same as the collection timestamp of each geographical coordinate point from the temperature sensing parameter subsequence, and arrange the instantaneous temperature sampling values ​​according to the order of appearance of the geographical coordinate points on the transport vehicle's running trajectory to generate a temperature parameter time sequence association list corresponding to the transport vehicle's running trajectory.

[0036] For each trajectory point in the transport vehicle's trajectory line, such as trajectory point Q_j, obtain its acquisition timestamp T_j. Find the instantaneous temperature sample value Temp_j with timestamp equal to T_j from the temperature sensing parameter subsequence. Since the acquisition frequency is consistent, theoretically each timestamp has a corresponding temperature value. Arrange the found Temp_j values ​​according to the order of trajectory points Q_j on the trajectory line, generating a temperature parameter time-series association list Temp_List that corresponds one-to-one with the trajectory line. The list length is M, and each element corresponds to the temperature value of a trajectory point.

[0037] Step S1272: Extract the instantaneous humidity sampling value that is the same as the collection timestamp of each geographical coordinate point from the humidity sensing parameter subsequence, and arrange the instantaneous humidity sampling values ​​according to the order of appearance of the geographical coordinate points on the transport vehicle's running trajectory to generate a humidity parameter time sequence association list corresponding to the transport vehicle's running trajectory.

[0038] Similarly, for trajectory point Q_j, find the instantaneous humidity sample value Hum_j with timestamp equal to T_j from the humidity sensing parameter subsequence. Arrange Hum_j in the order of trajectory point Q_j to generate a humidity parameter temporal association list Hum_List, with a list length of M.

[0039] Step S1273: Extract the instantaneous vibration sampling value that is the same as the acquisition timestamp of each of the geographical coordinate points from the vibration sensing parameter subsequence, and arrange the instantaneous vibration sampling values ​​according to the order of appearance of the geographical coordinate points on the transport vehicle's running trajectory to generate a vibration parameter time sequence association list corresponding to the transport vehicle's running trajectory.

[0040] Similarly, for trajectory point Q_j, find the instantaneous vibration sample value Vib_j with timestamp equal to T_j from the vibration sensing parameter subsequence. Arrange Vib_j according to the order of trajectory point Q_j to generate a vibration parameter temporal association list Vib_List, with a list length of M.

[0041] Step S1274: Perform temperature state interval mapping processing on each instantaneous temperature sample value in the temperature parameter time sequence association list. Convert each instantaneous temperature sample value into a corresponding temperature state label according to the preset temperature state division rule. The temperature state label is used to characterize the degree of deviation of the internal ambient temperature of the transport carrier from the required temperature range for drug storage.

[0042] Pre-defined temperature state classification rules. Assume the required storage temperature range for vaccines is 2°C to 8°C. Based on this range, several temperature state categories are defined: for example, when the instantaneous temperature sample value is below 2°C, it is marked as "low temperature deviation"; when the instantaneous temperature sample value is between 2°C and 8°C, it is marked as "normal"; when the instantaneous temperature sample value is above 8°C but below 10°C, it is marked as "slight overheating"; and when the instantaneous temperature sample value is above 10°C, it is marked as "severe overheating". For each temperature value Temp_j in the temporal association list Temp_List, it is judged according to the above rules and converted into a corresponding temperature state label, such as Temp_Label_j. The entire list is traversed to generate a temperature state label sequence Temp_Label_List.

[0043] Step S1275: Perform humidity state interval mapping processing on each instantaneous humidity sample value in the humidity parameter time sequence association list. Convert each instantaneous humidity sample value into a corresponding humidity state label according to the preset humidity state division rules. The humidity state label is used to characterize the degree of deviation of the internal environmental humidity of the transport carrier from the humidity range required for drug storage.

[0044] Pre-defined humidity state classification rules. Assume that vaccine storage requires a humidity range of 35% to 75%. Based on this range, several humidity state categories are defined: for example, when the instantaneous humidity sampling value is below 35%, it is marked as "dry"; when the instantaneous humidity sampling value is between 35% and 75%, it is marked as "normal"; and when the instantaneous humidity sampling value is above 75%, it is marked as "humid". For each humidity value Hum_j in the humidity parameter time-series association list Hum_List, it is judged according to the above rules and converted into a corresponding humidity state label Hum_Label_j. The entire list is traversed to generate the humidity state label sequence Hum_Label_List.

[0045] Step S1276: Perform vibration intensity level mapping processing on each instantaneous vibration sample value in the vibration parameter time sequence association list. According to the preset vibration intensity level classification rules, convert each instantaneous vibration sample value into a corresponding vibration intensity level label. The vibration intensity level label is used to characterize the level of mechanical impact intensity that the transport vehicle is subjected to during transportation.

[0046] A pre-defined vibration intensity level classification rule is established. Based on the amplitude of the instantaneous vibration sampling value, several vibration intensity levels are defined, such as levels 1 to 5. For example, when the vibration value is less than threshold A, it is level 1, indicating slight vibration; when the vibration value is between threshold A and threshold B, it is level 2; when the vibration value is between threshold B and threshold C, it is level 3; when the vibration value is between threshold C and threshold D, it is level 4; and when the vibration value is greater than threshold D, it is level 5, indicating severe impact. Thresholds A, B, C, and D can be dynamically set based on historical data and drug tolerance. For each vibration value Vib_j in the vibration parameter time-series association list Vib_List, it is judged according to the above rule and converted into the corresponding vibration intensity level label Vib_Label_j. The entire list is traversed to generate the vibration intensity level label sequence Vib_Label_List.

[0047] Step S1277: Combine and encapsulate the temperature status marker, humidity status marker, and vibration intensity level marker according to their association with the geographical coordinate points to generate a multi-dimensional environmental status marker set corresponding to each geographical coordinate point in the transport vehicle's trajectory line.

[0048] For each trajectory point Q_j, its corresponding temperature status label (Temp_Label_j), humidity status label (Hum_Label_j), and vibration intensity level label (Vib_Label_j) are combined to form a tuple: (Temp_Label_j, Hum_Label_j, Vib_Label_j). This tuple constitutes the multi-dimensional environmental status label for trajectory point Q_j. Arranging these tuples for all trajectory points in sequence generates a multi-dimensional environmental status label set that perfectly corresponds to the transport vehicle's trajectory. Each element in this set is a label vector containing status information for temperature, humidity, and vibration.

[0049] Step S130: Construct a logistics transfer node topology connection network along the trajectory of the transport vehicle. The logistics transfer node topology connection network includes multiple logistics transfer node units and directed transfer path edges connecting the logistics transfer node units. The logistics transfer node units are obtained by matching the geographical coordinates of the transport vehicle's trajectory with a preset logistics transfer node database.

[0050] Step S131: Extract all trajectory points marked as logistics node proximity points from the transport vehicle's running trajectory line, and cluster the logistics node proximity points belonging to the same logistics transfer node identifier according to the node association identifier corresponding to each logistics node proximity point, to obtain a set of node trajectory points corresponding to each logistics transfer node identifier.

[0051] From the transport vehicle's operational trajectory line, identify all trajectory points marked as proximity points to logistics nodes in step S125. Based on the node association identifier attached to each proximity point, group proximity points with the same node association identifier into one category. For example, all proximity points with the node association identifier of node A constitute the trajectory point set S_A for node A; all proximity points with the node association identifier of node B constitute the trajectory point set S_B for node B, and so on. Each set S_X contains all trajectory points when the vehicle stops or passes near the corresponding node X.

[0052] Step S132: Calculate the average value of the geographical coordinates of all neighboring points of the logistics node in each set of node trajectory points, and generate the actual passing position coordinates of the logistics transfer node unit represented by each logistics transfer node identifier in the running trajectory line of the transport vehicle.

[0053] For each set of trajectory points, such as S_A for node A, this set contains several trajectory points, each with its own longitude and latitude values. Calculate the arithmetic mean of these longitude values ​​to obtain the average longitude value; calculate the arithmetic mean of these latitude values ​​to obtain the average latitude value. The geographical coordinates formed by these average longitude and average latitude values ​​are the actual passing position coordinates P_A of node A in this transportation trajectory. The same calculation method is applied to all traversed nodes, such as node B and node C, to obtain the actual passing position coordinates for each node.

[0054] Step S133: Based on the time sequence in which each logistics transfer node unit is passed on the trajectory line of the transport vehicle, arrange the logistics transfer node units in chronological order to generate a sequence of logistics transfer nodes actually passed by the transport vehicle.

[0055] The arrival time of a vehicle in a given node's trajectory point set can be determined based on the earliest occurrence time of each trajectory point. All traversed nodes are then sorted according to the earliest timestamps in their corresponding trajectory point sets. For example, if the earliest timestamp in node A's set is T_A, and the earliest timestamp in node C's set is T_C, and T_A is earlier than T_C, then node A is placed before node C. This ultimately generates an ordered sequence of logistics transfer nodes, denoted as Node_Sequence, for example, (node ​​A, node C, node E, node H).

[0056] Step S134: For two adjacent logistics transfer node units in the logistics transfer node sequence, calculate the spatial straight-line distance between the two logistics transfer node units as the path distance parameter based on the actual passing position coordinates corresponding to the previous logistics transfer node unit and the actual passing position coordinates corresponding to the next logistics transfer node unit.

[0057] For each pair of adjacent nodes in the logistics transfer node sequence Node_Sequence, such as node A and node C, obtain the actual transit location coordinates P_A of node A and P_C of node C. Calculate the spatial straight-line distance between P_A and P_C, i.e., the Euclidean distance, using the same calculation method as the displacement distance parameter in step S121. Obtain the path distance parameter D_AC, which is used to characterize the geographical distance between node A and node C.

[0058] Step S135: Based on the order of adjacent logistics transfer node units in the logistics transfer node sequence, determine the directed transfer path edge connecting the previous logistics transfer node unit to the next logistics transfer node unit, and attach the path distance parameter to the directed transfer path edge as edge attribute information.

