Intelligent warehouse-oriented inventory whole-process ai tracing method and system

Abstract

The application provides a kind of storage whole process AI tracing method and system for wisdom warehousing, it is related to wisdom warehousing technical field, including by collecting multi-source data and dividing space-time slice, extract state evolution trajectory and construct state transition matrix to identify abnormal node;With abnormal node as root, build multi-level traceability tree, calculate node contribution intensity to determine root cause set;Based on root cause set, build simulation engine to generate space-time diffusion path, and judge whether other storage object trajectory intersects with it to get risk object;Finally, according to root cause and risk object, generate and execute control instruction.The application realizes efficient and accurate traceability and risk diffusion prediction of warehousing anomaly, improves the intelligent level and risk prevention and control ability of warehousing management.

Description

Technical Field

[0001] This invention relates to the field of smart warehousing technology, and in particular to an AI-based method and system for full-process inventory traceability in smart warehousing. Background Technology

[0002] In the field of smart warehousing, inventory traceability management is a key link in ensuring inventory accuracy, improving operational efficiency, and enhancing security. This typically relies on collecting and analyzing various data generated during the warehousing process to monitor inventory status and trace abnormal events.

[0003] Conventional traceability methods primarily rely on linear analysis of operational log data. For example, by recording key operational nodes such as inventory receipt, issuance, and relocation, a time-series-based chain of records is formed. When anomalies in inventory quantity or discrepancies in status are detected, the system traces this chain backwards to pinpoint the potentially problematic operational step. Another common approach is to incorporate sensor data, such as RFID, weight sensors, or visual recognition data, to assist in verifying the authenticity of operational records or automatically capturing partial status information.

[0004] However, existing methods often focus on analyzing discrete events or single-dimensional data, lacking modeling and understanding of the complete state evolution of inventory in a continuous spatiotemporal dimension. Operation logs can only reflect discrete event points recorded manually or triggered by the system, while sensor data is often processed independently, failing to be deeply integrated and correlated with business process logs within a unified spatiotemporal framework. This results in fragmented and disjointed descriptions of inventory state changes, making it difficult to depict the complete trajectory of its continuous and dynamic evolution.

[0005] Due to the lack of modeling for the continuous evolution trajectory of states, tracing the root cause of anomalies often only pinpoints directly related, explicit operational nodes. These methods struggle to effectively identify subtle, probabilistic anomalies during state transitions, and are even less capable of systematically analyzing complex operational interactions between multiple objects, the implicit influence of environmental factors, and the potential spread paths of anomalies across time and space. Therefore, conventional tracing methods lack depth and breadth, root cause localization may remain superficial, and they cannot proactively predict and prevent the spread of risks to similar inventory or adjacent areas, thus hindering the level of intelligent warehouse management and proactive risk control. Summary of the Invention

[0006] This invention provides an AI-based inventory traceability method and system for smart warehousing, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides an AI-based inventory traceability method for smart warehousing, comprising:

[0008] Collect multi-source sensor data and operation log data of inventory objects in the warehousing process;

[0009] Multi-source sensor data and operation log data are divided into multiple spatiotemporal slices according to the time and space dimensions. The state evolution trajectory of objects in each spatiotemporal slice is extracted. A state transition matrix is ​​constructed based on the state evolution trajectory. Abnormal state transition nodes are identified by analyzing the abnormal fluctuations of the transition probability in the state transition matrix.

[0010] Using the abnormal state transition node as the root node, a multi-level traceability tree is constructed based on the inventory object flow path, the operation interaction relationship between multiple inventory objects, and the influence of environmental factors. The contribution intensity of each node to the abnormal state transition node is calculated in the multi-level traceability tree, and the root cause node set is determined.

[0011] A state evolution simulation engine is constructed based on the root cause node set. A spatiotemporal diffusion path is generated by reproducing the state evolution process of the root cause node set. The state evolution trajectories of other inventory objects are input into the state evolution simulation engine to determine whether they intersect with the spatiotemporal diffusion path, thereby obtaining the risk inventory objects.

[0012] Control instructions are generated and executed based on the root cause node set and risk inventory objects.

[0013] In one optional embodiment, the multi-source sensor data and operation log data are divided into multiple spatiotemporal slices according to the time and space dimensions, and the state evolution trajectory of objects in each spatiotemporal slice is extracted, including:

[0014] Based on the timestamp information and spatial location information of inventory objects in multi-source sensor data and operation log data, multiple spatiotemporal slice units are divided in the spatiotemporal coordinate system according to preset time intervals and preset spatial ranges, and each data item is mapped to the corresponding spatiotemporal slice unit.

[0015] Within each spatiotemporal slice unit, the state feature sequence of the inventory object is extracted, the state entropy value of the state feature sequence within the spatiotemporal slice unit is calculated, and the moment when the state entropy value changes abruptly is marked as the state transition point.

[0016] By connecting the state transition points of the same inventory object in adjacent spatiotemporal slice units in chronological order, the state evolution trajectory of the inventory object is constructed. The state evolution trajectory records the state entropy value, timestamp, and spatial location of each state transition point.

[0017] In one optional embodiment, constructing a state transition matrix based on the state evolution trajectory, and identifying anomalous state transition nodes by analyzing anomalous fluctuations in the transition probabilities within the state transition matrix includes:

[0018] Extract the state feature vectors of each state transition point in the state evolution trajectory, calculate the distance metric between the state feature vectors, divide the state feature vectors with a distance metric less than a preset distance threshold into the same node cluster, assign the same state identifier to the state feature vectors belonging to the same node cluster, and obtain the state identifier sequence.

[0019] The transition frequency between each state identifier in the statistical state identifier sequence is counted, and a state transition matrix is ​​constructed. For each transition probability in the state transition matrix, a time-series tracking sequence of the transition probability is established. By periodically decomposing the time-series tracking sequence, the baseline periodic component and abnormal fluctuation component of the transition probability are extracted.

[0020] Calculate the energy spectral density of the abnormal fluctuation component, mark the transition probability when the energy spectral density exceeds the preset energy threshold as the abnormal transition probability, and extract the starting state identifier and target state identifier corresponding to the abnormal transition probability.

[0021] Locate the state transition point from the initial state identifier to the target state identifier in the state evolution trajectory, extract the operation log data within a preset time range before and after the state transition point, and mark the state transition point as an abnormal state transition node.

[0022] In one optional embodiment, using the abnormal state transition node as the root node, a multi-level traceability tree is constructed based on the inventory object flow path, the operational interaction relationships between multiple inventory objects, and the influence of environmental factors, including:

[0023] Set the abnormal state transfer node as the root node of the multi-level traceability tree, extract the timestamp of the abnormal state transfer node, trace back a preset time window with the timestamp as the endpoint, obtain the flow path of the inventory object, calculate the state fluctuation amplitude of each path segment in the flow path, filter the path segments whose state fluctuation amplitude exceeds the preset fluctuation threshold, extract the corresponding starting spatial position and ending spatial position as the first-level child node and connect them to the root node.

[0024] Traverse each first-level child node, obtain the operation log data of the corresponding spatial location within the corresponding time period, extract the identifiers of other inventory objects that have operation associations with the inventory object from the operation log data, calculate the operation association strength, and connect other inventory objects whose operation association strength exceeds the preset association threshold as second-level child nodes to the corresponding first-level child nodes.

[0025] Traverse the second-level child nodes to obtain the environmental parameter monitoring data of the spatial location of the corresponding other inventory objects at the time of operation association, calculate the abnormal deviation degree, and connect the environmental factors with abnormal deviation degree exceeding the preset deviation threshold as the third-level child nodes to the corresponding second-level child nodes to complete the construction of the multi-level traceability tree.

[0026] In one optional embodiment, the contribution strength of each node to the anomalous state transition node is calculated in a multi-level source tree, and the root cause node set is determined to include:

[0027] Extract the connection relationships of child nodes at each level in the multi-level source tree, construct a propagation path graph, extract the propagation path from each child node to the root node in the propagation path graph, obtain the state fluctuation amplitude sequence of the nodes on the propagation path, calculate the temporal similarity of the state fluctuation amplitude sequence, and count the number of propagation paths to which the temporal similarity of each child node exceeds the preset similarity threshold to determine the direct contribution value.

[0028] For each level of child node, extract the lower level child nodes connected to each level of child node, obtain the direct contribution value and state fluctuation amplitude of the lower level child node, calculate the causal response delay time between the state fluctuation amplitude of the lower level child node and the state fluctuation amplitude of the root node, and when the causal response delay time is less than the preset delay threshold, add the direct contribution value of the lower level child node to each level of child node to obtain the cumulative contribution intensity.

[0029] In a multi-level source tree, identify node clusters that form a closed-loop structure, extract the state evolution trajectory of each level of child nodes within the node cluster, calculate the mutual information between the state evolution trajectories as the cycle enhancement factor, determine the final contribution intensity based on the cumulative contribution intensity of each level of child nodes within the node cluster and the cycle enhancement factor, and select level child nodes that exceed the preset contribution threshold as root cause nodes to be added to the root cause node set.

[0030] In one optional embodiment, a state evolution simulation engine is constructed based on the root cause node set, and a spatiotemporal diffusion path is generated by reproducing the state evolution process of the root cause node set, including:

[0031] Extract the initial state feature vector and timestamp of each root node in the root node set, obtain the propagation path from each root node to the root node in the multi-level tracing tree, extract the state difference vector between each intermediate node and its adjacent intermediate node on the propagation path, arrange the state difference vectors in time sequence to construct the state transition tensor, and perform tensor decomposition to extract feature components, and construct the state evolution simulation engine.

