Fusion time sequence large model and multi-source heterogeneous data warehousing time efficient precise prediction method and system

Abstract

The application discloses a warehouse time limit accurate prediction method and system fusing a time sequence large model and multi-source heterogeneous data, the method firstly acquires environment state time sequences, goods turnover event streams and external influence event records, and maps to a unified warehouse time axis to generate mixed data streams; environment period base components, goods trend base components and external pulse excitation components are obtained through a time sequence decomposition network; environment period phase feature trajectories, goods turnover trend slope feature trajectories and external event pulse decay feature trajectories are respectively projected; a time sequence large model is inputted to perform cross-channel time dependence relationship joint modeling, and a warehouse time limit prediction result containing future time step retention distribution parameter sets is generated.

Description

Technical Field

[0001] This invention relates to the field of intelligent warehouse management technology, and more specifically, to a method and system for accurate prediction of warehouse timeliness by integrating a large time-series model and multi-source heterogeneous data. Background Technology

[0002] Warehouse timeliness forecasting is a crucial aspect of warehouse management and logistics scheduling. Existing methods often rely on historical inbound and outbound records, order volumes, or manual experience rules to predict cargo dwell time, which fails to fully reflect the combined impact of changes in the warehousing environment, cargo turnover rhythm, and external weather, traffic, and logistics events on warehouse timeliness.

[0003] In real-world warehousing scenarios, multi-source data often suffers from issues such as different sources, varying sampling frequencies, and inconsistent time granularities. This makes it difficult to collaboratively model environmental status data, cargo event data, and external influencing events on a unified timeline. Furthermore, the overlapping effects of environmental cyclical fluctuations, cargo turnover trends, and external unforeseen events make it challenging for traditional prediction models to distinguish the paths of action of different factors. Prediction results are typically only single-time values, failing to express the distribution range of future cargo dwell time and the risk of abnormal dwell times. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for accurate prediction of warehouse delivery timeliness by integrating large time-series models and multi-source heterogeneous data.

[0005] In a first aspect, embodiments of the present invention provide a method for accurately predicting warehouse delivery timeliness by integrating a large time-series model with multi-source heterogeneous data, comprising: Acquire the environmental status time series reported by warehouse sensor acquisition devices, the cargo turnover event stream reported by cargo entry and exit scanning devices, and the external impact event records pushed by external event publishing platforms; According to the preset timestamp mapping relationship, the environmental collection timestamp in the environmental state time series, the cargo event timestamp in the cargo turnover event stream, and the external event timestamp in the external influence event record are mapped to time anchors on the warehousing time axis, and a hybrid data stream with a unified time anchor mark is generated based on the time anchors. The pre-built temporal decomposition network is invoked to decouple the hybrid data stream in time series, generating environmental periodic basis components, cargo trend basis components, and external impulse excitation components; The environmental periodic base component, the cargo trend base component, and the external pulse excitation component are projected onto the periodic phase space, the trend slope space, and the pulse intensity decay space, respectively, to generate the environmental periodic phase feature trajectory, the cargo turnover trend slope feature trajectory, and the external event pulse decay feature trajectory. The environmental cycle phase characteristic trajectory, the cargo turnover trend slope characteristic trajectory, and the external event pulse decay characteristic trajectory are input into a pre-constructed time series large model to perform cross-channel time dependency joint modeling, generating warehousing timeliness prediction results. The warehousing timeliness prediction results include the set of retention distribution parameters corresponding to each time step within the future time window.

[0006] Secondly, embodiments of the present invention provide a warehouse timeliness accurate prediction system that integrates a large time series model and multi-source heterogeneous data, including at least one service node; The service node includes a storage unit and a computing unit; the storage unit is used to store program code; the computing unit is used to run the program code to execute the warehouse timeliness accurate prediction method that integrates time-series large models and multi-source heterogeneous data as described in the first aspect.

[0007] Compared with existing technologies, the beneficial effects provided by this invention include: by fusing multi-source heterogeneous data such as environmental conditions, cargo turnover and external events through a unified time anchor point, and by decoupling and modeling the effects of cycles, trends and impulses, the accuracy of warehousing timeliness prediction can be improved; at the same time, by using a large time series model to perform cross-channel time dependency joint analysis, the invention outputs a set of retention distribution parameters, enabling the prediction results to reflect the distribution changes of future retention risks, thereby improving prediction stability and the reference value for scheduling decisions. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A flowchart illustrating the steps of the method for accurate prediction of warehouse timeliness by fusing a large time-series model and multi-source heterogeneous data provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the inter-channel causal hysteresis constrained attention structure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the abnormal mutation identification and attribution tracing structure provided in an embodiment of the present invention; Figure 4 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0010] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0011] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the method for accurate prediction of warehouse delivery time by fusing a large time-series model and multi-source heterogeneous data, as provided in this embodiment. The following is a detailed description of this method for accurate prediction of warehouse delivery time by fusing a large time-series model and multi-source heterogeneous data.