[0059] Based on the order of the nodes in the sequence, node A comes first, followed by node C. Therefore, a directed transport path edge Edge_AC is defined from node A to node C. The path distance parameter D_AC calculated in step S134 is appended to this edge as its attribute information. For the next pair of nodes C and E, a directed edge Edge_CE is also defined, and the path distance parameter D_CE is appended.

[0060] Step S136: Take all the logistics transfer node units as network nodes, take all the directed transfer path edges as directed edges connecting the network nodes, and combine the edge attribute information to construct the logistics transfer node topology connection network.

[0061] All logistics transfer node units identified in step S133, namely node A, node C, node E, and node H, are designated as nodes in the network. All directed transfer path edges defined in step S135, namely Edge_AC, Edge_CE, and Edge_EH, are designated as directed edges connecting these nodes. Each edge is associated with a corresponding path distance attribute. This constructs a directed graph structure, namely the logistics transfer node topology connection network. This network clearly reflects the actual sequence of logistics nodes traversed by the vehicle and the spatial distance relationships between the nodes.

[0062] Step S140: Input the transport vehicle's trajectory line, the multi-dimensional environmental state marker set, and the logistics transfer node topology connection network into the pre-constructed spatiotemporal state feature fusion tracking model for joint analysis processing to generate a pharmaceutical cold chain logistics transportation full-process state tracking feature tensor.

[0063] Step S141: Input the trajectory line of the transport vehicle into the trajectory encoding module of the spatiotemporal state feature fusion tracking model, perform spatial position embedding representation processing on the sequence of geographical location coordinate points in the trajectory line of the transport vehicle, and generate the spatial trajectory embedding feature vector sequence corresponding to the trajectory line of the transport vehicle.

[0064] The spatiotemporal state feature fusion tracking model is a pre-built and trained deep learning model. First, the trajectory line of the transport vehicle, i.e., the sequence of effective geographical coordinate points Q, is input into the model's trajectory encoding module. This module first normalizes the longitude and latitude values ​​of each trajectory point Q_j, mapping the original longitude and latitude ranges to a unified numerical interval, such as -1 to 1. Then, a linear transformation layer maps the normalized two-dimensional coordinates (longitude and latitude values) into a higher-dimensional vector, such as a spatial trajectory embedding feature vector E_pos_j with dimension D1. This linear transformation layer is a fully connected network, and its weight parameters are learned during model training. For all M points on the trajectory line, after processing by the trajectory encoding module, a spatial trajectory embedding feature vector sequence E_pos is generated. This spatial trajectory embedding feature vector sequence is a matrix of shape M multiplied by D1, with each row corresponding to the spatial location embedding representation of a trajectory point.

[0065] Step S142: Input the multi-dimensional environmental state label set into the state encoding module of the spatiotemporal state feature fusion tracking model, perform joint embedding representation processing on the temperature state label, humidity state label and vibration intensity level label corresponding to each geographical coordinate point in the multi-dimensional environmental state label set, generate environmental state embedding feature vector corresponding to each geographical coordinate point, and arrange all the environmental state embedding feature vectors according to the order of appearance of the geographical coordinate points to generate an environmental state embedding feature vector sequence.

[0066] Next, the multi-dimensional environmental state label set, i.e., the triplet (Temp_Label_j, Hum_Label_j, Vib_Label_j) corresponding to each trajectory point Q_j, is input into the model's state encoding module. Temperature state labels, humidity state labels, and vibration intensity level labels are all discrete categorical variables. The state encoding module first establishes an independent embedding lookup table for each categorical variable. For example, for the temperature state label, an embedding table is established, mapping the four categories "low temperature deviation," "normal," "mild over-temperature," and "severe over-temperature" to vectors of dimension D2. Similarly, embedding tables are also established for the humidity state label and vibration intensity level label, mapping them to vectors of the same dimension D2. For trajectory point Q_j, its temperature embedding vector e_temp_j, humidity embedding vector e_hum_j, and vibration embedding vector e_vib_j are obtained by looking up the tables; all three vectors have a dimension of D2. Then, these three vectors are concatenated along the feature channel dimension to form an environmental state embedding feature vector E_env_j with a dimension of 3 times D2. Perform the same operation on all M trajectory points to obtain an environment state embedding feature vector sequence E_env, which is a matrix of shape M multiplied by (3*D2).

[0067] Step S143: Input the logistics transfer node topology connection network into the topology encoding module of the spatiotemporal state feature fusion tracking model, perform graph structure embedding representation processing on the logistics transfer node units and directed transfer path edges in the logistics transfer node topology connection network, and generate the network topology embedding feature representation corresponding to the logistics transfer node topology connection network. The network topology embedding feature representation includes the node embedding feature vector of each logistics transfer node unit and the edge embedding feature vector of each directed transfer path edge.

[0068] The constructed logistics transfer node topology network, i.e., a graph containing nodes A, C, E, H and edges Edge_AC, Edge_CE, and Edge_EH, is input into the topology encoding module. This module first generates an initial node feature vector for each node in the network, such as node A. This initial vector can be based on the node's inherent attributes, such as the node type (e.g., provincial hub, municipal distribution center) or the node's historical throughput. If this information is unavailable, random initialization or a zero vector can be used. Then, a graph neural network layer, such as a graph convolutional network layer or a graph attention network layer, aggregates information from neighboring nodes to update the feature representation of each node. In the graph neural network layer, the updated features of node A consider the information of its neighbor node C and the edge attributes (path distance parameter D_AC) of the edge Edge_AC between them. After processing by the graph neural network layer, each node is encoded as a node embedding feature vector of dimension D3; for example, the embedding vector for node A is Node_A_emb. Meanwhile, for each edge, such as Edge_AC, its path distance parameter D_AC can be transformed into a D4-dimensional edge embedding feature vector Edge_AC_emb through a linear transformation. Finally, the topology coding module outputs the network topology embedding feature representation, which contains the set of embedding feature vectors of all nodes and the set of embedding feature vectors of all edges.

[0069] Step S144: Align and splice the spatial trajectory embedded feature vector sequence and the environmental state embedded feature vector sequence on the time axis to generate a primary joint feature tensor that integrates spatiotemporal information. Each time step of the primary joint feature tensor contains both the spatial location information and the environmental state information of the corresponding time point.

[0070] Step S1441: Obtain the original acquisition timestamp corresponding to each spatial trajectory embedding feature vector in the spatial trajectory embedding feature vector sequence, and obtain the original acquisition timestamp corresponding to each environmental state embedding feature vector in the environmental state embedding feature vector sequence.

[0071] Since both the spatial trajectory embedding feature vector sequence E_pos and the environment state embedding feature vector sequence E_env are generated according to the order of the trajectory points Q_j, they naturally have the same order and a one-to-one correspondence. Each vector is associated with the original acquisition timestamp T_j of the same trajectory point Q_j. Therefore, this alignment step is straightforward.

[0072] Step S1442: Perform time-axis alignment verification on the spatial trajectory embedded feature vector sequence and the environmental state embedded feature vector sequence based on the original acquisition timestamp to confirm that the spatial trajectory embedded feature vector sequence and the environmental state embedded feature vector sequence have a one-to-one correspondence in the time dimension.

[0073] By simply comparing whether the lengths of the two sequences are equal (both are M) and whether the vectors corresponding to each position are all associated with the same trajectory point, it can be confirmed that they are in one-to-one correspondence in the time dimension.

[0074] Step S1443: The spatial trajectory embedding feature vector and the environmental state embedding feature vector at each time point after alignment are concatenated and combined in the feature channel dimension to generate a spatiotemporal combination feature vector corresponding to each time point.

[0075] For each trajectory point Q_j, the spatial trajectory embedding feature vector E_pos_j (dimension D1) and the environmental state embedding feature vector E_env_j (dimension 3 x D2) are concatenated along the feature channel dimension. The resulting vector F_j has a dimension of D1 plus 3 x D2. This F_j is the spatiotemporal combined feature vector corresponding to trajectory point Q_j, which simultaneously integrates the spatial location information and multidimensional environmental state information of that point.

[0076] Step S1444: Stack the spatiotemporal combined feature vectors corresponding to all time points in chronological order to form a two-dimensional feature matrix as the primary joint feature tensor. The row dimension of the two-dimensional feature matrix corresponds to the time step, and the column dimension corresponds to the total number of feature channels after concatenation.

[0077] Arrange the spatiotemporal combined feature vectors F_1 to F_M corresponding to all M trajectory points into a two-dimensional matrix in order from 1 to M. The first row of the matrix is ​​F_1, the second row is F_2, and so on, with the last row being F_M. The shape of this matrix is ​​M times (D1 plus 3 times D2), and it is called the primary joint feature tensor T_primary.

[0078] Step S1445: Perform feature dimension normalization processing on the primary joint feature tensor, adjust the numerical range of each feature channel in the primary joint feature tensor to the preset model input range, and generate a normalized primary joint feature tensor with a unified scale representation.

[0079] To ensure stability in subsequent model processing, the primary joint feature tensor T_primary is subjected to layer normalization. For each row in the matrix (i.e., the feature vector at each time step), the mean and variance of all eigenvalues ​​in that row are calculated. Then, each eigenvalue in that row is subtracted from the mean and divided by the square root of the variance, resulting in a mean of 0 and a variance of 1 for each row. After layer normalization, the normalized primary joint feature tensor T_primary_norm is obtained and serves as the input to subsequent modules.

[0080] Step S145: Input the primary joint feature tensor and the network topology embedding feature representation into the cross-attention fusion module of the spatiotemporal state feature fusion tracking model. The cross-attention fusion module uses the primary joint feature tensor as the query basis and the network topology embedding feature representation as the key value basis. It calculates the association weights between the features of each time step in the primary joint feature tensor and each node and edge in the logistics transfer node topology connection network through a multi-head cross-attention mechanism. It then performs weighted aggregation on the network topology embedding feature representation according to the association weights to generate an enhanced feature tensor with network topology context information.