[0032] The initial state feature vector of each root node is input into the state evolution simulation engine. Based on the feature components, the candidate state feature vector of each intermediate node at each time is deduced. The environmental parameter fluctuation of each intermediate node at the corresponding time is extracted from the candidate state feature vector. The environmental parameter fluctuation is converted into a correction amount and superimposed on the candidate state feature vector to generate the simulated state feature vector of each intermediate node.

[0033] The simulated state feature vectors of each intermediate node are extracted along the propagation path to determine the trigger time and spatial location of the anomaly judgment boundary. The time interval between adjacent trigger times and the spatial span between adjacent spatial locations are calculated. The time interval and spatial span are used as diffusion rate markers to mark each intermediate node. The intermediate nodes are connected in the order of trigger times to generate a spatiotemporal diffusion path.

[0034] In one optional embodiment, the state evolution trajectories of other inventory objects are input into the state evolution simulation engine to determine whether they intersect with the spatiotemporal diffusion path, thus obtaining risky inventory objects including:

[0035] Obtain the inventory objects other than the inventory object corresponding to the root node, extract the state evolution trajectory of the other inventory objects, extract the initial state feature vector, timestamp and spatial position of each state transition point from the state evolution trajectory, input the initial state feature vector into the state evolution simulation engine, infer the candidate state feature vector and corresponding spatial position of the other inventory objects at each time based on the feature components, and construct the predicted evolution trajectory of the other inventory objects.

[0036] Extract the trigger time, spatial location, and diffusion rate identifier of each intermediate node in the spatiotemporal diffusion path. Calculate the spatial distance between the spatial location of the predicted evolution trajectory of other inventory objects at each trigger time and the corresponding spatial location in the spatiotemporal diffusion path. Combine the diffusion rate identifier to calculate the diffusion influence range. When the spatial distance is within the diffusion influence range, mark it as the intersection time.

[0037] For other inventory objects with intersection times, extract the candidate state feature vectors of the predicted evolution trajectories of the other inventory objects at the intersection times, calculate the feature similarity with the simulated state feature vectors of the intermediate nodes at the corresponding intersection times in the spatiotemporal diffusion path, and when the feature similarity exceeds the preset similarity threshold, mark the other inventory object as a risk inventory object and add it to the risk inventory object set.

[0038] A second aspect of this invention provides an AI-powered inventory traceability system for smart warehousing, comprising:

[0039] The data acquisition unit is used to collect multi-source sensor data and operation log data of inventory objects in the warehousing process;

[0040] An anomaly identification unit is used to divide multi-source sensor data and operation log data into multiple spatiotemporal slices according to the time and space dimensions, extract the state evolution trajectory of objects in each spatiotemporal slice, construct a state transition matrix based on the state evolution trajectory, and identify abnormal state transition nodes by analyzing the abnormal fluctuations of the transition probability in the state transition matrix.

[0041] The root cause tracing unit is used to construct a multi-level tracing tree with the abnormal state transfer node as the root node, based on the inventory object flow path, the operation interaction relationship between multiple inventory objects and the influence of environmental factors. In the multi-level tracing tree, the contribution intensity of each node to the abnormal state transfer node is calculated, and the root cause node set is determined.

[0042] The risk prediction unit is used to construct a state evolution simulation engine based on the root cause node set. It generates a spatiotemporal diffusion path by reproducing the state evolution process of the root cause node set. The state evolution trajectory of other inventory objects is input into the state evolution simulation engine to determine whether it intersects with the spatiotemporal diffusion path, thereby obtaining the risky inventory objects.

[0043] The control execution unit is used to generate and execute control instructions based on the root cause node set and the risk inventory object.

[0044] A third aspect of the present invention provides an electronic device, comprising:

[0045] processor;

[0046] Memory used to store processor-executable instructions;

[0047] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0049] In this embodiment of the invention, by dividing the spacetime into spatiotemporal slices and extracting state evolution trajectories, the dynamic change characteristics of inventory objects during the warehousing process can be accurately captured. Analyzing abnormal fluctuations based on the state transition matrix can effectively identify key abnormal nodes that deviate from the normal pattern, significantly improving the sensitivity and timeliness of anomaly detection. A multi-level source tree is constructed starting from the abnormal nodes, integrating multi-dimensional correlation information such as flow paths, object interactions, and environmental factors. By calculating the contribution intensity of each node to the abnormal event, the set of root cause nodes leading to the anomaly can be accurately located. The state evolution simulation engine built based on the root cause set can reproduce the evolution process of abnormal states and generate potential spatiotemporal diffusion paths. By comparing the trajectories of other inventory objects with the diffusion paths, objects with similar risk patterns can be quickly identified. The final generated control instructions are directly associated with the root causes and risk objects, making the disposal measures highly targeted and effective. Executing these instructions can promptly block the spread of abnormal states and eliminate potential hazards. Attached Figure Description

[0050] Figure 1A flowchart illustrating an AI-powered inventory traceability method for smart warehousing.

[0051] Figure 2 This is a flowchart illustrating the process of tracing the origins of state evolution and spatiotemporal diffusion. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0054] Figure 1 This is a flowchart illustrating the AI-powered inventory traceability method for smart warehousing, as described in an embodiment of the present invention. Figure 1 As shown, the AI-based inventory traceability methods for smart warehousing include:

[0055] Collect multi-source sensor data and operation log data of inventory objects in the warehousing process;

[0056] Multi-source sensor data and operation log data are divided into multiple spatiotemporal slices according to the time and space dimensions. The state evolution trajectory of objects in each spatiotemporal slice is extracted. A state transition matrix is ​​constructed based on the state evolution trajectory. Abnormal state transition nodes are identified by analyzing the abnormal fluctuations of the transition probability in the state transition matrix.

[0057] Using the abnormal state transition node as the root node, a multi-level traceability tree is constructed based on the inventory object flow path, the operation interaction relationship between multiple inventory objects, and the influence of environmental factors. The contribution intensity of each node to the abnormal state transition node is calculated in the multi-level traceability tree, and the root cause node set is determined.

[0058] A state evolution simulation engine is constructed based on the root cause node set. A spatiotemporal diffusion path is generated by reproducing the state evolution process of the root cause node set. The state evolution trajectories of other inventory objects are input into the state evolution simulation engine to determine whether they intersect with the spatiotemporal diffusion path, thereby obtaining the risk inventory objects.

[0059] Control instructions are generated and executed based on the root cause node set and risk inventory objects.

[0060] In one optional embodiment, the multi-source sensor data and operation log data are divided into multiple spatiotemporal slices according to the time and space dimensions, and the state evolution trajectory of objects in each spatiotemporal slice is extracted, including:

[0061] Based on the timestamp information and spatial location information of inventory objects in multi-source sensor data and operation log data, multiple spatiotemporal slice units are divided in the spatiotemporal coordinate system according to preset time intervals and preset spatial ranges, and each data item is mapped to the corresponding spatiotemporal slice unit.

[0062] Within each spatiotemporal slice unit, the state feature sequence of the inventory object is extracted, the state entropy value of the state feature sequence within the spatiotemporal slice unit is calculated, and the moment when the state entropy value changes abruptly is marked as the state transition point.

[0063] By connecting the state transition points of the same inventory object in adjacent spatiotemporal slice units in chronological order, the state evolution trajectory of the inventory object is constructed. The state evolution trajectory records the state entropy value, timestamp, and spatial location of each state transition point.

[0064] In one specific implementation, in a warehousing environment, multi-source sensor data and operation log data exhibit a high degree of spatiotemporal coupling. The collected raw data includes various heterogeneous data sources such as tag identification records from RFID readers, environmental monitoring data from temperature and humidity sensors, image sequences from video surveillance equipment, location trajectories from forklift operation systems, and operation records from inbound and outbound management systems. These data have different sampling frequencies in the time dimension and are distributed across different functional areas of the warehouse in the spatial dimension, requiring the use of spatiotemporal slicing technology to achieve a unified data organization framework.

[0065] When performing spatiotemporal slicing on multi-source sensor data and operation log data, the first step is to extract the timestamp and spatial location information from each data record. Timestamps may exist in various formats, including Unix timestamps, standard time formats, or relative time offsets, and need to be uniformly converted to a standard time representation. Spatial location information varies depending on the data source type. Data generated by RFID devices carries the fixed location coordinates of the reader / writer, data generated by forklift equipment includes real-time GPS coordinates or indoor positioning coordinates, and sensor data from shelving areas is associated with fixed storage location codes. These heterogeneous spatial information are then uniformly mapped to the warehouse's three-dimensional coordinate system, establishing a transformation relationship from raw location identifiers to standard spatial coordinates.

[0066] In constructing the spatiotemporal coordinate system, the time dimension is set as the horizontal axis, and the spatial dimension is represented by two-dimensional coordinates in the vertical plane. Determining the preset time interval requires comprehensive consideration of data acquisition frequency and state change rate. If the time interval is too large, it will lead to the loss of state transition details; if the time interval is too small, it will generate a large number of redundant slices, increasing the computational burden. Through statistical analysis of historical data, the average time scale of inventory object state changes is calculated, and the preset time interval is set to 1 / 2 to 1 / 3 of this average time scale. The preset spatial range is divided according to the functional zoning of the warehouse, decomposing the warehouse space into multiple logical areas. Each area corresponds to a spatial slice unit, and the area boundaries are consistent with the shelving layout, work aisles, and functional zoning.

[0067] After completing the gridding of the spatiotemporal coordinate system, all data items are traversed, and their timestamp and spatial location information are read to determine the spatiotemporal slice unit to which the data item belongs. Specifically, the difference between the timestamp of the data item and the start time of the time axis is calculated, divided by a preset time interval, and rounded to obtain the slice index of the time dimension; the spatial coordinates of the data item are compared with the boundary range of each spatial region to determine the spatial region index to which it belongs. By combining the time slice index and the spatial region index, the spatiotemporal slice unit number to which the data item belongs is uniquely determined. A mapping table from data items to spatiotemporal slice units is established to complete the spatiotemporal organization of the data.