[0012] Step S201: Obtain the environmental status time series reported by the warehouse sensor acquisition device, the cargo turnover event stream reported by the cargo entry and exit scanning device, and the external impact event records pushed by the external event publishing platform; Step S202: According to the preset timestamp mapping relationship, the environmental collection timestamp in the environmental state time series, the cargo event timestamp in the cargo turnover event stream, and the external event timestamp in the external influence event record are mapped to time anchors on the storage time axis, and a hybrid data stream with unified time anchor mark is generated based on the time anchors. Step S203: Invoke the pre-built time series decomposition network to decouple the hybrid data stream in time series, and generate environmental periodic basis components, cargo trend basis components and external impulse excitation components. Step S204: Project the environmental periodic base component, the cargo trend base component, and the external pulse excitation component onto the periodic phase space, the trend slope space, and the pulse intensity decay space, respectively, to generate the environmental periodic phase feature trajectory, the cargo turnover trend slope feature trajectory, and the external event pulse decay feature trajectory. Step S205: Input the environmental cycle phase characteristic trajectory, the cargo turnover trend slope characteristic trajectory, and the external event pulse decay characteristic trajectory into the pre-constructed time series large model to perform cross-channel time dependency joint modeling and generate warehousing timeliness prediction results. The warehousing timeliness prediction results include the set of retention distribution parameters corresponding to each time step within the future time window.

[0013] In this embodiment of the invention, the server acts as the execution entity, establishing data communication connections with warehouse sensing and acquisition devices, goods entry and exit scanning devices, and an external event publishing platform. After acquiring the environmental state time series, goods turnover event stream, and external impact event records, the server maps the three types of data to the warehouse timeline according to a preset timestamp mapping relationship, and forms a hybrid recording unit at the time anchor point of the warehouse timeline, resulting in a hybrid data stream with a unified time anchor mark.

[0014] The server invokes a pre-built temporal decomposition network to decouple the hybrid data stream in time series, generating environmental periodic basis components, cargo trend basis components, and external impulse excitation components. The server further projects the environmental periodic basis components onto a periodic phase space to generate environmental periodic phase feature trajectories; projects the cargo trend basis components onto a trend slope space to generate cargo turnover trend slope feature trajectories; and projects the external impulse excitation components onto a pulse intensity decay space to generate external event pulse decay feature trajectories.

[0015] The server inputs the environmental cycle phase characteristic trajectory, the cargo turnover trend slope characteristic trajectory, and the external event pulse decay characteristic trajectory into a pre-constructed time-series large model to perform cross-channel time dependency joint modeling, generating a set of retention distribution parameters corresponding to each time step within the future time window, and obtaining the warehousing timeliness prediction result based on the retention distribution parameter set. The warehousing timeliness prediction result is used to characterize the probability distribution of cargo retention time within the future time window.

[0016] In this embodiment of the invention, the retention distribution parameter set is used to describe the distribution state of cargo retention time at each time step within a future time window in the form of a probability distribution. The server pre-divides the cargo retention time into multiple retention time intervals, each interval corresponding to an interval center value. The time-series large model's timeliness distribution prediction layer outputs the probability value corresponding to each retention time interval for each future time step, and the probability value and the retention time interval center value together form the retention distribution parameter set. The retention distribution parameter set includes at least one of the following: interval probability parameter, distribution expectation parameter, distribution variance parameter, distribution skewness parameter, distribution kurtosis parameter, and long retention probability parameter.

[0017] Specifically, for the t-th time step within a future time window, the server represents the set of retention distribution parameters as follows: , Including probability values ​​corresponding to multiple stay duration intervals and the center value of each interval corresponding to the length of stay. Where b represents the dwell time interval number, and B represents the total number of dwell time intervals. The server determines the dwell time interval based on the probability value. and the center value of the interval Calculate the expected distribution parameter, variance parameter, skewness parameter, and kurtosis parameter corresponding to the t-th time step. Through this method, the set of retention distribution parameters no longer represents only a single predicted value, but can represent the central tendency, dispersion, long-tail tendency, and degree of abnormal fluctuation of cargo retention time within future time steps, thus providing a calculable data foundation for subsequent analysis and processing of distribution pattern disturbances. Based on this, , , , , , , ,in, This represents the set of retention distribution parameters at time step t. Let represent the probability that the t-th time step falls into the b-th dwell time interval. Let B represent the center value of the b-th interval of stay duration, and let B represent the total number of intervals of stay duration. Represents the expected parameter of the distribution. Represents the distribution variance parameter. Indicates the distribution skewness parameter. Represents the kurtosis parameter of the distribution. This represents the probability parameter of long stay. This indicates a preset long dwell time threshold. In this embodiment of the invention, the pre-constructed time-series model is trained using historical warehousing time-efficiency samples. The server acquires historical environmental state time series, historical cargo turnover event streams, historical external influence event records, and historical cargo actual dwell time records within the historical warehousing business cycle. The server maps the historical environmental state time series, the historical cargo turnover event stream, and the historical external influence event records to the historical warehousing time axis according to the same timestamp mapping relationship as the prediction stage, generating a historical mixed data stream. The server then decouples the historical mixed data stream into time series using the pre-constructed time-series decomposition network or the time-series decomposition network corresponding to the training stage, obtaining historical environmental cycle basis components, historical cargo trend basis components, and historical external impulse excitation components.

[0018] The server performs feature trajectory generation processing on the historical environmental cycle base component, the historical cargo trend base component, and the historical external impulse excitation component to obtain the historical environmental cycle phase feature trajectory, the historical cargo turnover trend slope feature trajectory, and the historical external event impulse decay feature trajectory. The server then uses the historical environmental cycle phase feature trajectory, the historical cargo turnover trend slope feature trajectory, and the historical external event impulse decay feature trajectory as training input samples.

[0019] The server constructs training labels based on historical records of actual cargo dwell time. Specifically, the server divides the actual cargo dwell time within each historical time window into multiple dwell time intervals, calculates the proportion of samples whose actual cargo dwell time falls into each dwell time interval in each future time step, generates true interval probability labels, and calculates the true distribution expectation parameter, true distribution variance parameter, true distribution skewness parameter, and true distribution kurtosis parameter based on the true interval probability labels to form true dwell time distribution parameter labels.