[0081] Step S1451: Linearly map the primary joint feature tensor to the query matrix required by the multi-head cross-attention mechanism. The number of rows in the query matrix corresponds to the number of time steps of the primary joint feature tensor, and the number of columns in the query matrix corresponds to the preset query feature dimension.

[0082] The normalized primary joint feature tensor T_primary_norm, with a shape of M multiplied by (D1 + 3D2), is input into a linear mapping layer (a fully connected network) to map its feature dimension from (D1 + 3D2) to a predefined query feature dimension D_q. The result of the mapping is a matrix of shape M multiplied by D_q, which serves as the query matrix Q for the multi-head cross-attention mechanism.

[0083] Step S1452: Integrate all node embedding feature vectors and edge embedding feature vectors contained in the network topology embedding feature representation, and linearly map them into the key matrix and value matrix required by the multi-head cross-attention mechanism. The number of rows in the key matrix and the value matrix corresponds to the total number of node embedding feature vectors and edge embedding feature vectors, and the number of columns in the key matrix and the value matrix corresponds to the preset key feature dimension and value feature dimension, respectively.

[0084] All node embedding feature vectors (e.g., Node_A_emb, Node_C_emb) and all edge embedding feature vectors (e.g., Edge_AC_emb, Edge_CE_emb) from the network topology embedding feature representation are integrated into a list with L elements (L equals the number of nodes plus the number of edges). This list is then fed into another linear mapping layer, mapping to a preset key feature dimension D_k, resulting in a key matrix K with shape L multiplied by D_k. Simultaneously, the same list is fed into yet another linear mapping layer, mapping to a preset value feature dimension D_v, resulting in a value matrix V with shape L multiplied by D_v.

[0085] Step S1453: Perform a dot product operation on the query matrix and the key matrix to obtain an attention score matrix. Each element in the attention score matrix represents the original score of the association strength between a feature at a time step in the primary joint feature tensor and a node or an edge in the network topology embedding feature representation.

[0086] Calculate the dot product of the query matrix Q (M multiplied by D_q) and the transpose of the key matrix K (D_k multiplied by L). The result of the dot product is an attention score matrix S of shape M multiplied by L. The element S_ij in the i-th row and j-th column of matrix S represents the original association strength score between the trajectory feature at the i-th time step and the j-th network topology element (a node or an edge).

[0087] Step S1454: The attention score matrix is ​​scaled and normalized using softmax to generate an attention weight matrix. The sum of the elements in each row of the attention weight matrix is ​​one, which is used to represent the distribution of the degree of attention of the features at each time step to all nodes and edges.

[0088] Each element in the attention score matrix S is scaled by dividing it by the square root of D_k to prevent the dot product from becoming too large and causing gradient vanishing. Then, each row of the scaled matrix is ​​normalized using the softmax function. After softmax processing, the attention weight matrix A is obtained, still having the shape of M multiplied by L. In each row of matrix A, for example, the i-th row, the sum of all L elements is 1, and the value of each element is between 0 and 1, representing the attention weight of the trajectory feature at the i-th time step for each of the L network topology elements.

[0089] Step S1455: Perform a weighted summation operation on the attention weight matrix and the value matrix to obtain a weighted network topology context feature matrix. The number of rows in the weighted network topology context feature matrix is ​​the same as the number of time steps of the primary joint feature tensor, and the number of columns is the same as the dimension of the value feature.

[0090] Perform matrix multiplication on the attention weight matrix A (M multiplied by L) and the value matrix V (L multiplied by D_v). The result is a matrix C of shape M multiplied by D_v, which is the weighted network topology context feature matrix. The i-th row of matrix C is the weighted sum of all L row vectors in the value matrix V according to the weights of the i-th row of the attention weight matrix A. Therefore, each row of matrix C contains the context information most relevant to the trajectory features at the i-th time step, aggregated from the entire logistics transfer node topology connection network.

[0091] Step S1456: The primary joint feature tensor and the weighted network topology context feature matrix are concatenated and fused along the feature channel dimension to generate the enhanced feature tensor with network topology context information.

[0092] The original primary joint feature tensor T_primary_norm (shape M multiplied by (D1 + 3D2)) is concatenated with the newly generated weighted network topology context feature matrix C (shape M multiplied by D_v) along the feature channel dimension. The shape of the concatenated tensor T_enhanced is M multiplied by ((D1 + 3D2) + D_v). This T_enhanced is the enhanced feature tensor with network topology context information, which integrates the original spatiotemporal trajectory environment information and the context information obtained from the logistics node network.

[0093] Step S146: Input the enhanced feature tensor into the temporal convolutional coding layer of the spatiotemporal state feature fusion tracking model. The temporal convolutional coding layer performs multi-scale temporal receptive field feature extraction processing on the enhanced feature tensor through multiple temporal convolutional kernels with different dilation rates to generate the full-process state tracking feature tensor of the pharmaceutical cold chain logistics transportation.

[0094] The enhanced feature tensor T_enhanced is input into a temporal convolutional coding layer. This layer consists of multiple parallel temporal convolutional blocks, each using a convolutional kernel with a different dilation rate. For example, the first convolutional block uses a kernel with a dilation rate of 1, resulting in a smaller receptive field, used to capture local changes over a short period; the second convolutional block uses a kernel with a dilation rate of 2, resulting in a larger receptive field, used to capture patterns over a slightly longer period; and the third convolutional block uses a kernel with a dilation rate of 4, resulting in an even larger receptive field, used to capture periodic trends over a longer period. Each temporal convolutional block performs a convolution operation on the input tensor T_enhanced along the time dimension (M time steps) to extract features. The output of each convolutional block is a new feature map whose time step size remains unchanged (still M), but the number of feature channels may change. Finally, the outputs of all convolutional blocks are concatenated along the feature channel dimension to obtain a final output tensor. This final output tensor is the feature tensor T_final for tracking the entire state of pharmaceutical cold chain logistics transportation. The shape of T_final is M multiplied by D_final, where D_final is the sum of the number of output channels of all convolutional blocks. This tensor densely encodes a deep feature representation at each time step throughout the entire transportation process, incorporating spatial, environmental, and network topology context information.

[0095] Step S150: Perform abnormal state pattern matching processing on the pharmaceutical cold chain logistics transportation full-process status tracking feature tensor to obtain the abnormal state type identifier and abnormal state spatiotemporal distribution features corresponding to the pharmaceutical cold chain logistics transportation full-process status tracking feature tensor, and generate a pharmaceutical cold chain logistics transportation abnormal early warning instruction set based on the abnormal state type identifier and the abnormal state spatiotemporal distribution features.

[0096] Step S151: Input the pharmaceutical cold chain logistics transportation end-process status tracking feature tensor into a pre-trained abnormal state pattern recognition classifier. The abnormal state pattern recognition classifier performs global feature extraction and pattern classification processing on the pharmaceutical cold chain logistics transportation end-process status tracking feature tensor, and outputs the abnormal state type identifier to which the pharmaceutical cold chain logistics transportation end-process status tracking feature tensor belongs.

[0097] The end-to-end state tracking feature tensor T_final, generated in step S156, with shape M multiplied by D_final, is input into a pre-trained abnormal state pattern recognition classifier. It first passes through a global average pooling layer, performing global average pooling along the time dimension M on T_final. Specifically, for each feature channel, the arithmetic mean of the feature values ​​of that channel over all M time steps is calculated, resulting in a global feature vector G of dimension D_final. This global feature vector G encapsulates the deep state features of the entire transportation process.

[0098] Then, the global feature vector G is input into the classification head of the classifier. The classification head consists of multiple fully connected layers stacked together. For example, the first layer maps the D_final dimension G to a 512-dimensional intermediate vector and passes it through the ReLU activation function. The second layer maps the 512 dimensions to 256 dimensions and passes it through the ReLU activation function. The last layer maps the 256 dimensions to a C-dimensional output vector, where C equals the total number of predefined abnormal state types. These abnormal state types are not simple single-parameter out-of-limit categories, but rather a composite abnormal category system defined based on the spatiotemporal coupling features of multi-source sensor data. The system includes, but is not limited to, the following complex anomaly types: For example, the "progressive failure of the refrigeration unit accompanied by imbalance of heat load in the carriage" type, which manifests in the feature space as a synergistic pattern of continuously increasing energy in a specific frequency band of the vibration spectrum and a nonlinear acceleration of the temperature rise rate; another example is the "defect in the carriage's airtightness accompanied by the intrusion of hot and humid air from the outside," which exhibits a coupled characteristic of a saturated increase in humidity parameters and a fluctuating decrease in temperature parameters, often accompanied by geographical location parameters indicating that the vehicle is in a high-humidity external environment; yet another example is the "accumulation of micro-damage to pharmaceutical packaging due to bumpy transport routes," which manifests as the accumulated impact energy of vibration parameters exceeding a dynamic threshold and showing a statistical correlation with the road surface grade index of the route the vehicle travels on. The C-dimensional vector output from the final layer is normalized using the softmax function to obtain a C-dimensional probability distribution vector P, where each element is between 0 and 1, and the sum of all elements is 1, representing the probability that T_final belongs to each composite anomaly state type. Select the type with the highest probability value and output its corresponding identifier as the final abnormal state type identifier. For example, if the probability of "progressive failure of refrigeration unit accompanied by unbalanced heat load in the carriage" is the highest, then output the identifier "progressive failure of refrigeration unit accompanied by unbalanced heat load in the carriage".

[0099] Step S152: Simultaneously, the feature tensor of the entire state tracking of the pharmaceutical cold chain logistics transportation is input into the pre-trained anomaly spatiotemporal localization network. The anomaly spatiotemporal localization network performs time-step feature analysis on the feature tensor of the entire state tracking of the pharmaceutical cold chain logistics transportation to generate an anomaly confidence score sequence corresponding to each time step.