[0068] Within each spatiotemporal slice, an inventory object may generate multiple data records, reflecting its state changes within that spatiotemporal range. Extracting the state feature sequence of an inventory object requires extracting key features characterizing its state from multidimensional sensor data. For temperature-sensitive inventory, state features include temperature value and temperature change rate; for location-sensitive inventory, state features include spatial coordinates, movement speed, and dwell time; for operation-sensitive inventory, state features include operation type, operator, and operation duration. These features are arranged chronologically to form the state feature sequence within that spatiotemporal slice.

[0069] State entropy is used to quantify the degree of uncertainty in the state of an inventory object. The feature dimensions in the state feature sequence are discretized, and a histogram of feature value distribution is established. The information entropy of each feature dimension is calculated; it is obtained through the probability distribution of feature values. The more uniform the feature value distribution, the higher the information entropy; the more concentrated the feature value distribution, the lower the information entropy. The information entropies of multiple feature dimensions are weighted and fused, with weighting coefficients determined based on the importance of each feature to the state representation, resulting in a comprehensive state entropy value. This state entropy value reflects the stability of the inventory object's state within that spatiotemporal slice unit. A lower state entropy value indicates that the object is in a stable state, while a higher state entropy value indicates that the object's state is in a changing or uncertain state.

[0070] Identifying state transition points requires detecting abrupt changes in state entropy values ​​over time. The state entropy values ​​of continuous spatiotemporal slices are arranged chronologically to form a state entropy time series. A sliding window method is used to calculate the variance of state entropy values ​​within local intervals; a significant increase in variance indicates drastic fluctuations in state entropy values. Simultaneously, the difference in state entropy values ​​between adjacent spatiotemporal slices is calculated; when the difference exceeds a preset threshold, it is considered a state entropy abrupt change. The threshold is determined through statistical analysis of the range of state entropy fluctuations under normal conditions, typically taken as 2 to 3 times the standard deviation of normal fluctuations. The moment that meets the abrupt change condition is recorded as the state transition point; the spatiotemporal slice corresponding to this moment represents the spatiotemporal location where the state has significantly changed.

[0071] Constructing a state evolution trajectory requires tracking the state transition process of the same inventory object across different spatiotemporal slices. For a specific inventory object, its marked state transition points in each spatiotemporal slice are retrieved in chronological order. Since inventory objects undergo spatial movement during the warehousing process, their state transition points may be distributed across different spatial regions. Using the inventory object's unique identifier, the state transition points scattered across different spatiotemporal slices are associated, establishing connections in ascending order of timestamps. The lines connecting adjacent state transition points constitute a segment of the state evolution trajectory. A complete state evolution trajectory consists of multiple segments, presenting the evolution path of the inventory object in the spatiotemporal coordinate system.

[0072] The information recorded in the state evolution trajectory includes three core elements. The state entropy value describes the degree of uncertainty in the object's state at the transition point, used to quantify the drastic nature of the state change. The timestamp identifies the precise moment the state transition occurred, used to establish the temporal relationship of events. The spatial location marks the physical location where the state transition occurred, used to correlate environmental factors and operational behaviors. These three elements together constitute a complete description of the state transition point, providing foundational data for subsequent state transition matrix construction and anomaly identification. By comparing and analyzing the state evolution trajectories of multiple inventory objects, similarities and differences in state evolution patterns between different objects can be discovered, and abnormal evolution trajectories deviating from the normal pattern can be identified.

[0073] In one optional embodiment, constructing a state transition matrix based on the state evolution trajectory, and identifying anomalous state transition nodes by analyzing anomalous fluctuations in the transition probabilities within the state transition matrix includes:

[0074] Extract the state feature vectors of each state transition point in the state evolution trajectory, calculate the distance metric between the state feature vectors, divide the state feature vectors with a distance metric less than a preset distance threshold into the same node cluster, assign the same state identifier to the state feature vectors belonging to the same node cluster, and obtain the state identifier sequence.

[0075] The transition frequency between each state identifier in the statistical state identifier sequence is counted, and a state transition matrix is ​​constructed. For each transition probability in the state transition matrix, a time-series tracking sequence of the transition probability is established. By periodically decomposing the time-series tracking sequence, the baseline periodic component and abnormal fluctuation component of the transition probability are extracted.

[0076] Calculate the energy spectral density of the abnormal fluctuation component, mark the transition probability when the energy spectral density exceeds the preset energy threshold as the abnormal transition probability, and extract the starting state identifier and target state identifier corresponding to the abnormal transition probability.

[0077] Locate the state transition point from the initial state identifier to the target state identifier in the state evolution trajectory, extract the operation log data within a preset time range before and after the state transition point, and mark the state transition point as an abnormal state transition node.

[0078] In one specific implementation, within a smart warehousing system, inventory items undergo multiple state changes during various stages such as receiving, shelving, picking, and shipping. These state changes exhibit statistically regular patterns. To accurately identify abnormal state transitions, it is necessary to extract state features from the state evolution trajectory and construct a transition model.

[0079] State feature vectors are extracted from each state transition point in the state evolution trajectory. These vectors contain multi-dimensional information such as location coordinates, temperature, humidity, light intensity, operator code, shelf number, and equipment identification. For example, the state transition points for a batch of medicines in cold chain storage include location... ,temperature ,humidity Features such as operator employee number E1023 are used. These features are then normalized to form a state feature vector. , where m is the feature dimension. When calculating the distance metric between state feature vectors, a weighted Euclidean distance is used to reflect the differences in importance of different features. For vectors... and Calculate distance metric value ,in This represents the weight coefficient for the k-th feature. Distance metrics less than a preset distance threshold are considered. The state feature vectors are grouped into the same node cluster, a process implemented using a hierarchical clustering algorithm. For example, at a temperature of 2... Up to 3 The state transition points between and located in cold storage area A are divided into node clusters. .

[0080] When calculating the distance metric between state feature vectors, a weighted Euclidean distance is used to reflect the differences in importance of different features. For vectors and Calculate distance metric value ,in For the first Weight coefficients for each feature dimension Represents the state feature vector The Each feature dimension Represents the state feature vector The Each feature dimension. Distance metrics less than a preset distance threshold. The state feature vectors are grouped into the same node cluster, a process implemented using a hierarchical clustering algorithm. For example, temperature at... to The state transition points between and located in the cold storage area A are divided into node clusters C1.

[0081] Assign the same state identifier to state feature vectors belonging to the same node cluster. Generate a unique state identifier for each node cluster. Where p is the cluster number. Traverse the entire state evolution trajectory, mapping each state transition point to a corresponding state identifier, forming a state identifier sequence. ,in This represents a time index. For example, the status identifier sequence for a cold chain medicine might be: This indicates the state category of the drug at different times.

[0082] Count the transition frequency between state identifiers in the state identifier sequence. Traverse the state identifier sequence and count the transition frequency between states. Transition to state Number of occurrences Construct the state transition matrix. Matrix elements Indicates from state Transition to state The probability is calculated as follows: Where N is the total number of states. For example, from the refrigerated state... Transfer to room temperature The transition probability is This indicates that such a transfer is relatively rare under normal circumstances.

[0083] For each transition probability in the state transition matrix, a time-series tracking sequence of the transition probabilities is established. Then, a specific transition probability is selected from the state transition matrix. The transition probability is repeatedly calculated within different time windows to form a time-series tracking sequence. , where K is the number of observation time windows, and the length of each time window is set to 24 hours or 7 days. This sequence reflects the dynamic changes of a specific state transition pattern over time.

[0084] By performing periodic decomposition on the time-series tracking sequence, the baseline periodic component and abnormal fluctuation component of the transition probability are extracted. A seasonal trend decomposition method is used to decompose the time-series tracking sequence into... ,in As a trend component, it reflects long-term changing trends; It serves as the baseline periodic component, reflecting the periodic pattern; Abnormal fluctuation components reflect fluctuations that deviate from the normal pattern. For example, on a workday, the probability of moving from the waiting-to-pick state to the picking state exhibits periodic peaks. This peak pattern is extracted as the baseline periodic component, while sudden, unexpected moves are identified as abnormal fluctuation components.

[0085] Calculate the energy spectral density of the anomalous fluctuation component. For the anomalous fluctuation component... Perform a Fourier transform to obtain the frequency domain representation. Calculate the energy spectral density The total energy is obtained by integrating over the frequency domain. The energy spectral density exceeds a preset energy threshold. The probability of transfer is marked as an abnormal transfer probability. This threshold is determined based on the statistical distribution of historical data, and is usually set as the 95th quantile of the normal energy distribution. For example, if the transfer probability of transferring directly from a refrigerated state to a high-temperature state suddenly increases, and its abnormal fluctuation component energy significantly exceeds the threshold, it is marked as an abnormal transfer probability.

[0086] Extract the starting state identifier and target state identifier corresponding to the abnormal transition probability. For the marked abnormal transition probability... Record its row index i and column index j in the state transition matrix to obtain the starting state identifier. With target status identifier For example, the probability of abnormal transitions. Corresponding initial state (Refrigerated state) and target state (At high temperatures), such a sudden shift across temperature zones should not occur in normal processes.

[0087] Locate the state transition point from the initial state identifier to the target state identifier in the state evolution trajectory. Backtrack the original state identifier sequence and find all points that satisfy the conditions. and Time Index The physical transition points corresponding to these time points are precisely located in the state evolution trajectory, and information such as the timestamp, spatial location, and related equipment number of the transition points are obtained.