[0020] The server inputs the training input samples into the time-series model to be trained, and the time-series model outputs a set of predicted retention distribution parameters. The server calculates the training loss based on the difference between the predicted retention distribution parameter set and the actual retention distribution parameter labels. The training loss includes interval probability distribution prediction loss, moment parameter prediction loss, temporal continuity constraint loss, and anomalous time window weighted loss. The interval probability distribution prediction loss is used to constrain the consistency between the predicted probability distribution and the actual probability distribution; the moment parameter prediction loss is used to constrain the expected value, variance, skewness, and kurtosis of the predicted distribution to be consistent with the actual distribution shape; the temporal continuity constraint loss is used to suppress prediction jumps between adjacent future time steps that do not conform to the continuity of warehousing operations; the anomalous time window weighted loss is used to increase the model's learning weight for anomalous scenarios such as inbound congestion, outbound queuing, external transportation delays, and cold chain environment anomalies.

[0021] The server performs reverse parameter updates on the time-series large model to be trained based on the training loss until the training loss meets the preset convergence condition, thus obtaining the pre-constructed time-series large model. Through the above training process, the pre-constructed time-series large model can learn the correspondence between environmental periodic changes, cargo turnover trend changes, and external impulse event decay changes and future cargo dwell time distribution, making the dwell time distribution parameter set generated in the prediction stage have a trainable source and verifiable distribution label basis. Specifically, , , , , ,in, This represents the total number of time steps within a future time window. This indicates the total number of intervals in the stay duration. Indicates the true interval probability label. This represents the interval probability predicted by the model. , , , These represent the expected value, variance, skewness, and kurtosis of the predicted distribution, respectively. , , , Let these represent the expected value, variance, skewness, and kurtosis of the true distribution, respectively. This represents the set of parameters for predicting the retention distribution. Labels representing the true retention distribution parameters. This represents the weighting coefficient for the abnormal time window. , , , This represents the loss weight parameter.

[0022] In this embodiment of the invention, the step of mapping the environmental collection timestamp in the environmental state time series, the cargo event timestamp in the cargo turnover event stream, and the external event timestamp in the external influence event record to time anchors on the warehousing time axis according to the preset timestamp mapping relationship, and generating a hybrid data stream with a unified time anchor mark based on the time anchors, can be implemented through the following example.

[0023] Extract environmental collection timestamps one by one from the environmental state time series, input the environmental collection timestamps into a preset timestamp mapping relationship, and obtain the environmental time anchor point corresponding to the environmental collection timestamp on the storage time axis; Extract the timestamps of each cargo event from the cargo turnover event stream, input the timestamps of each cargo event into the preset timestamp mapping relationship, and obtain the cargo time anchor point corresponding to the timestamp of each cargo event on the warehousing time axis. Extract the timestamps of external events one by one from the external impact event records, input the timestamps of external events into the preset timestamp mapping relationship, and obtain the external time anchor points corresponding to the external event timestamps on the warehouse time axis. Traverse each time anchor point on the storage timeline. When the environmental time anchor point, the cargo time anchor point, or the external time anchor point hits the current time anchor point, mark the hit environmental state record, cargo turnover event record, or external impact event record into the data container corresponding to the current time anchor point. When there are only data records from a single source in the data container, a preset adjacent time anchor point interpolation compensation strategy is adopted to extract data records of the same type from the data containers corresponding to the previous and next time anchor points adjacent to the current time anchor point, and to fill the data container of the current time anchor point with data compensation. When there are two different sources of data records in the data container, a preset dual-source data alignment strategy is invoked. The time anchor point on the storage time axis is used as the alignment benchmark to merge the two different sources of data records into a dual-source mixed record according to the order of the time anchor points. When there are data records from three different sources in the data container, a preset three-source data alignment strategy is invoked. The time anchor point on the storage time axis is used as the alignment benchmark, and the data records from the three different sources are merged into a three-source mixed record according to the order of the time anchor points. The dual-source hybrid record or triple-source hybrid record corresponding to each time anchor point on the storage time axis is marked to generate a hybrid record unit with time anchor point marking. All hybrid record units with time anchor point marking corresponding to all time anchor points are arranged sequentially along the storage time axis to generate a hybrid data stream with unified time anchor point marking.

[0024] In this embodiment of the invention, when the server generates a mixed data stream with a unified time anchor mark, it reads the environmental collection timestamp in the environmental state time series, the cargo event timestamp in the cargo turnover event stream, and the external event timestamp in the external influence event record, and maps the three types of timestamps to environmental time anchors, cargo time anchors, and external time anchors respectively through a preset timestamp mapping relationship.

[0025] The server uses each time anchor point on the warehouse timeline as an alignment reference, writing the matched environmental state records, cargo turnover event records, or external impact event records into the corresponding data containers. When the current time anchor point corresponds to a data record from only a single source, the server extracts data records of the same type from adjacent time anchor points for compensation filling; when the current time anchor point corresponds to data records from two or three different sources, the server merges the corresponding data records into a mixed record unit according to the time anchor point order.

[0026] The server arranges all hybrid record units sequentially along the storage timeline, generating a hybrid data stream with uniform time anchors. This hybrid data stream preserves the temporal correspondence between environmental sources, cargo sources, and external event sources, providing an input basis for subsequent multi-source temporal decoupling by the time-series decomposition network.