[0100] Simultaneously, the same end-to-end state tracking feature tensor T_final is input into a parallel anomaly spatiotemporal localization network. This network employs an architecture combining temporal convolution and a bidirectional recurrent neural network. Specifically, T_final first passes through a one-dimensional temporal convolutional layer, which uses multiple convolutional kernels of size 3 to perform convolution operations along the time dimension to capture feature changes within a local time window, outputting a feature map T_conv, whose shape is maintained as M multiplied by a certain feature dimension. Next, T_conv is input into a bidirectional long short-term memory (LSTM) network layer. This bidirectional LSM network layer contains two LSM sublayers: a forward sublayer that processes T_conv sequentially from step 1 to step M, outputting a forward hidden state sequence; and a backward sublayer that processes T_conv in reverse temporal order from step M to step 1, outputting a backward hidden state sequence. For each time step j, the forward and backward hidden states are concatenated along the feature channel dimension to obtain a hidden state vector H_j that integrates bidirectional contextual information. The hidden state H_j at all time steps form a sequence of length M. This hidden state sequence is then input into an output layer, which is a fully connected layer with a sigmoid activation function. This output layer maps H_j at each time step to a scalar value s_j between 0 and 1, serving as the anomaly confidence score for that time step. Finally, the anomaly spatiotemporal localization network outputs an anomaly confidence score sequence S_list of length M, where S_list equals (s_1, s_2, ..., s_M). Each element s_j in the sequence corresponds to the anomaly confidence score at trajectory point Q_j. This anomaly confidence score represents the probability of the aforementioned composite anomaly state occurring within the local time window at that time point.

[0101] Step S153: Based on the abnormal confidence score of each time step in the abnormal confidence score sequence, filter out the continuous time step intervals in which the abnormal confidence score exceeds the preset abnormal judgment threshold, and determine the time range corresponding to the continuous time step intervals on the operating trajectory line of the transport vehicle as the time period of abnormal state occurrence.

[0102] An anomaly detection threshold, Th_abnormal, is set. This threshold is not a fixed constant but an adaptive threshold that is dynamically adjusted based on the anomaly type. For example, for anomalies like "progressive failure of the refrigeration unit accompanied by imbalance of heat load in the carriage," which develop slowly but have serious consequences, the threshold is set to a relatively low 0.7 to detect the anomaly earlier. For anomalies like "burden sealing failure accompanied by intrusion of hot and humid air," which are sudden, the threshold is set to a higher 0.9 to avoid frequent false alarms. Each score s_j in the anomaly confidence score sequence S_list is traversed, and time steps where s_j is greater than the corresponding dynamic threshold are marked as candidate anomaly time steps. Then, consecutive segments in these candidate anomaly time steps are merged into a single interval. For example, suppose in a transportation scenario, for the anomaly type of "progressive failure of the refrigeration unit accompanied by imbalance of heat load in the carriage," all values ​​of s_j from second 800 to second 1200 are greater than 0.7, forming a continuous interval [800, 1200]; all values ​​of s_j from second 2000 to second 2150 are greater than 0.7, forming another continuous interval [2000, 2150]. For each of these continuous intervals, based on its start and end time steps, it corresponds to a specific time range on the trajectory of the transportation vehicle. Assuming the start time step is step a, with a corresponding timestamp T_a, and the end time step is step b, with a corresponding timestamp T_b, then the time period for an anomaly is determined to be from T_a to T_b. In the example above, two time periods for anomalies are obtained: time period 1 is from T_800 to T_1200, and time period 2 is from T_2000 to T_2150.

[0103] Step S154: Based on the time period of the abnormal state occurrence, extract the trajectory segment corresponding to the time period of the abnormal state occurrence from the operating trajectory line of the transport vehicle as the abnormal trajectory segment, and extract the environmental state marker subset corresponding to the time period of the abnormal state occurrence from the multi-dimensional environmental state marker set as the abnormal environmental state marker subset.

[0104] For each time period in which an abnormal state occurs, for example, time period 1 from T_800 to T_1200, all trajectory points corresponding to this time period are extracted from the transport vehicle's trajectory line. The trajectory points start from point Q_800 at the 800th time step and end at point Q_1200 at the 1200th time step, forming an abnormal trajectory segment 1, denoted as Traj_seg_1. Simultaneously, all environmental state label triples corresponding to this time period are extracted from the multi-dimensional environmental state label set, starting from (Temp_Label_800, Hum_Label_800, Vib_Label_800) at the 800th time step and ending at (Temp_Label_1200, Hum_Label_1200, Vib_Label_1200) at the 1200th time step, forming an abnormal environmental state label subset 1, denoted as Env_subset_1. For time period 2, the same operation is performed to obtain abnormal trajectory segment 2 and abnormal environment state marker subset 2.

[0105] Step S155: Calculate the geographical area covered by the abnormal state during the time period of the abnormal state occurrence based on the sequence of geographical coordinate points in the abnormal trajectory segment, and generate a description of the abnormal change trend of the environmental state during the time period of the abnormal state occurrence based on the distribution of each state marker in the subset of abnormal environmental state markers.

[0106] For each anomalous trajectory segment, such as Traj_seg_1, extract the longitude values ​​of all trajectory points and find the minimum longitude Lon_min_1 and the maximum longitude Lon_max_1; extract the latitude values ​​of all trajectory points and find the minimum latitude Lat_min_1 and the maximum latitude Lat_max_1. Use these four extreme points (Lon_min_1, Lat_min_1) and (Lon_max_1, Lat_max_1) to delineate a rectangular geographical area, which is designated as the geographical area Area_1 covered by the anomalous state. For the corresponding anomalous environmental state label subset Env_subset_1, perform multidimensional time series analysis. Taking the temperature state label as an example, map it to a numerical sequence (e.g., "normal" is mapped to 0, "slight overheating" to 1, and "severe overheating" to 2), calculate the first difference of the sequence to obtain the temperature change rate sequence; calculate the sliding window variance of the sequence to obtain the temperature fluctuation sequence. Perform similar processing for humidity state labels and vibration intensity level labels. Then, key features are extracted from these derived sequences: for example, a consistently positive rate of temperature change with acceleration exceeding a threshold indicates an accelerating temperature increase; periodic impact patterns appear in the vibration level sequence, indicating that the vibration has specific frequency characteristics. These features are combined to generate a structured description of the abnormal change trend, Trend_1, such as "During time period 1, the temperature state increased from normal to severely overheated at an increasing rate, with a positive acceleration; the humidity state remained within the normal range but showed a weak negative correlation fluctuation opposite to the temperature change; the vibration level changed from a stable level 2 state to a periodic impact pattern of levels 3 to 4, with an impact interval of approximately 15 to 20 seconds."

[0107] Step S156: Combine and encapsulate the time period of the abnormal state occurrence, the abnormal trajectory segment, the abnormal environmental state marker subset, the geographical area covered by the abnormal state, and the description of the abnormal change trend to generate the spatiotemporal distribution characteristics of the abnormal state.

[0108] All the time periods of abnormal states, the corresponding abnormal trajectory segments for each time period, the subset of abnormal environmental state markers for each time period, the geographical area for each time period, and the description of the abnormal change trends for each time period are organized chronologically to form a structured data object. For example, for time period 1, a record item Item_1 is generated, which equals (time period 1, Traj_seg_1, Env_subset_1, Area_1, Trend_1); for time period 2, Item_2 is generated, which equals (time period 2, Traj_seg_2, Env_subset_2, Area_2, Trend_2). All these record items are arranged chronologically to constitute the spatiotemporal distribution characteristics of abnormal states. This characteristic fully describes when, where, and what kind of environmental anomalies with complex internal relationships occurred, as well as the multidimensional evolutionary patterns of the anomalies.

[0109] Step S157: Generate a set of early warning instructions for abnormal medical cold chain logistics transportation based on the abnormal state type identifier and the spatiotemporal distribution characteristics of the abnormal state.

[0110] Step S1571: Parse the abnormal state type identifier and search for the warning level code corresponding to the abnormal state type identifier from the preset mapping relationship library. The warning level code is used to indicate the severity level of the abnormal state.

[0111] The abnormal state type identifier output in step S151 is analyzed, assuming it is "progressive failure of refrigeration unit accompanied by unbalanced heat load in the carriage". A search is performed from a pre-defined mapping database. This database is a multi-dimensional lookup table built based on historical accident severity statistics and drug loss risk assessment. It predefines the correspondence between each composite abnormality type and multiple warning levels, including: drug quality impact level, emergency response time window level, and economic loss risk level. For example, the drug quality impact level corresponding to "progressive failure of refrigeration unit accompanied by unbalanced heat load in the carriage" is "high impact" (may lead to a large-scale reduction in drug efficacy), the emergency response time window level is "urgent" (intervention must be carried out within 2 hours), and the economic loss risk level is "significant loss". Combining these three dimensions, a pre-defined fusion rule (e.g., taking the highest level or a weighted average) is used to obtain the final warning level code, "Level 1 Warning" (the highest level).

[0112] Step S1572: Extract the start and end time points of the time period during which the abnormal state occurs from the spatiotemporal distribution features of the abnormal state, and extract the start and end geographical coordinates of the abnormal trajectory segment from the spatiotemporal distribution features of the abnormal state as the warning trigger location coordinates.

[0113] From the spatiotemporal distribution characteristics of the abnormal state, extract the start and end times of the first abnormal time period (i.e., the earliest occurring time period). For example, from time period 1, extract the start time T_start_alarm as T_800 and the end time T_end_alarm as T_1200. Simultaneously, from the abnormal trajectory segment Traj_seg_1 corresponding to this time period, extract the geographic coordinates of the first trajectory point Q_800 as the starting warning position P_start_alarm, and extract the geographic coordinates of the last trajectory point Q_1200 as the ending warning position P_end_alarm.

[0114] Step S1573: Extract a subset of abnormal environmental state markers from the spatiotemporal distribution features of the abnormal state, and generate an environmental state abnormal evolution description vector based on the change sequence of temperature state markers, humidity state markers and vibration intensity level markers in the subset of abnormal environmental state markers during the time period of the abnormal state occurrence.