[0088] Extract operation log data within a preset time range before and after the state transition point. The preset time range is set to before the state transition point. After the conversion point For example, a 30-minute window before and after the event. Within this time window, the operation log database is searched to extract all relevant operation records, including barcode scanning records, handling instructions, equipment start / stop records, personnel entry / exit records, and environmental monitoring alarms. These operation logs are analyzed to identify operational behaviors that may lead to abnormal state transitions, such as refrigeration equipment malfunctions, human error, or the mixing of abnormal goods. State transition points that meet the above conditions are marked as abnormal state transition nodes, and the associated operation log summary, abnormal intensity index, and impact range are recorded in the node attributes to provide a data foundation for subsequent source tracing analysis.

[0089] In one optional embodiment, using the abnormal state transition node as the root node, a multi-level traceability tree is constructed based on the inventory object flow path, the operational interaction relationships between multiple inventory objects, and the influence of environmental factors, including:

[0090] Set the abnormal state transfer node as the root node of the multi-level traceability tree, extract the timestamp of the abnormal state transfer node, trace back a preset time window with the timestamp as the endpoint, obtain the flow path of the inventory object, calculate the state fluctuation amplitude of each path segment in the flow path, filter the path segments whose state fluctuation amplitude exceeds the preset fluctuation threshold, extract the corresponding starting spatial position and ending spatial position as the first-level child node and connect them to the root node.

[0091] Traverse each first-level child node, obtain the operation log data of the corresponding spatial location within the corresponding time period, extract the identifiers of other inventory objects that have operation associations with the inventory object from the operation log data, calculate the operation association strength, and connect other inventory objects whose operation association strength exceeds the preset association threshold as second-level child nodes to the corresponding first-level child nodes.

[0092] Traverse the second-level child nodes to obtain the environmental parameter monitoring data of the spatial location of the corresponding other inventory objects at the time of operation association, calculate the abnormal deviation degree, and connect the environmental factors with abnormal deviation degree exceeding the preset deviation threshold as the third-level child nodes to the corresponding second-level child nodes to complete the construction of the multi-level traceability tree.

[0093] In one specific implementation, after identifying the abnormal state transition node, it is necessary to trace the underlying cause of the abnormality. This tracing process is achieved by constructing a multi-level source tree. The root node of the source tree is the detected abnormal state transition node, which records the critical moment when the inventory object transitions from a normal state to an abnormal state. When constructing the source tree, the complete timestamp information corresponding to the abnormal state transition node is first extracted from the spatiotemporal information database. This timestamp accurately marks the time when the abnormality occurred, denoted as […]. .

[0094] Using this timestamp as the endpoint, a preset time window is traced back. The length of this time window is dynamically set according to the characteristics of different types of inventory objects, typically ranging from 2 to 48 hours. For perishable goods, the time window may be shorter, set to 2 to 6 hours; for ordinary industrial products, the time window may be set to 12 to 24 hours; for goods with a long shelf life, the time window can be extended to 48 hours or even longer. (The last sentence appears to be incomplete and possibly refers to a step backward within the time window.) The system retrieves the complete flow path record of the inventory object from the warehouse management system. This record includes all spatial location nodes that the inventory object passes through, the dwell time, the movement trajectory, and the status parameter collection values ​​of each location node.

[0095] For each path segment in the flow path, the fluctuation amplitude of the state parameters of the objects within that path segment is calculated. The calculation of the state fluctuation amplitude uses a combination of standard deviation and mean change rate. Specifically, for a state parameter sequence collected within a certain path segment i... First, calculate the standard deviation of the sequence. Then calculate the relative rate of change of the state values ​​at the beginning and end of the path segment. The state fluctuation amplitude index is obtained by combining the results. ,in and The weighting coefficient is usually determined based on historical data statistics, and common values ​​are... , The amplitude of state fluctuations in each path segment is compared with a preset fluctuation threshold. When comparing, At that time, it was determined that the path segment had significant abnormal changes in its state and needed to be included in the scope of in-depth tracing.

[0096] For these selected key path segments, their starting and ending spatial locations are extracted. Each spatial location is represented in three-dimensional coordinates, including warehouse area number, shelf number, storage location number, and corresponding GPS or indoor positioning coordinates. These spatial locations serve as the first-level child nodes of a multi-level traceability tree, connected to the root node via directed edges. The edges are labeled with the state fluctuation amplitude and time interval information of the path segment, facilitating subsequent quantitative analysis of the contribution of each node to the anomaly.

[0097] After constructing the first-level child nodes, it is necessary to further explore the operational interaction information at these spatial locations. For each first-level child node, based on its corresponding spatial coordinates and time interval, all operation records within that spatiotemporal range are retrieved from the operation log database. The operation log data includes information such as operation type, operation time, involved inventory object identifier, operating device identifier, and operator identifier. By parsing these operation logs, other inventory objects that have operational associations with the currently analyzed inventory object in the same spatiotemporal environment can be identified.

[0098] The criteria for determining operational association include several dimensions: First, spatial proximity, meaning that the spatial distance between other inventory objects and the target inventory object is less than a preset threshold, usually set at 3 to 10 meters, depending on the density of the warehouse racking layout; second, temporal overlap, meaning that the operation time of other inventory objects overlaps with the operation time of the target inventory object or the time difference is less than a preset time threshold, which is usually set at 5 to 30 minutes; and third, operational relevance, meaning that the operation types involved in the two inventory objects have a logical relationship, such as simultaneous receiving operations, simultaneous quality inspection operations, or transmission operations on the same conveyor line.

[0099] For other inventory objects that meet the above association criteria, calculate the operational association strength between them and the target inventory object. The calculation of operational association strength comprehensively considers the inverse of spatial distance, temporal overlap, and operational type similarity. The closer the spatial distance, the higher the association strength; the greater the temporal overlap, the higher the association strength; the more similar the operational types, the higher the association strength. Specifically, a weighted summation method can be used, assuming the spatial distance is... The time overlap is The similarity of operation types is Then the operational correlation strength ,in , , For weight parameters, To prevent the calculation of tiny positive numbers with a denominator of zero, the calculated operational correlation strength is compared with a preset correlation threshold. When comparing, When this happens, the other inventory object is treated as a second-level child node and connected to the corresponding first-level child node through directed edges. The operation association strength value and specific operation type information are marked on the edges.

[0100] After constructing the second-level child nodes, it is necessary to further trace the impact of environmental factors. For each second-level child node, iterate through it to obtain the specific spatial location of the other inventory objects corresponding to that child node at the time of the operation association. For that spatial location, extract the corresponding environmental parameter monitoring data from the environmental monitoring database. These environmental parameters include, but are not limited to, temperature, humidity, light intensity, gas concentration, vibration frequency, and dust concentration. Different types of inventory objects have different sensitivities to environmental factors, requiring the selection of key environmental parameters based on the attributes of the inventory objects.

[0101] For each environmental parameter, its abnormal deviation at the time of operation is calculated. The calculation of abnormal deviation is based on the historical normal value distribution of the environmental parameter. First, the mean value of the environmental parameter under normal conditions at that spatial location is statistically analyzed from historical data. and standard deviation Then calculate the actual monitored value at the time of operation association. The degree of deviation from the normal mean is expressed using standardized deviation. .when When the deviation is significantly greater than the normal fluctuation range, the environmental factor is considered abnormal. The calculated abnormal deviation is then compared with a preset deviation threshold. For comparison, this threshold is typically set at 2 to 3 times the standard deviation level. At that time, it was believed that this environmental factor might be one of the potential causes of the abnormality.

[0102] Identified abnormal environmental factors are treated as third-level child nodes, connected to their corresponding second-level child nodes via directed edges. These edges are labeled with the abnormal deviation value and the specific type and numerical information of the environmental parameters. This completes the tracing process from the abnormal state transition node, through key locations in the flow path, operation-related inventory objects, and environmental factor anomalies, constructing a complete multi-level tracing tree. This tracing tree clearly depicts the multi-dimensional causal chain leading to the anomaly, providing a structured data foundation for subsequent root cause identification and risk diffusion analysis. In practical applications, this tracing tree can be intuitively displayed through a graphical interface, allowing warehouse managers to quickly understand the propagation path and scope of impact of the anomaly, thus enabling them to take targeted measures.

[0103] In one optional embodiment, the contribution strength of each node to the anomalous state transition node is calculated in a multi-level source tree, and the root cause node set is determined to include:

[0104] Extract the connection relationships of child nodes at each level in the multi-level source tree, construct a propagation path graph, extract the propagation path from each child node to the root node in the propagation path graph, obtain the state fluctuation amplitude sequence of the nodes on the propagation path, calculate the temporal similarity of the state fluctuation amplitude sequence, and count the number of propagation paths to which the temporal similarity of each child node exceeds the preset similarity threshold to determine the direct contribution value.

[0105] For each level of child node, extract the lower level child nodes connected to each level of child node, obtain the direct contribution value and state fluctuation amplitude of the lower level child node, calculate the causal response delay time between the state fluctuation amplitude of the lower level child node and the state fluctuation amplitude of the root node, and when the causal response delay time is less than the preset delay threshold, add the direct contribution value of the lower level child node to each level of child node to obtain the cumulative contribution intensity.

[0106] In a multi-level source tree, identify node clusters that form a closed-loop structure, extract the state evolution trajectory of each level of child nodes within the node cluster, calculate the mutual information between the state evolution trajectories as the cycle enhancement factor, determine the final contribution intensity based on the cumulative contribution intensity of each level of child nodes within the node cluster and the cycle enhancement factor, and select level child nodes that exceed the preset contribution threshold as root cause nodes to be added to the root cause node set.

[0107] In one specific implementation, the contribution strength of each node to the abnormal state transition node is quantitatively evaluated in the constructed multi-level traceability tree. The connection relationships between child nodes at each level are extracted from the structure of the multi-level traceability tree. These connections reflect the upstream and downstream dependencies, operational sequences, and environmental impact propagation paths of inventory objects in the warehousing process. By traversing all edges of the traceability tree and recording the directed connections between nodes, a propagation path graph is formed. This propagation path graph contains all possible paths from leaf nodes to the root node (i.e., the abnormal state transition node).