[0027] In this embodiment of the invention, the invocation of a pre-built time-series decomposition network to decouple the hybrid data stream in time series and generate environmental periodic basis components, cargo trend basis components and external impulse excitation components can be implemented through the following example.

[0028] The mixed data stream with unified time anchor marks is input into the mixed signal separation layer of the pre-constructed time-series decomposition network. The environmental sensing signal extraction channel in the mixed signal separation layer is called to filter the environmental dimension features of the mixed data stream and extract the environmental state signal subsequence. The cargo turnover signal extraction channel in the hybrid signal separation layer is invoked to filter the cargo dimension features of the hybrid data stream and extract the cargo turnover signal subsequence; The external event signal extraction channel in the hybrid signal separation layer is invoked to perform external dimension feature filtering on the hybrid data stream and extract external event signal subsequences; The environmental state signal subsequence is input into the periodic component analysis layer of the pre-constructed time-series decomposition network. The multi-scale convolution decomposition structure in the periodic component analysis layer is called to separate the environmental state signal subsequence layer by layer, and the periodic fluctuation component in the environmental state signal subsequence is extracted as the environmental periodic basis component. The cargo turnover signal subsequence is input into the trend component analysis layer of the pre-constructed time series decomposition network. The adaptive sliding window decomposition structure in the trend component analysis layer is called to extract the segmented trend of the cargo turnover signal subsequence. The results of each segmented trend extraction are aggregated to obtain the cargo trend base component. The external event signal subsequence is input into the pulse component parsing layer of the pre-constructed time-series decomposition network. The pulse event localization structure in the pulse component parsing layer is called to perform event trigger boundary detection on the external event signal subsequence, and the pulse signal segments in the external event signal subsequence are segmented as external pulse excitation components.

[0029] In this embodiment of the invention, the server inputs a hybrid data stream with uniform time anchor markers into a pre-constructed temporal decomposition network. The hybrid signal separation layer of the temporal decomposition network extracts environmental state signal subsequences, cargo turnover signal subsequences, and external event signal subsequences according to the source markers and field types in the hybrid record units.

[0030] The server inputs the environmental state signal subsequence into the periodic component analysis layer, and uses a multi-scale convolution decomposition structure to remove periodic fluctuation components to obtain the environmental periodic base component; it inputs the cargo turnover signal subsequence into the trend component analysis layer, and uses an adaptive sliding window decomposition structure to extract segmented trends to obtain the cargo trend base component; it inputs the external event signal subsequence into the pulse component analysis layer, and uses a pulse event localization structure to detect event trigger boundaries and segment pulse signal segments to obtain the external pulse excitation component.

[0031] Through the above decoupling process, the server separates the periodic, trend-based, and sudden effects superimposed in the mixed data stream into three basic components with different temporal meanings. In this embodiment of the invention, the environmental periodic basis components are projected onto the periodic phase space through a pre-constructed periodic phase alignment network to generate environmental periodic phase feature trajectories. This can be implemented through the following example.

[0032] The environmental periodic basis components are input into the basis frequency extraction layer of the pre-constructed periodic phase alignment network, and the autocorrelation function calculation structure in the basis frequency extraction layer is called to perform autocorrelation operation on the environmental periodic basis components to generate an autocorrelation coefficient sequence of the environmental periodic basis components under different time delays. The periodic peak detection structure in the base frequency extraction layer is invoked to perform local peak search on the autocorrelation coefficient sequence, extract the delay step parameter corresponding to the peak position in the autocorrelation coefficient sequence, and use the frequency corresponding to the delay step parameter as the environmental base frequency. The environmental periodic basis component and the environmental basis frequency are input into the phase mapping layer of the pre-constructed periodic phase alignment network. The sine and cosine projection structure in the phase mapping layer is called to perform joint projection of the sine and cosine phase components of each time point in the environmental periodic basis component, generating a phase coordinate pair corresponding to each time point. The phase expansion structure in the phase mapping layer is invoked to perform phase angle continuity processing on the phase coordinate pairs, converting the phase coordinate pairs into a continuous phase angle sequence; The continuous phase angle sequence is input into the phase trajectory generation layer of the pre-constructed periodic phase alignment network, and the phase state space encoding structure in the phase trajectory generation layer is called to reconstruct the phase space of the continuous phase angle sequence, so as to obtain the environmental periodic phase feature trajectory arranged in time order in the periodic phase space.

[0033] In this embodiment of the invention, the server inputs the environmental periodic basis components into a pre-constructed periodic phase alignment network. The basis frequency extraction layer of the periodic phase alignment network performs autocorrelation analysis on the environmental periodic basis components and obtains the environmental basis frequency through periodic peak detection.

[0034] The server invokes the phase mapping layer based on the environmental base frequency to perform joint projection of the sine and cosine phase components at each time point in the environmental periodic base component, obtaining phase coordinate pairs. The server further invokes the phase expansion structure to perform phase angle continuity processing on the phase coordinate pairs, generating a continuous phase angle sequence.

[0035] The server inputs the continuous phase angle sequence into the phase trajectory generation layer, and combines and encodes the continuous phase angle, time anchor point and environmental periodic intensity through the phase state space coding structure to obtain the environmental periodic phase feature trajectory arranged in time order in the periodic phase space.

[0036] In this embodiment of the invention, the cargo trend base component is projected onto the trend slope space through a pre-built trend parsing network to generate a cargo turnover trend slope feature trajectory, which can be implemented through the following example.