[0115] From the spatiotemporal distribution characteristics of abnormal states, an abnormal environmental state label subset Env_subset_1 corresponding to the first abnormal time period is extracted. This subset contains temperature state labels, humidity state labels, and vibration intensity level labels for each time step from time step 800 to 1200. First, the discrete state labels are converted into numerical vectors according to a pre-defined complex encoding rule. For the temperature state label, not only is its category encoded (e.g., "normal" is 0, "slight overheating" is 1, and "severe overheating" is 2), but its trend characteristics are also encoded. For example, the direction of change is encoded by calculating the difference between consecutive time steps (increasing is +1, stable is 0, decreasing is -1), and the volatility is encoded by calculating the variance within a local window (high volatility is 1, low volatility is 0). Similar multivariate encoding rules are established for the humidity state label and the vibration intensity level label. In this way, the original triples of each time step are expanded into a higher-dimensional numerical vector, such as an 8-dimensional vector, containing the original category encoding, the direction of change encoding, the volatility encoding, etc. Then, these high-dimensional vectors from all time steps within the time period are stacked in chronological order to form a two-dimensional numerical matrix with 401 rows (time steps) and 8 columns. This matrix is ​​then flattened to form a one-dimensional sequence, which serves as the Evo_Vector, describing the abnormal evolution of the environmental state, with dimensions 401 x 8.

[0116] Step S1574: Combine the warning level code, the start time point, the end time point, the warning trigger location coordinates, and the environmental state abnormal evolution description vector to generate a basic warning instruction unit for the currently detected abnormal state.

[0117] The information obtained in steps S1571 to S1673 is combined as follows: the warning level is coded as "Level 1 Warning", the start time is T_start_alarm, the end time is T_end_alarm, the warning trigger location coordinates are (P_start_alarm, P_end_alarm), and the evolution description vector is Evo_Vector. These five parts are serialized and packaged according to a predefined data encapsulation format to form a basic warning instruction unit Unit_base.

[0118] Step S1575: Determine whether the abnormal state type identifier belongs to a preset chain reaction abnormal type set. If the abnormal state type identifier belongs to the chain reaction abnormal type set, predict the downstream logistics transfer node units that the abnormal state may affect based on the logistics transfer node topology connection network, and generate a predictive early warning instruction unit for the downstream logistics transfer node units.

[0119] Check whether the abnormal state type identifier "Progressive failure of refrigeration unit accompanied by unbalanced heat load in the carriage" output in step S151 is in a preset set of chain reaction abnormal types. This set of chain reaction abnormal types is a composite abnormal category system constructed based on post-event analysis of historical transportation accident data and the physical mechanism of pharmaceutical cold chain transportation. It includes composite abnormal types that have a clear causal propagation path or a high probability of causing secondary abnormalities. "Progressive failure of refrigeration unit accompanied by unbalanced heat load in the carriage" is pre-classified into this set because it is highly likely to lead to a series of subsequent secondary abnormalities: First, the uneven temperature distribution in the carriage will intensify, leading to temperature exceeding limits in local hot spots; temperature exceeding limits may cause the performance of some pharmaceutical packaging materials to deteriorate, releasing trace amounts of moisture, thus causing an abnormal increase in humidity; the abnormal increase in humidity coupled with continuous vibration may lead to a decrease in the sealing of pharmaceutical packaging, ultimately causing a risk of pharmaceutical contamination.

[0120] Prediction is performed based on the logistics transfer node topology network constructed in step S146. The current anomaly occurs between nodes A and C in the logistics transfer node sequence, and the vehicle is traveling from node A to node C. Attribute information of node C is obtained from the network structure, including its type as a "regional distribution center," its temperature control capability level as "Level 2," the historical average dwell time of similar medicines passing through this node as "several hours," and the historical statistical index of temperature field uniformity in the internal storage area of ​​node C as "standard deviation 0.3 degrees Celsius." Simultaneously, the attribute information of the downstream node of node C, node E, is obtained. Node E's type is a "last-mile delivery station," its temperature control capability level is "Level 3," its storage space type is a "small cold storage," and its historical temperature and humidity fluctuation record shows a "frequency of temperature exceedance events" of "several times per year."

[0121] Based on the above information, a predictive early warning instruction unit, Unit_pred_C, is generated for node C. This unit contains the following specific contents:

[0122] First, a sequence of predicted anomaly types is generated. This sequence is a vector containing the possible secondary anomaly types and their probabilities of occurrence. For example, the first element is "deterioration of temperature field uniformity," with an occurrence probability P_temp_field calculated based on the refrigeration failure level parameter, node C temperature control level parameter, and node C's historical temperature field uniformity index. P_temp_field is equal to the weighted linear combination of the sigmoid function applied to the above parameters. The second element is "temperature exceeding limits in local hotspot areas," with an occurrence probability P_hotspot calculated based on the probability of deterioration of temperature field uniformity and the cargo packing density parameter inside node C (extracted from historical data). The third element is "humidity anomaly caused by moisture absorption of packaging materials," with an occurrence probability P_humidity calculated based on the probability of temperature exceeding limits in local hotspot areas and the basic humidity parameter of node C's environment (obtained from meteorological data).

[0123] Second, the estimated time window for arrival at node C is an interval [T_arrive_C_early, T_arrive_C_late]. It is calculated based on the spatial distance between the current vehicle position and the actual passing coordinates of node C, the current average vehicle speed, and the historical travel time fluctuation range of the same route. For example, T_arrive_C_early is the current time plus 3.5 hours, and T_arrive_C_late is the current time plus 4.2 hours.

[0124] Third, the recommended measure code set contains specific operational suggestions for different functional positions at Node C. For example, the measure code for temperature control technicians is "Start the standby refrigeration unit at Node C in advance and preheat to the set temperature range, while checking the calibration status of the temperature sensor array"; the measure code for warehouse management personnel is "Clear a dedicated isolation area before the arrival of medicines, which must have independent temperature monitoring and recording functions"; and the measure code for quality inspection personnel is "Prepare portable temperature and humidity recorders, packaging integrity testing equipment, and sampling tools, and conduct online testing immediately after the medicines arrive."

[0125] Similarly, a predictive warning instruction unit, Unit_pred_E, is generated for node E. This predictive warning instruction unit includes:

[0126] First, a predicted anomaly type identifier sequence is generated, which is further propagated based on the prediction results at node C. For example, the first element is "decreased drug packaging seal," with an occurrence probability P_seal calculated based on the humidity anomaly probability in the node C prediction results, the transportation distance parameter from node E to node C, and the historical road surface grade index of that road segment. The second element is "drug contamination risk," with an occurrence probability P_contamination calculated based on the probability of decreased packaging seal and the cleanliness grade parameter of the storage environment at node E (obtained from the facility archive).

[0127] Second, the expected time window for arrival at node E is calculated based on the expected time window for arrival at node C, the spatial distance between node C and node E, and the average dwell time at node C. For example, T_arrive_E_early is T_arrive_C_early plus the average dwell time at node C (4 hours) plus the travel time, and T_arrive_E_late is T_arrive_C_late plus the maximum dwell time at node C (6 hours) plus the travel time.

[0128] Third, a set of recommended measures codes is provided. For example, the measure code for the receiving personnel at node E is "Prepare a dedicated isolated storage area and increase the sampling inspection ratio of this batch of drugs before warehousing to three times the standard ratio, with a focus on testing packaging sealing and water activity"; the measure code for the quality traceability personnel is "Create an abnormal batch tracking tag in the quality traceability system and automatically associate it with the entire process of sensor data summary for this transportation, so as to facilitate subsequent quality investigations".

[0129] Step S1576: Summarize and sort all the generated basic early warning instruction units and predictive early warning instruction units according to time sequence and spatial correlation to form the pharmaceutical cold chain logistics transportation anomaly early warning instruction set.

[0130] The basic early warning instruction unit (Unit_base) is aggregated with all predictive early warning instruction units (Unit_pred_C and Unit_pred_E). Following chronological order, Unit_base corresponds to the most immediate anomaly, being the earliest and listed first; Unit_pred_C corresponds to the next upcoming node, being the next in time; and Unit_pred_E corresponds to a further downstream node, being the last in time. Spatially, these are associated with their corresponding node identifiers, along with path distance information between nodes, ultimately forming an early warning instruction set list. Each element in this list is a structured early warning instruction unit, containing complete early warning levels, spatiotemporal information, anomaly evolution characteristics, predictive information, and specific recommended measures. This instruction set can be serialized into a standard format data packet and sent via encrypted communication links to the monitoring center's large-screen display system, the transport driver's mobile terminal application, the early warning receiving devices of on-site management personnel at nodes C and E, and the quality management department system of the drug recipient. This allows all parties to take coordinated actions in advance based on the early warning level and predictive information, minimizing drug loss and quality risks.

[0131] For example, the method further includes:

[0132] Step S210: Obtain multiple historical original sensor tracking data samples collected during the historical pharmaceutical cold chain logistics transportation process, and label each of the historical original sensor tracking data samples with the corresponding historical full-process status tracking feature tensor ground value and historical abnormal status type identifier ground value.

[0133] During the model training phase, a large number of historical raw sensor tracking data samples were first obtained from the historical database. Each sample contains a subsequence of temperature sensor parameters, a subsequence of humidity sensor parameters, a subsequence of vibration sensor parameters, and a subsequence of geolocation parameters for a complete pharmaceutical cold chain transportation process. These samples cover diverse scenarios with different seasons, different transportation routes, different types of medicines, and different transportation durations.

[0134] Meanwhile, domain experts, through in-depth post-event analysis and joint assessment with multi-dimensional data, labeled each historical sample with the true value of the full-process state tracking feature tensor under the corresponding ideal state, as well as the true value of the abnormal state type identifier that actually occurred during the entire transportation process.