[0108] For each path in the propagation path graph, the state fluctuation amplitude sequence of all nodes on the path is extracted. The state fluctuation amplitude is obtained by comparing the rate of change of the node's state parameters at adjacent time points, such as the temperature change rate recorded by a temperature sensor, the humidity change rate recorded by a humidity sensor, and the movement rate of the inventory object's location coordinates. The state fluctuation amplitudes of the nodes on each propagation path are arranged in chronological order to form a time series. To measure the similarity of the impact of different propagation paths on nodes transitioning from abnormal states, a dynamic time warping algorithm is used to calculate the temporal similarity between the state fluctuation amplitude sequences of each path and the root node's state fluctuation amplitude sequence. The dynamic time warping algorithm can handle the nonlinear scaling of the time series on the time axis. By calculating the optimal matching distance between the two sequences, a normalized temporal similarity value is obtained, ranging from 0 to 1.

[0109] A preset similarity threshold of 0.75 was set, and the number of propagation paths originating from each level's child nodes with a temporal similarity exceeding this threshold was counted. This number reflects how many significantly related propagation paths the child node at that level influences the anomalous state transition node, and is defined as the direct contribution value of that child node. A larger direct contribution value indicates more strongly correlated propagation channels between the node and the anomalous state transition node.

[0110] Considering only the direct contribution value cannot fully reflect the true influence of a node in the source tree, because some nodes may indirectly affect nodes that transition to abnormal states through their lower-level child nodes. Therefore, it is necessary to calculate the cumulative contribution strength. For each level of child node in the source tree, traverse downwards to all its directly connected lower-level child nodes. Obtain the direct contribution value of each of these lower-level child nodes and their state fluctuation amplitude data.

[0111] To determine whether the influence of lower-level child nodes on the current-level child nodes is effectively transmitted to the root node, the concept of causal response delay is introduced. The Granger causality test is used to analyze the time lag relationship between the state fluctuation amplitude sequences of lower-level child nodes and the state fluctuation amplitude sequences of the root node. Specifically, a vector autoregressive model is constructed, representing the current state of the root node as a linear combination of its historical states and the historical states of its lower-level child nodes. An F-test is used to determine whether adding lag terms from lower-level child nodes significantly improves the model's explanatory power. If significant, a causal relationship is considered to exist, and the lag step that maximizes the F-statistic is recorded as the causal response delay.

[0112] A preset delay threshold of 5 time units is set, determined based on the actual operational cycle of the warehousing process. When the causal response delay of a lower-level child node to the root node is less than 5 time units, the influence of that lower-level child node is considered to be quickly transmitted to the root node, and its direct contribution value should be accumulated to the current-level child node. The direct contribution values ​​of all lower-level child nodes that meet the condition are summed, and then added to the direct contribution value of the current-level child node itself to obtain the cumulative contribution intensity of that level of child node. This calculation process is performed layer by layer from bottom to top, ensuring that the influence of lower-level nodes is fully accumulated to upper-level nodes.

[0113] In warehousing processes, certain operations or environmental factors can create cyclical relationships, such as circular handling paths, repetitive temperature control adjustments, and multiple quality inspections and rework. To further explain, while the basic multi-level traceability tree logically maintains an acyclic tree structure, cyclical operations in actual warehousing business processes generate additional relationships outside the tree structure. To comprehensively capture these cyclical effects, while analyzing based on the tree structure, we identify and handle these cyclical relationships that cross tree boundaries, treating them as supplementary connections to the basic tree structure. These cyclical structures manifest as closed-loop paths in the multi-level traceability tree, meaning they start from a node, pass through a series of intermediate nodes, and return to the original node. Specifically, closed-loop paths are formed by overlaying cyclical relationship edges from the business process onto the tree structure, thus constructing a local directed graph structure. Identifying these closed-loop structures is crucial for accurately assessing their contribution intensity, as nodes within the loop reinforce each other's influence.

[0114] A depth-first search algorithm is used to traverse the source tree, while also considering cyclic edges in the business process, maintaining an access stack to record the current search path. When a node is found to already exist in the access stack, it indicates that a closed-loop structure has been detected. All nodes from this node to the top node of the stack are extracted, forming a node cluster. For each identified node cluster, the state evolution trajectories of all child nodes within the cluster are extracted. These trajectories contain state parameter vectors of the nodes at multiple time points, such as multi-dimensional features like temperature, humidity, location, and operation type.

[0115] Calculate the mutual information between the state evolution trajectories of any two nodes within a node cluster. Mutual information measures the degree of interdependence between two random variables; a higher value indicates more synchronized state changes and stronger mutual influence between the two nodes. The joint probability distribution and marginal probability distribution of the state evolution trajectories are estimated using the kernel density estimation method, and the mutual information is calculated using the Shannon entropy formula. ,in and These are the entropies of the state evolution trajectories of the two nodes, respectively. Let be the joint entropy. The average mutual information between all pairs of nodes within a node cluster is used as the cycle enhancement factor for that node cluster.

[0116] For a hierarchical child node belonging to a certain node cluster, its final contribution strength is determined by both the cumulative contribution strength and the cyclic enhancement factor. Specifically, the calculation method involves multiplying the node's cumulative contribution strength by an enhancement factor, where the enhancement factor equals... ,in The cyclic enhancement factor of the node cluster to which this node belongs. To adjust the parameters, they are set to 0.3 based on the actual warehousing scenario. For hierarchical child nodes that do not belong to any closed-loop node cluster, their final contribution strength is their cumulative contribution strength.

[0117] After calculating the final contribution intensity of all hierarchical child nodes, a preset contribution threshold is set for screening. This threshold is determined through statistical distribution analysis of the final contribution intensity of all nodes, selecting the 90th percentile of the distribution as the threshold to ensure that the set of nodes with the most significant impact on anomalous state transition nodes is selected. All hierarchical child nodes with a final contribution intensity exceeding this threshold are marked as root cause nodes and added to the root cause node set. The root cause node set contains the key factors leading to anomalous state transitions, including upstream nodes that directly affect anomalous nodes, as well as deep nodes that exert significant influence through multi-level transmission or cyclical reinforcement. This set provides a precise starting point for subsequent state evolution simulation and risk diffusion analysis, supporting the formulation of targeted control measures.

[0118] like Figure 2 As shown, a flowchart illustrating the state evolution and spatiotemporal diffusion origin tracing process is presented.

[0119] In one optional embodiment, a state evolution simulation engine is constructed based on the root cause node set, and a spatiotemporal diffusion path is generated by reproducing the state evolution process of the root cause node set, including:

[0120] Extract the initial state feature vector and timestamp of each root node in the root node set, obtain the propagation path from each root node to the root node in the multi-level tracing tree, extract the state difference vector between each intermediate node and its adjacent intermediate node on the propagation path, arrange the state difference vectors in time sequence to construct the state transition tensor, and perform tensor decomposition to extract feature components, and construct the state evolution simulation engine.

[0121] The initial state feature vector of each root node is input into the state evolution simulation engine. Based on the feature components, the candidate state feature vector of each intermediate node at each time is deduced. The environmental parameter fluctuation of each intermediate node at the corresponding time is extracted from the candidate state feature vector. The environmental parameter fluctuation is converted into a correction amount and superimposed on the candidate state feature vector to generate the simulated state feature vector of each intermediate node.

[0122] The simulated state feature vectors of each intermediate node are extracted along the propagation path to determine the trigger time and spatial location of the anomaly judgment boundary. The time interval between adjacent trigger times and the spatial span between adjacent spatial locations are calculated. The time interval and spatial span are used as diffusion rate markers to mark each intermediate node. The intermediate nodes are connected in the order of trigger times to generate a spatiotemporal diffusion path.

[0123] In one specific implementation, when extracting the initial state feature vector of each root node in the root node set, it is first necessary to determine the dimensional composition of the feature vector. The initial state feature vector includes the physical state parameters, environmental exposure parameters, and operation history encoding of the inventory object at the corresponding moment of the root node. The physical state parameters cover the values ​​directly collected by sensors such as temperature readings, humidity readings, vibration amplitude, and light intensity of the inventory object. The environmental exposure parameters record spatial environmental information such as ventilation status, shelf level, and surrounding goods density of the storage area where the inventory object is located. The operation history encoding converts the operation information such as the number of handling operations, dwell time, and number of personnel contacted by the inventory object before arriving at the root node into numerical features through one-hot encoding or embedding vectors. These features of different categories are concatenated into a unified high-dimensional feature vector after standardization, and the timestamp corresponding to the feature vector is recorded. The timestamp adopts the Unix timestamp format with accuracy to the second to ensure the accuracy of subsequent time series analysis.

[0124] When obtaining the propagation path from each root cause node to the root node in a multi-level source tree, path tracing is required along the edges of the source tree. Starting from any root cause node, the path is traversed upwards layer by layer according to the directed edge relationships already constructed in the source tree until the root node is reached. All nodes encountered during the traversal are recorded in sequence to form a complete propagation path. Nodes on the propagation path other than the starting and ending points are defined as intermediate nodes. These intermediate nodes represent the carriers of abnormal states in the warehousing process, which may be other goods interacting with the inventory objects, equipment involved in the operation, or factors affecting the environment.

[0125] When extracting the state difference vectors between each intermediate node and its adjacent intermediate nodes along the propagation path, it is necessary to perform element-wise difference operations on the state feature vectors corresponding to the adjacent nodes. Assume the state feature vector of a certain intermediate node along the propagation path is... The state feature vector of its next adjacent intermediate node is The state difference vector is calculated as follows: The state difference vector characterizes the evolutionary increment of the anomalous state during propagation, with each dimension reflecting the magnitude of change in different state parameters. For all adjacent node pairs included in the propagation path, the corresponding state difference vectors are calculated sequentially and arranged in timestamp order.