[0037] The cargo trend base component is input into the trend window segmentation layer of the pre-built trend parsing network, and the variable length window generation structure in the trend window segmentation layer is called to perform sliding window division on the cargo trend base component, generating multiple trend analysis window segments with variable lengths and partial overlap. The multiple trend analysis window segments are input into the local slope estimation layer of the pre-constructed trend parsing network. The linear fitting structure in the local slope estimation layer is called to perform local linear fitting processing on each trend analysis window segment to obtain the local trend slope parameter corresponding to each trend analysis window segment. The slope direction quantization structure in the local slope estimation layer is invoked to perform sign determination and amplitude normalization on the local trend slope parameter, and the local trend slope parameter is converted into a trend direction category label and a trend intensity label. The trend direction category label and the trend intensity label are input into the slope trajectory tracking layer of the pre-built trend parsing network. The temporal chain connection structure in the slope trajectory tracking layer is called to concatenate the trend direction category label and trend intensity label of adjacent trend analysis window segments in time order to generate a cargo turnover trend slope feature trajectory containing the slope change path.

[0038] In this embodiment of the invention, for example, after obtaining the cargo trend base component, the server inputs the cargo trend base component into a pre-constructed trend analysis network. The cargo trend base component originates from the trend component analysis results of the cargo turnover signal subsequence and is used to characterize the changing trends of operations such as warehousing, shelving, picking, verification, and outbound at continuous time anchor points within the warehouse. The server first calls the trend window segmentation layer of the trend analysis network and performs sliding window division on the cargo trend base component through the variable-length window generation structure in the trend window segmentation layer. The server dynamically determines the window length based on the cargo turnover rate and event density. Shorter trend analysis window segments are generated during periods of concentrated release of promotional orders, concentrated arrival of warehousing vehicles, or rapid accumulation of outbound tasks to capture rapid changes in cargo turnover pressure; longer trend analysis window segments are generated during periods of low nighttime activity, stable replenishment, or stable outbound volume to extract a smoother overall trend. Some overlapping areas are retained between adjacent trend analysis window segments, enabling the server to maintain trend continuity during window switching and avoid misjudgment due to trend abrupt changes caused by window boundaries.

[0039] The server inputs multiple trend analysis window segments of variable length with partial overlap into the local slope estimation layer of the trend parsing network. The local slope estimation layer invokes a linear fitting structure to perform local linear fitting on each trend analysis window segment. The server uses the inventory turnover intensity value corresponding to each time anchor point within the trend analysis window segment as the fitting object, calculating the slope of the inventory turnover state as it increases or decreases over time within that window segment. For example, during a morning period, the number of incoming scans and the number of goods waiting to be shelved continuously increase, and the server obtains a positive and large local trend slope parameter through linear fitting; after the concentrated outbound shipments in the afternoon, the number of picking tasks gradually decreases, and the server obtains a negative local trend slope parameter; during regular inventory consolidation periods, inventory turnover changes little, and the server obtains a local trend slope parameter close to zero. The local trend slope parameter can reflect the rate of increase, release, or stability of inventory turnover pressure within different time windows.

[0040] The server further invokes the slope direction quantization structure in the local slope estimation layer to perform sign determination and amplitude normalization on the local trend slope parameters. Based on the slope sign, the server converts the local trend slope parameters into trend direction category labels, which include upward trend, downward trend, and stationary trend. Based on the slope amplitude, the server normalizes the intensity of change in different window segments to a uniform scale and converts it into trend intensity labels, which represent the strength of changes in cargo turnover pressure. For example, the server labels window segments with a rapid increase in outbound scan volume within multiple consecutive time anchor points as strong upward trends, window segments with a slow decrease in inbound tasks as weak downward trends, and window segments with insignificant changes in turnover volume as stationary trends. Through direction quantization and intensity labeling, the server converts the original numerical slope into trend semantic features suitable for understanding by subsequent large-scale time-series models.

[0041] Finally, the server inputs the trend direction category label and the trend intensity label into the slope trajectory tracking layer of the trend analysis network. The slope trajectory tracking layer invokes a time-series chain connection structure to concatenate the trend direction category labels and trend intensity labels corresponding to adjacent trend analysis window segments according to the warehousing time axis sequence. It also smoothly connects trend turning points by combining overlapping areas between adjacent windows, generating a cargo turnover trend slope feature trajectory containing the slope change path. This cargo turnover trend slope feature trajectory can represent the complete change process of cargo turnover pressure from a slow increase to rapid accumulation, and then from concentrated release to gradual recovery to stability, providing a basis for trend changes in subsequent time-series large-scale models to determine the cargo retention distribution parameter set within future time windows.

[0042] In this embodiment of the invention, the external pulse excitation component is projected onto the pulse intensity decay space through a pre-constructed impulse response analysis network to generate an external event pulse decay feature trajectory, which can be implemented through the following example.