[0135] The abnormal state type identifier includes, but is not limited to, the following complex abnormal types: For example, the "temperature gradient anomaly accompanied by high-frequency low-amplitude vibration coupling" type, characterized by a continuous and non-linear upward trend in temperature within a specific time window, while vibration sensors capture persistent micro-vibrations within a specific frequency range. This composite state often indicates a progressive failure of the refrigeration unit in the refrigerated truck. Another example is the "path deviation accompanied by periodic temperature fluctuations" type, characterized by the transport vehicle deviating from the planned route, with periodic temperature fluctuations exceeding normal thresholds during the deviation period. This composite state may mean the transport vehicle has been staying in an unauthorized area for an extended period and repeatedly opening and closing the truck doors. Yet another example is the "humidity oversaturation accompanied by impact vibration" type, characterized by a continuous rise in humidity parameters reaching saturation, while multiple instantaneous high-amplitude impact signals appear in the vibration parameters. This composite state often indicates that the truck's sealing has been compromised and it has encountered severe road bumps, potentially leading to moisture damage to the pharmaceutical packaging. The abnormal state type identifier for each historical sample is selected from one or more of the most matching category identifiers from this complex composite abnormality category system.

[0136] Step S220: Generate a historical transport vehicle trajectory line sample based on the historical geographic location parameter subsequence in each of the historical original sensor tracking data samples, and generate a historical multi-dimensional environmental state marker set sample associated with each historical geographic location coordinate point in the historical transport vehicle trajectory line sample based on the historical temperature sensor parameter subsequence, historical humidity sensor parameter subsequence, and historical vibration sensor parameter subsequence in each of the historical original sensor tracking data samples.

[0137] For each historical sample, the complete process of steps S121 to S137 is repeated. Specifically, a sequence of historical geographic location coordinates is extracted from the historical geographic location parameter subsequence. After removing redundant points, a sequence of historical valid geographic location coordinates is obtained. Then, the motion state change description vector is calculated and spliced ​​to generate a historical transport vehicle trajectory line sample. At the same time, instantaneous temperature, humidity, and vibration sample values ​​with the same timestamp as each historical valid geographic location coordinate point are extracted from the historical temperature sensor parameter subsequence, historical humidity sensor parameter subsequence, and historical vibration sensor parameter subsequence.

[0138] Then, according to the complex state classification rules preset in steps S134 to S137, the above instantaneous sampled values ​​are converted into corresponding historical temperature state markers, historical humidity state markers, and historical vibration intensity level markers. The state classification rules are not simple threshold judgments, but rather a complex mapping based on statistical distribution and dynamic benchmarks. For example, the temperature state markers include not only basic categories such as "low temperature deviation," "normal," "slight over-temperature," and "severe over-temperature," but also categories based on changing trends such as "rapid temperature rise," "slow temperature rise," and "periodic temperature fluctuation." The humidity state markers include categories such as "dry," "normal," "humid," and "saturated condensation." The vibration intensity level markers are refined into multiple levels such as "stable," "slight bumps," "moderate impact," "severe impact," and "high-frequency resonance." Finally, the above markers are combined and encapsulated in chronological order to generate a set of historical multi-dimensional environmental state markers associated with each historical geographical coordinate point in the historical transport vehicle trajectory sample.

[0139] Step S230: Construct a historical logistics transfer node topology connection network sample that each historical transport vehicle trajectory line sample passes through. The historical logistics transfer node topology connection network sample includes multiple historical logistics transfer node units and historical directed transfer path edges connecting the historical logistics transfer node units.

[0140] For each historical sample, repeat steps S141 to S146. First, extract all trajectory points marked as neighboring points of historical logistics nodes from the historical transport vehicle trajectory line sample. Cluster them according to the historical node identifier associated with each neighboring point to obtain a set of historical node trajectory points corresponding to each historical logistics transfer node. Then, calculate the mean of the geographical coordinates of all trajectory points in each set to generate the actual passing position coordinates of each historical logistics transfer node in this historical transport. Next, generate a sequence of historical logistics transfer nodes actually traversed by the historical transport based on the order of appearance of these actual passing position coordinates on the time axis. For each pair of adjacent nodes in this sequence, calculate the spatial straight-line distance between their actual passing position coordinates as the historical path distance parameter, and define the directed historical transfer path edge connecting the preceding and following nodes based on this, while attaching the historical path distance parameter as an attribute of the edge. Finally, construct a topological connection network sample of historical logistics transfer nodes corresponding to each historical sample by treating all historical logistics transfer nodes as network nodes and all historical directed transfer path edges as directed edges connecting the nodes. This network sample not only reflects the order of nodes passed through during historical transportation, but also includes spatial distance information between nodes, as well as other attribute information such as node type and node size that may be implied.

[0141] Step S240: Using the historical transport vehicle trajectory line sample, the historical multi-dimensional environmental state marker set sample, and the historical logistics transfer node topology connection network sample as input, and the corresponding historical full-process state tracking feature tensor ground truth value as output, the initial spatiotemporal state feature fusion tracking model is subjected to the first stage of supervised learning training, and the network parameters of the spatiotemporal state feature fusion tracking model are updated to obtain the first intermediate state spatiotemporal state feature fusion tracking model.

[0142] The historical sample data generated in steps S220 and S230—namely, historical transport vehicle trajectory line samples, historical multi-dimensional environmental state marker set samples, and historical logistics transfer node topology connection network samples—are used as input data and fed into the initialized spatiotemporal state feature fusion tracking model. The model performs forward propagation computation according to the architecture described in steps S151 to S156: the trajectory encoding module encodes historical trajectory lines into a sequence of historical spatial trajectory embedding feature vectors; the state encoding module encodes the historical multi-dimensional environmental state marker set into a sequence of historical environmental state embedding feature vectors; and the topology encoding module encodes the historical logistics transfer node topology connection network into a historical network topology embedding feature representation. Then, a historical primary joint feature tensor is generated through feature alignment and concatenation. This tensor is then fused with network topology information through a cross-attention fusion module to generate a historical enhanced feature tensor. Finally, a predicted historical full-process state tracking feature tensor is output through a temporal convolutional coding layer. This predicted historical feature tensor is compared with the ground truth value of the historical full-process state tracking feature tensor obtained in step S210, and the reconstruction error between the two is calculated, for example, using a mean squared error loss function. Using the backpropagation algorithm and optimizer, such as the Adam optimizer, all trainable parameters within the model are updated based on the loss function value. After iterative training on tens of thousands of historical samples, the predicted feature tensor output by the model increasingly approximates the labeled ground truth. After completing this stage of training, the first intermediate spatiotemporal state feature fusion tracking model is obtained.

[0143] Step S250: Input the historical transport vehicle trajectory line sample, the historical multi-dimensional environmental state marker set sample, and the historical logistics transfer node topology connection network sample into the first intermediate state spatiotemporal state feature fusion tracking model to generate a predicted full-process state tracking feature tensor, and input the predicted full-process state tracking feature tensor into a preset abnormal state pattern recognition classifier to obtain a predicted abnormal state type identifier.

[0144] The same historical training samples are forward-propagated again using the first intermediate state model to generate a predicted full-process state tracking feature tensor. This predicted feature tensor is then input into the abnormal state pattern recognition classifier described in step S151. The architecture of this classifier is consistent with that described in step S151, including a global average pooling layer, multiple fully connected layers, and a softmax output layer. The classifier processes the input predicted feature tensor and outputs a probability distribution vector. Each element in the vector corresponds to the predicted probability of a composite anomaly category (e.g., "temperature gradient anomaly accompanied by high-frequency low-amplitude vibration coupling," "path deviation accompanied by periodic temperature fluctuations," "humidity oversaturation accompanied by impact vibration," etc.). The category with the highest probability value is selected from this probability distribution to obtain the predicted abnormal state type identifier, such as "temperature gradient anomaly accompanied by high-frequency low-amplitude vibration coupling."

[0145] Step S260: Construct a classification loss function based on the difference between the predicted abnormal state type identifier and the true value of the historical abnormal state type identifier, and construct a reconstruction loss function based on the difference between the predicted full-process state tracking feature tensor and the true value of the historical full-process state tracking feature tensor.

[0146] The classification loss function is calculated using the cross-entropy loss function to measure the difference between the predicted probability distribution of the composite anomaly category and the true anomaly category labels of historical samples. Specifically, the true values ​​of the historical anomaly state type labels are converted into one-hot encoded vectors, and cross-entropy is calculated with the probability distribution vector output by the classifier to obtain the classification loss value L_cls. Simultaneously, the reconstruction loss between the predicted full-process state tracking feature tensor and the true values ​​of the historical full-process state tracking feature tensor is calculated again, for example, using mean squared error loss, to obtain the reconstruction loss value L_rec. These two loss functions measure the quality of the model output from different perspectives.

[0147] Step S270: The classification loss function and the reconstruction loss function are weighted and combined to generate a joint optimization objective function. The network parameters of the first intermediate spatiotemporal state feature fusion tracking model are then jointly optimized and trained in the second stage using the backpropagation algorithm and optimizer until the joint optimization objective function converges, thus obtaining the trained spatiotemporal state feature fusion tracking model.

[0148] The classification loss L_cls and reconstruction loss L_rec are weighted and summed according to preset weight coefficients to generate a joint optimization objective function L_total, which equals α multiplied by L_rec plus β multiplied by L_cls. The values ​​of α and β are dynamically adjusted based on actual training performance. For example, in the initial stage, α can be set to 0.9 and β to 0.1, allowing the model to prioritize feature reconstruction. As training progresses, the ratio of α and β is gradually adjusted, for example, eventually to α to 0.6 and β to 0.4, so that the model simultaneously considers the accuracy of feature representation and adaptability to downstream anomaly classification tasks. Using this joint objective function, backpropagation and parameter updates are performed on the entire spatiotemporal state feature fusion tracking model (including underlying network layers that may be shared with the classifier). After multiple rounds of iterative optimization, the training process terminates when the value of the joint objective function stabilizes and no longer decreases significantly, or when the preset maximum number of iterations is reached, resulting in the finally trained spatiotemporal state feature fusion tracking model. This model possesses the ability to extract deep fusion features from complex multi-source sensor data.