[0126] When constructing the state transition tensor by temporally arranging the state difference vectors, the state difference vectors along the propagation paths of all root nodes are integrated into a unified three-dimensional tensor structure. The first dimension of the state transition tensor corresponds to different propagation path numbers, the second dimension corresponds to the time step number along the propagation path, and the third dimension corresponds to the feature dimension of the state difference vectors. This tensor organization method allows for the simultaneous capture of state evolution patterns along multiple propagation paths and the correlations between different time steps. The constructed state transition tensor is then decomposed using Tucker decomposition or CP decomposition methods, decomposing the high-dimensional tensor into a core tensor and a combination of several factor matrices. The resulting core tensor contains the main evolution patterns during the state transition process, while each factor matrix corresponds to a basis vector along the propagation path dimension, time dimension, and feature dimension. Feature components are extracted from the core tensor; these feature components are represented as several typical state evolution pattern vectors, which can be used to approximately reconstruct the original state transition tensor through linear combination.

[0127] When constructing the state evolution simulation engine, the extracted feature components are used as the core evolution operators. The engine employs a recurrent neural network architecture. The input layer receives the current node's state feature vector, the hidden layer integrates the feature components obtained from tensor decomposition through a gating mechanism, and the output layer generates the candidate state feature vector for the next time step. The structural information of the feature components is incorporated into the weight matrix of the hidden layer during initialization to ensure that the network's evolutionary pattern remains consistent with historically observed state transition patterns. The network's training data comes from the recorded real state transition sequences in a multi-level source tree, and model parameter optimization is achieved by minimizing the mean squared error between the predicted state and the actual observed state.

[0128] After inputting the initial state feature vectors of each root node into the state evolution simulation engine, the network iteratively extrapolates according to time steps. In the first time step, the initial state feature vectors of the root nodes are used as input, and forward propagation is used to calculate the candidate state feature vector of the first intermediate node at the corresponding time. This candidate state feature vector is then used as input for the next time step, and the process continues to extrapolate the candidate state feature vector of the next intermediate node. This iterative process is repeated until all intermediate nodes along the propagation path are traversed, obtaining a sequence of candidate state feature vectors for each intermediate node at each time step. The candidate state feature vectors represent the theoretical evolution results considering only the intrinsic state transition laws, without incorporating the real-time influence of external environmental factors.

[0129] When extracting the environmental parameter fluctuations of each intermediate node at a given time from the candidate state feature vector, it is necessary to query the spatial coordinates of that intermediate node at a specific time and the monitoring data of its surrounding environment from the warehouse management system. The environmental parameter fluctuation is defined as the deviation between the actual monitored environmental parameter values ​​and the historical average values ​​for that area. Specific environmental parameters include local temperature deviations, humidity deviations, airflow speed changes, and light intensity fluctuations. These deviation values ​​constitute the environmental parameter fluctuation vector, which has the same feature dimension as the candidate state feature vector, facilitating direct superposition of corresponding elements.

[0130] When converting environmental parameter fluctuations into correction values, it is necessary to establish a mapping relationship between environmental parameters and state characteristics. By fitting a regression model to historical data, the influence coefficients of each dimension of the environmental parameter fluctuations on each dimension of the state feature vector are determined. The correction value is calculated as the product of the environmental parameter fluctuation vector and the influence coefficient matrix, resulting in a correction vector with dimensions consistent with the candidate state feature vector. The correction value is then superimposed onto the candidate state feature vector, and element-wise addition yields a simulated state feature vector that incorporates the real-time environmental impact. This simulated state feature vector more accurately reflects the actual state evolution of intermediate nodes in a real warehousing environment.

[0131] When extracting the trigger time and spatial location of the simulated state feature vectors of each intermediate node along the propagation path to break through the anomaly judgment boundary, it is necessary to predefine the judgment rules for the anomaly judgment boundary. The anomaly judgment boundary is determined by statistical analysis of historical normal state data, and upper and lower limit thresholds for each state parameter are set using the box plot method or the three-standard-deviation principle based on the Gaussian distribution. Each component of the simulated state feature vector of each intermediate node is checked one by one to see if it exceeds the corresponding threshold range. Once any component exceeds the threshold, the intermediate node is judged to have triggered an anomaly at the current moment. The timestamp of the first anomaly trigger is recorded as the trigger time, and the spatial coordinates of the intermediate node at that moment are also recorded. The spatial coordinates include the warehouse area number, shelf row and column number, and the precise location in the three-dimensional Cartesian coordinate system.

[0132] When calculating the time interval between adjacent trigger moments, the trigger moments arranged sequentially along the propagation path are subjected to differential operations. Let the trigger moment of the Kth intermediate node be... , No. The triggering time of each intermediate node is The time interval is calculated as The unit is seconds or minutes. The size of the time interval reflects the speed at which an anomaly spreads during its propagation; a shorter time interval means that the anomaly is transmitted quickly between adjacent nodes.

[0133] When calculating the spatial span between adjacent spatial locations, the spatial coordinates of adjacent triggering nodes are extracted, and the straight-line distance between the two points is calculated using Euclidean distance. Let the spatial coordinates of the Kth triggering node be... , No. The spatial coordinates of the trigger nodes are: The spatial span is calculated as follows: Spatial span characterizes the distance an anomaly travels in the physical space of a warehouse, and when combined with time intervals, the spatial diffusion rate of the anomaly can be calculated.

[0134] When time intervals and spatial spans are used as diffusion rate identifiers to label each intermediate node, attribute fields are added to each intermediate node to record its corresponding time interval and spatial span values. The diffusion rate can be further calculated as the ratio of spatial span to time interval, in meters per second or meters per minute. This indicator quantifies the dynamic characteristics of anomaly propagation in the spatiotemporal domain. When generating a spatiotemporal diffusion path by connecting intermediate nodes in the order of triggering time, all intermediate nodes that triggered the anomaly are sorted from earliest to latest according to their timestamps, and adjacent nodes are connected sequentially with directed edges. The spatiotemporal diffusion path is expressed in the form of a graph structure, where nodes represent intermediate nodes that triggered the anomaly, and edges represent the spatiotemporal connections of the anomaly propagation. Each directed edge is accompanied by three attribute identifiers: time interval, spatial span, and diffusion rate. These identifiers provide a quantitative basis for subsequent anomaly propagation pattern analysis. The visualization of the spatiotemporal diffusion path uses a three-dimensional coordinate system. The horizontal and vertical axes represent the planar coordinates of the storage space, and the vertical axis represents the time dimension. The path in three-dimensional space is presented as a curved trajectory extending from the root cause node to the root node.

[0135] In generating spatiotemporal diffusion paths, special handling is required for propagation cases with branches. When an intermediate node simultaneously triggers anomalies in multiple downstream nodes, the spatiotemporal diffusion path will fork, forming a tree-like diffusion structure. In this case, a diffusion rate identifier needs to be calculated for each branch, and each branch path needs to be distinguished by different colors or line types in the path diagram. When multiple propagation paths converge at the same node, this node needs to be marked as the convergence point, and the time difference between the arrival times of each propagation path at the convergence point needs to be recorded. This time difference reflects the difference in diffusion speed between different propagation paths.

[0136] To verify the accuracy of the spatiotemporal diffusion path, the simulated path is compared with the actual distribution of anomalous nodes recorded in a multi-level source tree. The recall and precision are calculated between the set of nodes on the simulated path and the set of real anomalous nodes. Recall represents the proportion of real anomalous nodes covered by the simulated path, and precision represents the proportion of nodes on the simulated path that actually exhibit anomalies. When both recall and precision exceed a preset threshold (e.g., 80%), the simulation of the spatiotemporal diffusion path is considered to have reached a usable standard. If the verification results do not meet the standard, it is necessary to return to adjust the mapping coefficients of the network parameters or environmental parameter fluctuations in the state evolution simulation engine and re-perform the simulation until the accuracy requirements are met.

[0137] Once the spatiotemporal diffusion path is generated, key features of anomaly propagation can be extracted for subsequent early warning. These key features include statistical indicators such as average diffusion rate, maximum spatial span, total path duration, and number of branch nodes. The average diffusion rate, calculated by dividing the total spatial distance of the path by the total path duration, reflects the overall speed of anomaly propagation. The maximum spatial span identifies the farthest distance a single jump can cover during anomaly propagation; this indicator is closely related to the spatial connectivity of the warehouse layout. The total path duration is the time difference between the triggering time of the root cause node and the triggering time of the root node, reflecting the delay from the initial occurrence of the anomaly to its final detection. The number of branch nodes counts the number of nodes where the path branches, characterizing the breadth of anomaly propagation.

[0138] Different types of inventory exhibit varying spatiotemporal diffusion paths. For temperature-sensitive inventory, the diffusion path typically extends along the temperature gradient within the storage space, with the diffusion rate highly correlated with the operational status of the area's ventilation system. For humidity-sensitive inventory, the diffusion path tends to spread between adjacent shelves within the same humidity zone, with a relatively small spatial span but short time intervals. For vibration-sensitive inventory, the diffusion path highly overlaps with the operating trajectories of material handling equipment, and the diffusion rate significantly accelerates during periods of high-frequency equipment operation. By clustering the spatiotemporal diffusion paths of different types of inventory, a knowledge base mapping inventory types to abnormal propagation patterns can be established, providing a basis for developing targeted risk control measures.