[0043] The external pulse excitation component is input into the pulse leading edge detection layer of the pre-constructed pulse response parsing network. The rising edge triggering recognition structure in the pulse leading edge detection layer is called to locate the pulse start boundary of the external pulse excitation component and extract the pulse start timestamp and pulse start amplitude parameters. The external pulse excitation component is input into the pulse decay modeling layer of the pre-constructed pulse response analysis network. The decay envelope fitting structure in the pulse decay modeling layer is called to fit the amplitude decay envelope curve of the external pulse excitation component from the pulse start timestamp, generating a pulse decay envelope sequence. The attenuation parameter extraction structure in the pulse attenuation modeling layer is invoked to extract the features of the attenuation rate parameter and the attenuation steady-state parameter of the pulse attenuation envelope sequence, thereby obtaining the pulse attenuation rate parameter and the pulse attenuation steady-state parameter. The pulse start amplitude parameter, the pulse decay rate parameter, and the pulse decay steady-state parameter are input into the decay state projection layer of the pre-constructed pulse response analysis network. The decay state space mapping structure in the decay state projection layer is called to jointly map the pulse start amplitude parameter, the pulse decay rate parameter, and the pulse decay steady-state parameter to the decay state coordinate point in the pulse intensity decay space. Starting from the attenuation state coordinate point, the attenuation state coordinates corresponding to each time step are gradually deduced in the pulse intensity attenuation space along the time step direction according to the preset attenuation state update function. The attenuation state coordinates obtained by gradual deduction are connected in time order to generate the attenuation characteristic trajectory of external event pulse.

[0044] In this embodiment of the invention, the server inputs the external pulse excitation component into a pre-constructed impulse response analysis network. The pulse leading edge detection layer of the impulse response analysis network locates the pulse start boundary of the external pulse excitation component and extracts the pulse start timestamp and pulse start amplitude parameters.

[0045] The server inputs the external pulse excitation component into the pulse decay modeling layer, performs decay envelope fitting starting from the pulse start timestamp, generates a pulse decay envelope sequence, and extracts the pulse decay rate parameter and pulse decay steady-state parameter from the pulse decay envelope sequence.

[0046] The server maps the pulse start amplitude parameter, the pulse decay rate parameter, and the pulse decay steady-state parameter to the pulse intensity decay space, and performs decay state deduction along the time step direction to generate the external event pulse decay characteristic trajectory.

[0047] In this embodiment of the invention, the step of inputting the environmental cycle phase feature trajectory, the cargo turnover trend slope feature trajectory, and the external event pulse decay feature trajectory into a pre-constructed time series model to perform cross-channel time dependency joint modeling and generate warehousing timeliness prediction results can be implemented through the following example.

[0048] The environmental periodic phase feature trajectory, the cargo turnover trend slope feature trajectory, and the external event pulse decay feature trajectory are respectively input into the first channel encoding structure, the second channel encoding structure, and the third channel encoding structure of the pre-constructed time series large model. The first channel encoding structure is called to perform periodic phase encoding on the environmental periodic phase feature trajectory to generate an environmental periodic phase encoding vector sequence. The second channel encoding structure is called to perform trend slope encoding on the cargo turnover trend slope feature trajectory to generate a cargo trend slope encoding vector sequence. The third channel encoding structure is called to perform pulse decay encoding on the external event pulse decay feature trajectory to generate an external pulse decay encoding vector sequence. The environmental cycle phase encoding vector sequence, the cargo trend slope encoding vector sequence, and the external pulse decay encoding vector sequence are input into the cross-channel time dependency interaction layer of the pre-built time series model. The multi-head cross-attention structure in the cross-channel time dependency interaction layer is called to perform cross-attention calculation between the three channels of the environmental cycle phase encoding vector sequence, the cargo trend slope encoding vector sequence, and the external pulse decay encoding vector sequence, generating an interaction encoding vector sequence containing inter-channel dependencies. The interactive encoded vector sequence is input into the temporal state evolution layer of the pre-built temporal large model, and the gated temporal evolution structure in the temporal state evolution layer is called to perform time-step recursive state update processing on the interactive encoded vector sequence to generate a temporal state vector sequence carrying cross-channel historical memory. The time-series state vector sequence is input into the time-effect distribution prediction layer of the pre-built time-series large model. The probability distribution generation structure in the time-effect distribution prediction layer is called to perform time-effect probability distribution mapping on the time-series state vector sequence, generating the retention distribution parameter set corresponding to each time step in the future time window. The set of retention distribution parameters corresponding to each time step in the future time window is aggregated, and the aggregated probability distribution parameter set is used as the storage timeliness prediction result.

[0049] In this embodiment of the invention, the server inputs the environmental cycle phase feature trajectory, the cargo turnover trend slope feature trajectory, and the external event pulse decay feature trajectory into a pre-constructed time-series large model. The time-series large model generates environmental cycle phase encoded vector sequences, cargo trend slope encoded vector sequences, and external pulse decay encoded vector sequences through a first-channel encoded structure, a second-channel encoded structure, and a third-channel encoded structure, respectively.

[0050] The server inputs three types of encoded vector sequences into the cross-channel time dependency interaction layer to obtain an interactive encoded vector sequence; the interactive encoded vector sequence is then input into the time-series state evolution layer to obtain a time-series state vector sequence; and the time-series state vector sequence is input into the time-series distribution prediction layer to generate a set of retention distribution parameters corresponding to each time step in the future time window.

[0051] The server performs parameter aggregation on the set of retention distribution parameters corresponding to each time step within a future time window to obtain the warehousing timeliness prediction result. This warehousing timeliness prediction result is used to characterize the distribution of cargo retention time within the future time window.