[0149] In one exemplary embodiment, a pharmaceutical cold chain logistics transportation end-to-end tracking system is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2 As shown, the pharmaceutical cold chain logistics transportation end-to-end tracking system includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a pharmaceutical cold chain logistics transportation end-to-end tracking method. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the shell of the pharmaceutical cold chain logistics transportation tracking system, or an external keyboard, touchpad, or mouse, etc.

[0150] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for end-to-end tracking of pharmaceutical cold chain logistics transportation, characterized in that, The method includes: The raw sensor tracking data stream is obtained from the sensor cluster deployed inside the transport vehicle during the cold chain logistics transportation of pharmaceuticals within a continuous acquisition time window. The raw sensor tracking data stream includes a temperature sensor parameter subsequence, a humidity sensor parameter subsequence, a vibration sensor parameter subsequence, and a geolocation parameter subsequence marked with acquisition timestamps. The transport vehicle's trajectory is generated based on the movement order of the geographic location coordinates in the geographic location parameter subsequence, and a multi-dimensional environmental state marker set associated with each geographic location coordinate in the transport vehicle's trajectory is generated based on the physical parameters in the temperature sensor parameter subsequence, the humidity sensor parameter subsequence, and the vibration sensor parameter subsequence that have the same timestamp as each geographic location coordinate. Construct a logistics transfer node topology connection network along the trajectory of the transport vehicle. The logistics transfer node topology connection network includes multiple logistics transfer node units and directed transfer path edges connecting the logistics transfer node units. The logistics transfer node units are obtained by matching the geographical coordinates of the transport vehicle's trajectory with a preset logistics transfer node database. The transport vehicle's trajectory line, the multi-dimensional environmental state marker set, and the logistics transfer node topology connection network are input into a pre-constructed spatiotemporal state feature fusion tracking model for joint analysis and processing to generate a pharmaceutical cold chain logistics transportation full-process state tracking feature tensor. Anomaly pattern matching processing is performed on the feature tensor of the entire state tracking of pharmaceutical cold chain logistics transportation to obtain the anomaly type identifier and spatiotemporal distribution characteristics of the anomaly state corresponding to the feature tensor of the entire state tracking of pharmaceutical cold chain logistics transportation. Anomaly warning instruction set for pharmaceutical cold chain logistics transportation is generated based on the anomaly type identifier and the spatiotemporal distribution characteristics of the anomaly state.

2. The method for full-process tracking of pharmaceutical cold chain logistics transportation according to claim 1, characterized in that, The step of generating the transport vehicle trajectory line based on the movement order of geographical coordinate points in the geographic positioning parameter subsequence includes: Multiple geographic location coordinate points arranged in chronological order of collection time are extracted from the geographic location parameter subsequence, and redundant points are removed from the geographic location coordinate points to obtain an effective geographic location coordinate point sequence for constructing the operating trajectory of the transport vehicle. The redundant point removal process is achieved by calculating the displacement distance parameter between adjacent geographic location coordinate points and filtering out geographic location coordinate points whose displacement distance parameter is less than a preset static discrimination threshold. Based on the acquisition timestamp corresponding to each valid geographic location coordinate point in the valid geographic location coordinate point sequence, calculate the time interval parameter and displacement direction angle parameter between adjacent valid geographic location coordinate points; normalize the time interval parameter and displacement direction angle parameter respectively, and generate a motion state change description vector between adjacent valid geographic location coordinate points based on the normalized time interval parameter and displacement direction angle parameter. The motion state change description vector includes a standardized motion duration component and a standardized motion direction change component. The effective geographic location coordinate point sequence is associated and combined with the motion state change description vector between the adjacent effective geographic location coordinate points to generate an initial trajectory segment set with temporal continuity. The initial trajectory segment set consists of consecutive effective geographic location coordinate points in the effective geographic location coordinate point sequence and their corresponding motion state change description vectors. The initial trajectory segment set is subjected to trajectory segment splicing processing. Based on the matching degree of the geographical coordinates of the endpoints of adjacent trajectory segments in the initial trajectory segment set and the continuity parameter calculated based on the motion state change description vector, the adjacent trajectory segments that meet the preset splicing conditions are connected to generate the transport vehicle running trajectory line. For each trajectory point in the operation trajectory line of the transport vehicle, logistics node proximity marking processing is performed. The spatial distance parameter between each trajectory point and all logistics transfer nodes in the preset logistics transfer node database is calculated. The trajectory points whose spatial distance parameter is less than the preset proximity distance threshold are marked as logistics node proximity points. At the same time, the identifier of the logistics transfer node closest to the trajectory point is recorded as the node association identifier. Based on the distribution density and order of the logistics node proximity points on the transport vehicle's trajectory line, the sequence of logistics transfer nodes along the transport vehicle's trajectory line is determined. The sequence of logistics transfer nodes is composed of logistics transfer node identifiers arranged sequentially according to the travel direction of the transport vehicle's trajectory line.

3. The method for full-process tracking of pharmaceutical cold chain logistics transportation according to claim 1, characterized in that, The step of generating a multi-dimensional environmental state marker set associated with each geographical coordinate point in the transport vehicle's trajectory line based on physical parameters in the temperature sensing parameter subsequence, humidity sensing parameter subsequence, and vibration sensing parameter subsequence that have the same timestamp as each geographical coordinate point includes: Extract instantaneous temperature sampling values ​​that are the same as the collection timestamp of each geographical coordinate point from the temperature sensing parameter subsequence, and arrange the instantaneous temperature sampling values ​​according to the order in which the geographical coordinate points appear on the transport vehicle's running trajectory to generate a time-series association list of temperature parameters corresponding to the transport vehicle's running trajectory. Extract the instantaneous humidity sampling value that is the same as the collection timestamp of each of the geographical coordinate points from the humidity sensing parameter subsequence, and arrange the instantaneous humidity sampling values ​​according to the order in which the geographical coordinate points appear on the transport vehicle's running trajectory to generate a humidity parameter time sequence association list corresponding to the transport vehicle's running trajectory. Extract the instantaneous vibration sampling value that is the same as the acquisition timestamp of each of the geographical coordinate points from the vibration sensing parameter subsequence, and arrange the instantaneous vibration sampling values ​​according to the order in which the geographical coordinate points appear on the transport vehicle's running trajectory to generate a vibration parameter time sequence association list corresponding to the transport vehicle's running trajectory. Each instantaneous temperature sample value in the time-series association list of temperature parameters is subjected to temperature state interval mapping processing. According to the preset temperature state division rules, each instantaneous temperature sample value is converted into a corresponding temperature state label. The temperature state label is used to characterize the degree of deviation of the internal ambient temperature of the transport carrier from the required temperature range for drug storage. Each instantaneous humidity sample value in the humidity parameter time-series association list is subjected to humidity state interval mapping processing. Each instantaneous humidity sample value is converted into a corresponding humidity state label according to the preset humidity state division rules. The humidity state label is used to characterize the degree of deviation of the humidity inside the transport carrier from the humidity range required for drug storage. Each instantaneous vibration sample value in the vibration parameter time-series association list is subjected to vibration intensity level mapping processing. According to the preset vibration intensity level classification rules, each instantaneous vibration sample value is converted into a corresponding vibration intensity level label. The vibration intensity level label is used to characterize the level of mechanical impact intensity that the transport vehicle is subjected to during transportation. The temperature status marker, humidity status marker, and vibration intensity level marker are combined and encapsulated according to their association with the geographical coordinate points to generate a multi-dimensional environmental status marker set corresponding to each geographical coordinate point in the trajectory line of the transport vehicle.

4. The method for full-process tracking of pharmaceutical cold chain logistics transportation according to claim 1, characterized in that, The construction of the logistics transfer node topology connection network along the trajectory of the transport vehicle includes: Extract all trajectory points marked as logistics node proximity points from the operation trajectory line of the transport vehicle, and cluster the logistics node proximity points belonging to the same logistics transfer node identifier according to the node association identifier corresponding to each logistics node proximity point to obtain the node trajectory point set corresponding to each logistics transfer node identifier. Calculate the average of the geographical coordinates of all neighboring points of the logistics node in each set of node trajectory points, and generate the actual passing position coordinates of the logistics transfer node unit represented by each logistics transfer node identifier in the running trajectory line of the transport vehicle; Based on the time sequence in which each logistics transfer node unit is passed on the trajectory line of the transport vehicle, the logistics transfer node units are arranged in chronological order to generate a sequence of logistics transfer nodes actually passed by the transport vehicle. For two adjacent logistics transfer node units in the logistics transfer node sequence, the spatial straight-line distance between the two logistics transfer node units is calculated as the path distance parameter based on the actual passing position coordinates corresponding to the previous logistics transfer node unit and the actual passing position coordinates corresponding to the next logistics transfer node unit. Based on the sequential order of adjacent logistics transfer node units in the logistics transfer node sequence, the directed transfer path edge connecting the previous logistics transfer node unit to the next logistics transfer node unit is determined, and the path distance parameter is appended to the directed transfer path edge as edge attribute information. All the logistics transfer node units are treated as network nodes, all the directed transfer path edges are treated as directed edges connecting the network nodes, and the topology connection network of the logistics transfer nodes is constructed by combining the edge attribute information.