[0139] The spatiotemporal diffusion path can also be correlated with other data sources in the warehouse management system to uncover potential factors influencing anomaly propagation. By aligning the trigger times of each node on the spatiotemporal diffusion path with warehouse operation logs, synchronous operational events occurring during anomaly propagation can be identified, such as goods handling, equipment maintenance, and temperature control system adjustments. Statistical analysis shows that certain operational events significantly alter the anomaly propagation rate; for example, activating the ventilation system accelerates the spatial diffusion of temperature anomalies, while emergency activation of refrigeration equipment can temporarily block the propagation path. These operational events are labeled as control nodes on the diffusion path, providing a reference for optimizing emergency response strategies.

[0140] In one optional embodiment, the state evolution trajectories of other inventory objects are input into the state evolution simulation engine to determine whether they intersect with the spatiotemporal diffusion path, thus obtaining risky inventory objects including:

[0141] Obtain the inventory objects other than the inventory object corresponding to the root node, extract the state evolution trajectory of the other inventory objects, extract the initial state feature vector, timestamp and spatial position of each state transition point from the state evolution trajectory, input the initial state feature vector into the state evolution simulation engine, infer the candidate state feature vector and corresponding spatial position of the other inventory objects at each time based on the feature components, and construct the predicted evolution trajectory of the other inventory objects.

[0142] Extract the trigger time, spatial location, and diffusion rate identifier of each intermediate node in the spatiotemporal diffusion path. Calculate the spatial distance between the spatial location of the predicted evolution trajectory of other inventory objects at each trigger time and the corresponding spatial location in the spatiotemporal diffusion path. Combine the diffusion rate identifier to calculate the diffusion influence range. When the spatial distance is within the diffusion influence range, mark it as the intersection time.

[0143] For other inventory objects with intersection times, extract the candidate state feature vectors of the predicted evolution trajectories of the other inventory objects at the intersection times, calculate the feature similarity with the simulated state feature vectors of the intermediate nodes at the corresponding intersection times in the spatiotemporal diffusion path, and when the feature similarity exceeds the preset similarity threshold, mark the other inventory object as a risk inventory object and add it to the risk inventory object set.

[0144] In one specific implementation, after the smart warehousing system reproduces the state evolution process of the root cause node set using a state evolution simulation engine, it needs to further identify other inventory objects that may be affected. Complete flow information for other inventory objects besides those corresponding to the root nodes is obtained; these objects may intersect with the anomaly source in time or space. For each other inventory object, its state evolution trajectory is extracted from the warehouse management system. This trajectory records all state changes the inventory object undergoes from the time it enters the warehouse. Detailed information for each state transition point is extracted from the state evolution trajectory. A state transition point refers to the moment when the inventory object's state changes, such as from a pending inspection state to a qualified state, or from a stored state to an outbound state.

[0145] For each state transition point, an initial state feature vector of the inventory object at that moment is extracted. This vector contains multi-dimensional feature components such as temperature, humidity, packaging integrity, and quality parameters. Simultaneously, the timestamp and spatial location corresponding to this state transition point are recorded. The timestamp uses a unified time base, and the spatial location is marked using a three-dimensional coordinate system, including shelf number, layer number, and specific location coordinates. The extracted initial state feature vector is input into the state evolution simulation engine. Based on the established state transition rules, the engine infers the state of the inventory object at subsequent moments. During the inference process, candidate state feature vectors are calculated for each moment based on the interaction relationships between feature components and the influence weights of environmental factors.

[0146] The derivation of candidate state feature vectors adopts a time-series recursive approach. Based on the initial state feature vector, and considering warehousing environment parameters, operational influences, and time decay factors, the feature vector values ​​for subsequent time points are calculated progressively. For the temperature feature component, ambient temperature fluctuations, the heat dissipation characteristics of the goods themselves, and the heat conduction effects of adjacent stored goods are considered. The humidity feature component comprehensively considers warehouse ventilation conditions, the permeability of goods packaging materials, and seasonal humidity variations. During the derivation process, the corresponding spatial location of the stored goods at each time point is calculated simultaneously. If a stored goods are moved or moved in / out of the warehouse, the spatial location is updated based on the movement records in the operation log. Finally, the predicted evolution trajectory of other stored goods is formed, which records the evolution of the candidate state feature vectors and their spatial locations in time-series form.

[0147] To determine whether other inventory items are affected by the anomaly, it is necessary to compare the intersection between their predicted evolution trajectories and spatiotemporal diffusion paths. Key parameters for each intermediate node are extracted from the spatiotemporal diffusion path, including trigger time, spatial location, and diffusion rate identifier. The trigger time indicates the point in time when the anomaly affects that node, and the spatial location identifies the node's coordinates within the storage space. The diffusion rate identifier reflects the speed at which the anomaly propagates to its surroundings; different types of anomalies have different diffusion rates. For example, the diffusion rate of temperature anomalies is related to airflow speed, while the diffusion rate of pollutants is related to their volatility characteristics.

[0148] For each trigger moment in the spatiotemporal diffusion path, the spatial position of the predicted evolution trajectory of other inventory objects at that moment is calculated. By querying the time series records of the predicted evolution trajectory, the time point closest to the trigger moment is located, and the spatial coordinates of that time point are obtained. The spatial distance between this spatial position and the corresponding spatial position in the spatiotemporal diffusion path at the trigger moment is calculated. The spatial distance is measured using a three-dimensional Euclidean distance method, taking into account both horizontal distance and vertical height differences. In actual calculations, if two inventory objects are located on different layers of the same shelf, although the horizontal distance may be close, the vertical distance may be large, requiring an accurate reflection of the true distance in three-dimensional space.

[0149] The diffusion impact range is calculated based on the diffusion rate identifier. The diffusion impact range is defined as the spatial boundary where an abnormal state can have a significant impact. During calculation, the diffusion rate is multiplied by the time interval from the trigger moment to the current judgment moment to obtain the diffusion radius. The diffusion radius is not uniformly distributed but is adjusted according to the physical structure of the storage space. For example, obstacles such as shelves and walls can hinder the diffusion path, while ventilation ducts may accelerate diffusion in certain directions. By comprehensively considering these factors, an irregular diffusion impact range boundary is constructed. It is then determined whether the spatial location of other stored objects is within the diffusion impact range. When the spatial distance is less than or equal to the diffusion radius and there is no physical isolation, the trigger moment is marked as the intersection moment.

[0150] For other inventory objects at the intersection time, further analysis is needed to determine whether they are truly affected by the anomalous state. Spatiotemporal intersection alone is insufficient to determine the substantial impact; comparison of state feature similarities is also required. Candidate state feature vectors of the predicted evolution trajectories of other inventory objects at the intersection time are extracted. These vectors reflect the expected state of the inventory objects in the absence of anomalous influence. Simultaneously, simulated state feature vectors of intermediate nodes at the corresponding intersection time are obtained from the spatiotemporal diffusion path. These vectors represent the characteristic manifestation of the anomalous state at that spatiotemporal location.

[0151] The similarity between two feature vectors is calculated using a weighted cosine similarity method, assigning different weights to each feature component. Environmental features such as temperature and humidity have higher weights because they are easily affected by the surrounding environment. Intrinsic attributes such as quality parameters and appearance features have lower weights because changes in these features typically require longer periods or stronger external influences. During the calculation, the feature vectors are normalized to eliminate the influence of differences in the dimensions of different feature components. The normalized feature vectors are then used to obtain a similarity value through a dot product operation, with the value ranging from 0 to 1; the closer the value is to 1, the higher the similarity.

[0152] The calculated feature similarity is compared with a preset similarity threshold, which is determined based on statistical analysis of historical data and is typically set between 0.75 and 0.85. When the feature similarity exceeds this threshold, it indicates that the actual state characteristics of other inventory objects are highly consistent with the state characteristics on the abnormal diffusion path, and it is determined that the inventory object has been affected by the abnormal state. This other inventory object is marked as a risk inventory object, and its inventory number, intersection time, spatial location, and feature similarity value are recorded and added to the risk inventory object set. The risk inventory object set adopts a dynamic update mechanism. As the spatiotemporal diffusion path continues to evolve and new inventory objects are added, the risk object list is judged and updated in real time, providing an accurate target object range for subsequent control measures.

[0153] A second aspect of this invention provides an AI-powered inventory traceability system for smart warehousing, comprising:

[0154] The data acquisition unit is used to collect multi-source sensor data and operation log data of inventory objects in the warehousing process;

[0155] An anomaly identification unit is used to divide multi-source sensor data and operation log data into multiple spatiotemporal slices according to the time and space dimensions, extract the state evolution trajectory of objects in each spatiotemporal slice, construct a state transition matrix based on the state evolution trajectory, and identify abnormal state transition nodes by analyzing the abnormal fluctuations of the transition probability in the state transition matrix.

[0156] The root cause tracing unit is used to construct a multi-level tracing tree with the abnormal state transfer node as the root node, based on the inventory object flow path, the operation interaction relationship between multiple inventory objects and the influence of environmental factors. In the multi-level tracing tree, the contribution intensity of each node to the abnormal state transfer node is calculated, and the root cause node set is determined.

[0157] The risk prediction unit is used to construct a state evolution simulation engine based on the root cause node set. It generates a spatiotemporal diffusion path by reproducing the state evolution process of the root cause node set. The state evolution trajectory of other inventory objects is input into the state evolution simulation engine to determine whether it intersects with the spatiotemporal diffusion path, thereby obtaining the risky inventory objects.

[0158] The control execution unit is used to generate and execute control instructions based on the root cause node set and the risk inventory object.