[0052] In this embodiment of the invention, the method further includes: Obtain the attribute description information of the stored goods, which includes the shelf life constraint parameters, storage temperature zone requirement parameters, and stacking restriction parameters. Input the shelf life constraint parameters in the warehouse goods attribute description information into a preset shelf life decay weight mapper to generate a shelf life sensitivity weight curve. Input the cargo storage temperature zone requirement parameters from the cargo attribute description information into a preset temperature zone constraint mapper to generate temperature zone adaptation tolerance boundary parameters. The cargo stacking limitation parameters in the warehouse cargo attribute description information are input into a preset stacking pressure mapper to generate a stacking pressure attenuation coefficient sequence; The sensitivity weight curve of the shelf life of goods, the boundary parameter of the temperature zone adaptation tolerance, and the sequence of stacking pressure attenuation coefficients are superimposed on the set of retention distribution parameters corresponding to each time step in the storage timeliness prediction result to generate a storage timeliness prediction result that includes correction of goods attribute constraints.

[0053] In this embodiment of the invention, the method further includes: Obtain the spatial topology information of the warehouse area, which includes warehouse shelf node identifiers, warehouse logistics channel network, and warehouse picking path distance matrix; Input the warehouse rack node identifiers and warehouse logistics channel network from the warehouse area spatial topology information into a preset warehouse space graph builder to generate a warehouse space connectivity graph with warehouse rack node identifiers as graph nodes and warehouse logistics channel network as graph edges. The warehouse space connectivity graph and the warehouse picking path distance matrix are input into a preset graph attention network. The graph attention network is then called to perform neighbor node distance weighted aggregation processing on each graph node in the warehouse space connectivity graph to generate a spatial topology embedding vector for each graph node. The spatial topology of each graph node is embedded into a vector input to a preset outbound efficiency mapping layer to generate an outbound time distribution curve corresponding to each graph node. The outbound time distribution curve corresponding to the graph node where the target goods are located in the warehousing timeliness prediction result is located is jointly convolved with the retention distribution parameter set to generate a warehousing timeliness prediction result that incorporates spatial topological constraints.

[0054] In this embodiment of the invention, for example, after generating the warehousing timeliness prediction result, the server further obtains the spatial topology information of the warehouse area and superimposes the impact of the warehouse area spatial layout on outbound efficiency into the warehousing timeliness prediction process. The warehouse area spatial topology information comes from the warehouse management system, the storage location management system, and the picking route planning system, including warehouse shelf node identifiers, the warehouse logistics channel network, and the warehouse picking route distance matrix. The server reads the warehouse shelf node identifiers corresponding to each shelf, temporary storage area, verification station, sorting port, outbound port, and loading platform, and reads the passable logistics channel connections between each node to form the warehouse logistics channel network. The server also reads the path distances from different warehouse shelf nodes to the picking start point, verification area, packing area, and outbound port to form the warehouse picking route distance matrix. Through the above processing, the server can clearly determine the location, connectivity, and path cost of the target goods' storage location within the entire warehouse space to reach the outbound stage.

[0055] The server inputs the warehouse rack node identifiers and the warehouse logistics channel network into a preset warehouse space map builder to generate a warehouse space connectivity graph. This connectivity graph uses warehouse rack node identifiers as graph nodes and the warehouse logistics channel network as graph edges to represent the spatial connectivity relationships between various racks, aisles, picking areas, and outbound areas within the warehouse. For example, the server constructs cold chain zone 1 rack A03, ambient temperature zone 2 rack B12, verification station R01, and outbound exit D02 as graph nodes, and establishes graph edges based on the connections between forklift aisles, manual picking aisles, and conveyor lines. When a warehouse area has narrow aisles, long detour distances, or few connecting paths to the outbound exit, the spatial accessibility of the corresponding graph node in the warehouse space connectivity graph is relatively low.

[0056] The server inputs the warehouse space connectivity graph and the warehouse picking path distance matrix into a preset graph attention network. The graph attention network performs weighted aggregation of neighbor node distances on each graph node in the warehouse space connectivity graph. Based on the path distance between the current graph node and its neighboring graph nodes, the number of connected passages, and the reachability distance to the outbound node, the server calculates the spatial influence weight of neighboring graph nodes on the current graph node, and performs weighted aggregation of the spatial features of neighboring nodes to generate a spatial topology embedding vector for each graph node. For shelf nodes that are close to the verification station and the outbound exit and have sufficient passage connectivity, the server generates a spatial topology embedding vector indicating a high level of outbound convenience; for shelf nodes located deep in the warehouse area, with many detours or requiring passage through congested passages, the server generates a spatial topology embedding vector indicating a high level of outbound time.

[0057] The server inputs the spatial topology embedding vector of each graph node into a preset outbound efficiency mapping layer to generate an outbound time distribution curve corresponding to each graph node. This outbound time distribution curve characterizes the probability distribution of the time required for a target product to be picked, handled, checked, packaged, and arrive at the outbound exit from its corresponding graph node. For example, based on the spatial topology embedding vector, the server identifies that shelf A03 in cold chain zone 1 is close to the outbound exit and has smooth passageways, generating an outbound time distribution curve with high concentration and short processing time. Conversely, the server identifies that shelf B12 in ambient temperature zone 2 is located deep within the warehouse and requires traversing multiple public passageways, generating an outbound time distribution curve with a higher mean and a wider fluctuation range.

[0058] Finally, the server performs a joint convolution process with the set of storage time distribution parameters corresponding to the graph node where the target goods are located, based on the storage time prediction results. In each future time step, the server fuses the set of storage time distribution parameters, jointly predicted by environmental cycles, goods turnover trends, and external event impulses, with the set of outbound time distribution parameters corresponding to the spatial node where the target goods are located. This allows the prediction results to simultaneously reflect storage risk in the time dimension and outbound path cost in the spatial dimension. Through this joint convolution process, the server generates a storage time prediction result that incorporates spatial topology constraints. This storage time prediction result, incorporating spatial topology constraints, can distinguish the differences in outbound time for goods in different storage locations within the same time window, providing a more accurate predictive basis for subsequent storage location adjustments, picking wave planning, route scheduling, and priority outbound control.