5. The method for full-process tracking of pharmaceutical cold chain logistics transportation according to claim 1, characterized in that, The process of inputting the transport vehicle's trajectory line, the multi-dimensional environmental state marker set, and the logistics transfer node topology network into a pre-constructed spatiotemporal state feature fusion tracking model for joint analysis generates a pharmaceutical cold chain logistics transportation end-to-end state tracking feature tensor, including: The transport vehicle's trajectory line is input into the trajectory encoding module of the spatiotemporal state feature fusion tracking model. The sequence of geographical location coordinates in the transport vehicle's trajectory line is processed by spatial location embedding representation to generate a spatial trajectory embedding feature vector sequence corresponding to the transport vehicle's trajectory line. The multi-dimensional environmental state label set is input into the state encoding module of the spatiotemporal state feature fusion tracking model. The temperature state label, humidity state label and vibration intensity level label corresponding to each geographic location coordinate point in the multi-dimensional environmental state label set are jointly embedded and represented to generate an environmental state embedding feature vector corresponding to each geographic location coordinate point. All environmental state embedding feature vectors are arranged in the order of appearance of the geographic location coordinate points to generate an environmental state embedding feature vector sequence. The logistics transfer node topology connection network is input into the topology coding module of the spatiotemporal state feature fusion tracking model. The logistics transfer node units and directed transfer path edges in the logistics transfer node topology connection network are processed by graph structure embedding representation to generate the network topology embedding feature representation corresponding to the logistics transfer node topology connection network. The network topology embedding feature representation includes the node embedding feature vector of each logistics transfer node unit and the edge embedding feature vector of each directed transfer path edge. The spatial trajectory embedded feature vector sequence and the environmental state embedded feature vector sequence are aligned and spliced ​​on the time axis to generate a primary joint feature tensor that integrates spatiotemporal information. Each time step of the primary joint feature tensor contains both spatial location information and environmental state information at the corresponding time point. The primary joint feature tensor and the network topology embedding feature representation are input into the cross-attention fusion module of the spatiotemporal state feature fusion tracking model. The cross-attention fusion module uses the primary joint feature tensor as the query basis and the network topology embedding feature representation as the key value basis. It calculates the association weight between the features of each time step in the primary joint feature tensor and each node and edge in the logistics transfer node topology connection network through a multi-head cross-attention mechanism. Based on the association weight, it performs weighted aggregation on the network topology embedding feature representation to generate an enhanced feature tensor with network topology context information. The enhanced feature tensor is input into the temporal convolutional coding layer of the spatiotemporal state feature fusion tracking model. The temporal convolutional coding layer performs multi-scale temporal receptive field feature extraction processing on the enhanced feature tensor through multiple temporal convolutional kernels with different dilation rates to generate the full-process state tracking feature tensor of the pharmaceutical cold chain logistics transportation.

6. The method for full-process tracking of pharmaceutical cold chain logistics transportation according to claim 5, characterized in that, The step of aligning and concatenating the spatial trajectory embedded feature vector sequence and the environmental state embedded feature vector sequence on the time axis to generate a primary joint feature tensor that fuses spatiotemporal information includes: Obtain the original acquisition timestamp corresponding to each spatial trajectory embedding feature vector in the spatial trajectory embedding feature vector sequence, and obtain the original acquisition timestamp corresponding to each environmental state embedding feature vector in the environmental state embedding feature vector sequence; Based on the original acquisition timestamp, the spatial trajectory embedded feature vector sequence and the environmental state embedded feature vector sequence are aligned and verified on the time axis to confirm that the spatial trajectory embedded feature vector sequence and the environmental state embedded feature vector sequence have a one-to-one correspondence in the time dimension. The spatial trajectory embedding feature vector and the environmental state embedding feature vector at each time point after alignment are concatenated and combined in the feature channel dimension to generate a spatiotemporal combination feature vector corresponding to each time point. The spatiotemporal combined feature vectors corresponding to all time points are stacked in chronological order to form a two-dimensional feature matrix as the primary joint feature tensor. The row dimension of the two-dimensional feature matrix corresponds to the time step, and the column dimension corresponds to the total number of feature channels after concatenation. The primary joint feature tensor is subjected to feature dimension normalization processing, and the numerical range of each feature channel in the primary joint feature tensor is adjusted to the preset model input range to generate a normalized primary joint feature tensor with a unified scale representation.

7. The method for full-process tracking of pharmaceutical cold chain logistics transportation according to claim 5, characterized in that, The step of inputting the primary joint feature tensor and the network topology embedding feature representation into the cross-attention fusion module of the spatiotemporal state feature fusion tracking model to generate an enhanced feature tensor with network topology context information includes: The primary joint feature tensor is linearly mapped to the query matrix required by the multi-head cross-attention mechanism. The number of rows in the query matrix corresponds to the number of time steps of the primary joint feature tensor, and the number of columns in the query matrix corresponds to the preset query feature dimension. All node embedding feature vectors and edge embedding feature vectors contained in the network topology embedding feature representation are integrated and linearly mapped to the key matrix and value matrix required by the multi-head cross-attention mechanism. The number of rows in the key matrix and the value matrix corresponds to the total number of node embedding feature vectors and edge embedding feature vectors, and the number of columns in the key matrix and the value matrix corresponds to the preset key feature dimension and value feature dimension, respectively. Perform a dot product operation between the query matrix and the key matrix to obtain an attention score matrix. Each element in the attention score matrix represents the original score of the association strength between a feature at a time step in the primary joint feature tensor and a node or an edge in the network topology embedding feature representation. The attention score matrix is ​​scaled and normalized using softmax to generate an attention weight matrix. The sum of the elements in each row of the attention weight matrix is ​​one, which is used to represent the distribution of the degree of attention of the features at each time step to all nodes and edges. The attention weight matrix and the value matrix are weighted and summed to obtain a weighted network topology context feature matrix. The number of rows in the weighted network topology context feature matrix is ​​the same as the number of time steps in the primary joint feature tensor, and the number of columns is the same as the dimension of the value feature. The primary joint feature tensor and the weighted network topology context feature matrix are concatenated and fused along the feature channel dimension to generate the enhanced feature tensor with network topology context information.

8. The method for full-process tracking of pharmaceutical cold chain logistics transportation according to claim 1, characterized in that, The abnormal state pattern matching processing of the feature tensor for the entire state tracking of pharmaceutical cold chain logistics transportation is performed to obtain the abnormal state type identifier and the spatiotemporal distribution characteristics of the abnormal state corresponding to the feature tensor for the entire state tracking of pharmaceutical cold chain logistics transportation, including: The pharmaceutical cold chain logistics transportation status tracking feature tensor is input into a pre-trained abnormal state pattern recognition classifier. The abnormal state pattern recognition classifier performs global feature extraction and pattern classification processing on the pharmaceutical cold chain logistics transportation status tracking feature tensor and outputs the abnormal state type identifier to which the pharmaceutical cold chain logistics transportation status tracking feature tensor belongs. Simultaneously, the feature tensor of the entire state tracking of the pharmaceutical cold chain logistics transportation is input into the pre-trained anomaly spatiotemporal localization network. The anomaly spatiotemporal localization network performs time-step feature analysis on the feature tensor of the entire state tracking of the pharmaceutical cold chain logistics transportation to generate an anomaly confidence score sequence corresponding to each time step. Based on the abnormal confidence score of each time step in the abnormal confidence score sequence, the continuous time step intervals in which the abnormal confidence score exceeds the preset abnormal judgment threshold are selected, and the time range corresponding to the continuous time step intervals on the operating trajectory line of the transport vehicle is determined as the time period of abnormal state occurrence. Based on the time period of the abnormal state occurrence, the trajectory segment corresponding to the time period of the abnormal state occurrence is extracted from the operating trajectory line of the transport vehicle as the abnormal trajectory segment, and the environmental state marker subset corresponding to the time period of the abnormal state occurrence is extracted from the multi-dimensional environmental state marker set as the abnormal environmental state marker subset. The geographical area covered by the abnormal state during the time period of the abnormal state is calculated based on the sequence of geographical coordinate points in the abnormal trajectory segment, and a description of the abnormal change trend of the environmental state during the time period of the abnormal state is generated based on the distribution of each state marker in the subset of abnormal environmental state markers. The time period of the abnormal state, the abnormal trajectory segment, the subset of abnormal environmental state markers, the geographical area covered by the abnormal state, and the description of the abnormal change trend are combined and encapsulated to generate the spatiotemporal distribution characteristics of the abnormal state.

9. The method for full-process tracking of pharmaceutical cold chain logistics transportation according to claim 1, characterized in that, The step of generating a set of early warning instructions for pharmaceutical cold chain logistics transportation anomalies based on the anomaly type identifier and the spatiotemporal distribution characteristics of the anomaly includes: The abnormal state type identifier is parsed, and the warning level code corresponding to the abnormal state type identifier is searched from the preset mapping relationship library. The warning level code is used to indicate the severity level of the abnormal state. The start and end times of the time period during which the abnormal state occurs are extracted from the spatiotemporal distribution features of the abnormal state, and the starting and ending geographical coordinates of the abnormal trajectory segments are extracted from the spatiotemporal distribution features of the abnormal state as early warning trigger location coordinates. Extract a subset of abnormal environmental state markers from the spatiotemporal distribution features of the abnormal state, and generate an environmental state abnormal evolution description vector based on the change sequence of temperature state markers, humidity state markers and vibration intensity level markers in the subset of abnormal environmental state markers during the time period of the abnormal state occurrence. The warning level code, the start time point, the end time point, the warning trigger location coordinates, and the environmental state abnormal evolution description vector are combined to generate a basic warning instruction unit for the currently detected abnormal state. Determine whether the abnormal state type identifier belongs to a preset chain reaction abnormal type set. If the abnormal state type identifier belongs to the chain reaction abnormal type set, predict the downstream logistics transfer node units that the abnormal state may affect based on the logistics transfer node topology connection network, and generate a predictive early warning instruction unit for the downstream logistics transfer node units. All the generated basic early warning instruction units and predictive early warning instruction units are summarized and sorted according to time sequence and spatial correlation to form the pharmaceutical cold chain logistics transportation anomaly early warning instruction set.

10. A pharmaceutical cold chain logistics transportation end-to-end tracking system, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the pharmaceutical cold chain logistics transportation end-to-end tracking method according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images