[0159] A third aspect of the present invention provides an electronic device, comprising:

[0160] processor;

[0161] Memory used to store processor-executable instructions;

[0162] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0163] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0164] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based inventory traceability method for smart warehousing, characterized in that, include: Collect multi-source sensor data and operation log data of inventory objects in the warehousing process; Multi-source sensor data and operation log data are divided into multiple spatiotemporal slices according to the time and space dimensions. The state evolution trajectory of objects in each spatiotemporal slice is extracted. A state transition matrix is ​​constructed based on the state evolution trajectory. Abnormal state transition nodes are identified by analyzing the abnormal fluctuations of the transition probability in the state transition matrix. Using the abnormal state transition node as the root node, a multi-level traceability tree is constructed based on the inventory object flow path, the operation interaction relationship between multiple inventory objects, and the influence of environmental factors. The contribution intensity of each node to the abnormal state transition node is calculated in the multi-level traceability tree, and the root cause node set is determined. A state evolution simulation engine is constructed based on the root cause node set. A spatiotemporal diffusion path is generated by reproducing the state evolution process of the root cause node set. The state evolution trajectories of other inventory objects are input into the state evolution simulation engine to determine whether they intersect with the spatiotemporal diffusion path, thereby obtaining the risk inventory objects. Control instructions are generated and executed based on the root cause node set and risk inventory objects.

2. The method according to claim 1, characterized in that, The multi-source sensor data and operation log data are divided into multiple spatiotemporal slices based on time and space dimensions. The state evolution trajectory of objects in memory within each spatiotemporal slice is extracted, including: Based on the timestamp information and spatial location information of inventory objects in multi-source sensor data and operation log data, multiple spatiotemporal slice units are divided in the spatiotemporal coordinate system according to preset time intervals and preset spatial ranges, and each data item is mapped to the corresponding spatiotemporal slice unit. Within each spatiotemporal slice unit, the state feature sequence of the inventory object is extracted, the state entropy value of the state feature sequence within the spatiotemporal slice unit is calculated, and the moment when the state entropy value changes abruptly is marked as the state transition point. By connecting the state transition points of the same inventory object in adjacent spatiotemporal slice units in chronological order, the state evolution trajectory of the inventory object is constructed. The state evolution trajectory records the state entropy value, timestamp, and spatial location of each state transition point.

3. The method according to claim 1, characterized in that, A state transition matrix is ​​constructed based on the state evolution trajectory. Abnormal state transition nodes are identified by analyzing the abnormal fluctuations in the transition probabilities within the state transition matrix. Extract the state feature vectors of each state transition point in the state evolution trajectory, calculate the distance metric between the state feature vectors, divide the state feature vectors with a distance metric less than a preset distance threshold into the same node cluster, assign the same state identifier to the state feature vectors belonging to the same node cluster, and obtain the state identifier sequence. The transition frequency between each state identifier in the statistical state identifier sequence is counted, and a state transition matrix is ​​constructed. For each transition probability in the state transition matrix, a time-series tracking sequence of the transition probability is established. By periodically decomposing the time-series tracking sequence, the baseline periodic component and abnormal fluctuation component of the transition probability are extracted. Calculate the energy spectral density of the abnormal fluctuation component, mark the transition probability when the energy spectral density exceeds the preset energy threshold as the abnormal transition probability, and extract the starting state identifier and target state identifier corresponding to the abnormal transition probability. Locate the state transition point from the initial state identifier to the target state identifier in the state evolution trajectory, extract the operation log data within a preset time range before and after the state transition point, and mark the state transition point as an abnormal state transition node.

4. The method according to claim 1, characterized in that, Using the abnormal state transition node as the root node, a multi-level traceability tree is constructed based on the inventory object flow path, the operational interaction relationships between multiple inventory objects, and the influence of environmental factors, including: Set the abnormal state transfer node as the root node of the multi-level traceability tree, extract the timestamp of the abnormal state transfer node, trace back a preset time window with the timestamp as the endpoint, obtain the flow path of the inventory object, calculate the state fluctuation amplitude of each path segment in the flow path, filter the path segments whose state fluctuation amplitude exceeds the preset fluctuation threshold, extract the corresponding starting spatial position and ending spatial position as the first-level child node and connect them to the root node. Traverse each first-level child node, obtain the operation log data of the corresponding spatial location within the corresponding time period, extract the identifiers of other inventory objects that have operation associations with the inventory object from the operation log data, calculate the operation association strength, and connect other inventory objects whose operation association strength exceeds the preset association threshold as second-level child nodes to the corresponding first-level child nodes. Traverse the second-level child nodes to obtain the environmental parameter monitoring data of the spatial location of the corresponding other inventory objects at the time of operation association, calculate the abnormal deviation degree, and connect the environmental factors with abnormal deviation degree exceeding the preset deviation threshold as the third-level child nodes to the corresponding second-level child nodes to complete the construction of the multi-level traceability tree.

5. The method according to claim 1, characterized in that, Calculate the contribution strength of each node to the abnormal state transition node in the multi-level source tree, and determine the root cause node set including: Extract the connection relationships of child nodes at each level in the multi-level source tree, construct a propagation path graph, extract the propagation path from each child node to the root node in the propagation path graph, obtain the state fluctuation amplitude sequence of the nodes on the propagation path, calculate the temporal similarity of the state fluctuation amplitude sequence, and count the number of propagation paths to which the temporal similarity of each child node exceeds the preset similarity threshold to determine the direct contribution value. For each level of child node, extract the lower level child nodes connected to each level of child node, obtain the direct contribution value and state fluctuation amplitude of the lower level child node, calculate the causal response delay time between the state fluctuation amplitude of the lower level child node and the state fluctuation amplitude of the root node, and when the causal response delay time is less than the preset delay threshold, add the direct contribution value of the lower level child node to each level of child node to obtain the cumulative contribution intensity. In a multi-level source tree, identify node clusters that form a closed-loop structure, extract the state evolution trajectory of each level of child nodes within the node cluster, calculate the mutual information between the state evolution trajectories as the cycle enhancement factor, determine the final contribution intensity based on the cumulative contribution intensity of each level of child nodes within the node cluster and the cycle enhancement factor, and select level child nodes that exceed the preset contribution threshold as root cause nodes to be added to the root cause node set.

6. The method according to claim 1, characterized in that, A state evolution simulation engine is built based on a root node set. By reproducing the state evolution process of the root node set, spatiotemporal diffusion paths are generated, including: Extract the initial state feature vector and timestamp of each root node in the root node set, obtain the propagation path from each root node to the root node in the multi-level tracing tree, extract the state difference vector between each intermediate node and its adjacent intermediate node on the propagation path, arrange the state difference vectors in time sequence to construct the state transition tensor, and perform tensor decomposition to extract feature components, and construct the state evolution simulation engine. The initial state feature vector of each root node is input into the state evolution simulation engine. Based on the feature components, the candidate state feature vector of each intermediate node at each time is deduced. The environmental parameter fluctuation of each intermediate node at the corresponding time is extracted from the candidate state feature vector. The environmental parameter fluctuation is converted into a correction amount and superimposed on the candidate state feature vector to generate the simulated state feature vector of each intermediate node. The simulated state feature vectors of each intermediate node are extracted along the propagation path to determine the trigger time and spatial location of the anomaly judgment boundary. The time interval between adjacent trigger times and the spatial span between adjacent spatial locations are calculated. The time interval and spatial span are used as diffusion rate markers to mark each intermediate node. The intermediate nodes are connected in the order of trigger times to generate a spatiotemporal diffusion path.

7. The method according to claim 1, characterized in that, The state evolution trajectories of other inventory objects are input into the state evolution simulation engine to determine whether they intersect with the spatiotemporal diffusion path, thus identifying the risky inventory objects, including: Obtain the inventory objects other than the inventory object corresponding to the root node, extract the state evolution trajectory of the other inventory objects, extract the initial state feature vector, timestamp and spatial position of each state transition point from the state evolution trajectory, input the initial state feature vector into the state evolution simulation engine, infer the candidate state feature vector and corresponding spatial position of the other inventory objects at each time based on the feature components, and construct the predicted evolution trajectory of the other inventory objects. Extract the trigger time, spatial location, and diffusion rate identifier of each intermediate node in the spatiotemporal diffusion path. Calculate the spatial distance between the spatial location of the predicted evolution trajectory of other inventory objects at each trigger time and the corresponding spatial location in the spatiotemporal diffusion path. Combine the diffusion rate identifier to calculate the diffusion influence range. When the spatial distance is within the diffusion influence range, mark it as the intersection time. For other inventory objects with intersection times, extract the candidate state feature vectors of the predicted evolution trajectories of the other inventory objects at the intersection times, calculate the feature similarity with the simulated state feature vectors of the intermediate nodes at the corresponding intersection times in the spatiotemporal diffusion path, and when the feature similarity exceeds the preset similarity threshold, mark the other inventory object as a risk inventory object and add it to the risk inventory object set.

8. An AI-powered inventory traceability system for smart warehousing, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is used to collect multi-source sensor data and operation log data of inventory objects in the warehousing process; An anomaly identification unit is used to divide multi-source sensor data and operation log data into multiple spatiotemporal slices according to the time and space dimensions, extract the state evolution trajectory of objects in each spatiotemporal slice, construct a state transition matrix based on the state evolution trajectory, and identify abnormal state transition nodes by analyzing the abnormal fluctuations of the transition probability in the state transition matrix. The root cause tracing unit is used to construct a multi-level tracing tree with the abnormal state transfer node as the root node, based on the inventory object flow path, the operation interaction relationship between multiple inventory objects and the influence of environmental factors. In the multi-level tracing tree, the contribution intensity of each node to the abnormal state transfer node is calculated, and the root cause node set is determined. The risk prediction unit is used to construct a state evolution simulation engine based on the root cause node set. It generates a spatiotemporal diffusion path by reproducing the state evolution process of the root cause node set. The state evolution trajectory of other inventory objects is input into the state evolution simulation engine to determine whether it intersects with the spatiotemporal diffusion path, thereby obtaining the risky inventory objects. The control execution unit is used to generate and execute control instructions based on the root cause node set and the risk inventory object.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images