[0059] In this embodiment of the invention, the method further includes: Obtain warehousing operation scheduling plan information, which includes inbound operation scheduling table, outbound operation scheduling table, and intra-warehouse transfer operation scheduling table; Input the inbound operation schedule, the outbound operation schedule, and the in-warehouse transfer operation schedule into a preset operation conflict relationship graph builder to generate an operation conflict relationship graph with operation tasks as nodes and resource contention relationships between operation tasks as connecting edges. The job conflict relationship diagram is input into a preset job resource scheduling simulator. The conflict resolution strategy in the job resource scheduling simulator is called to perform resource allocation simulation on the job conflict relationship diagram and generate the actual execution time sequence plan of the job. The actual execution time schedule of the job is input into a preset job delay estimator to generate a job delay distribution curve for each job task; The storage timeliness prediction result is generated by time-series superposition and fusion of the set of retention distribution parameters corresponding to each time step in the storage timeliness prediction result with the operation delay distribution curve.

[0060] In this embodiment of the invention, the method further includes: Obtain warehousing operation scheduling plan information, which includes inbound operation scheduling table, outbound operation scheduling table, and intra-warehouse transfer operation scheduling table; Input the inbound operation schedule, the outbound operation schedule, and the in-warehouse transfer operation schedule into a preset operation conflict relationship graph builder to generate an operation conflict relationship graph with operation tasks as nodes and resource contention relationships between operation tasks as connecting edges. The job conflict relationship diagram is input into a preset job resource scheduling simulator. The conflict resolution strategy in the job resource scheduling simulator is called to perform resource allocation simulation on the job conflict relationship diagram and generate the actual execution time sequence plan of the job. The actual execution time schedule of the job is input into a preset job delay estimator to generate a job delay distribution curve for each job task; The storage timeliness prediction result is generated by time-series superposition and fusion of the set of retention distribution parameters corresponding to each time step in the storage timeliness prediction result with the operation delay distribution curve.

[0061] In this embodiment of the invention, the method further includes: The environmental periodic basis component is subjected to frequency domain decomposition granularity adaptive adjustment processing. The environmental periodic basis component is input into a preset multi-granularity frequency domain decomposition structure. The wavelet packet transform layer in the multi-granularity frequency domain decomposition structure is called to perform multi-level frequency band division on the environmental periodic basis component, generating a sub-band energy distribution sequence of the environmental periodic basis component in different frequency bands. The energy concentration evaluation layer in the multi-granularity frequency domain decomposition structure is invoked to perform energy concentration analysis on the sub-band energy distribution sequence, locate the target frequency band number with the highest energy concentration, and feed back the target frequency band number to the frequency band division parameter register of the wavelet packet transform layer, triggering the wavelet packet transform layer to adjust the division granularity parameter of the next round of frequency band division according to the target frequency band number; The sub-band energy distribution sequence regenerated after adjusting the frequency band division granularity parameters is input into the preset main frequency band tracking layer. The continuous time window correlation analysis structure in the main frequency band tracking layer is called to perform frequency band migration trajectory tracking on the sub-band energy distribution sequences of adjacent time windows, and generate the main frequency band migration trajectory sequence. The main frequency band migration trajectory sequence is input into a preset frequency band drift compensation layer. The frequency band phase correction structure in the frequency band drift compensation layer is called to perform phase continuity compensation processing on the frequency band jump points in the main frequency band migration trajectory sequence, thereby generating a frequency band drift compensated main frequency band phase sequence. The frequency-drift-compensated main frequency band phase sequence replaces the original input of the environmental periodic basis component in the pre-constructed periodic phase alignment network. The frequency-drift-compensated main frequency band phase sequence is then projected onto the periodic phase space through the pre-constructed periodic phase alignment network to generate the frequency-adaptive calibrated environmental periodic phase feature trajectory. The environmental periodic phase feature trajectory after frequency adaptive correction is used to replace the environmental periodic phase feature trajectory and input into the pre-constructed time series large model for joint modeling of cross-channel time dependency.

[0062] In this embodiment of the invention, for example, after obtaining the environmental cycle basis components, the server further performs frequency domain decomposition granularity adaptive adjustment processing on the environmental cycle basis components, so that the cyclical changes in the storage environment can be analyzed at a more suitable frequency scale. The server inputs the environmental cycle basis components into a preset multi-granularity frequency domain decomposition structure and calls the wavelet packet transform layer in the multi-granularity frequency domain decomposition structure to perform multi-level frequency band division on the environmental cycle basis components. According to the continuous time window on the storage time axis, the server decomposes the temperature cycle fluctuation, humidity cycle fluctuation, short-period disturbances caused by the opening and closing of the warehouse door, and medium-period disturbances caused by the start and stop of the refrigeration unit into different frequency bands, generating a sub-band energy distribution sequence of the environmental cycle basis components in different frequency bands. For example, when the cold chain warehouse is running stably at low frequency in the early morning, the environmental cycle energy is mainly concentrated in the low frequency band; when the warehouse is receiving goods in the morning, the warehouse door is frequently opened, and the refrigeration unit is running continuously for compensation, the environmental cycle energy shifts to the mid-to-high frequency band. The server expresses this frequency domain change through the sub-band energy distribution sequence.

[0063] The server invokes the energy concentration evaluation layer in the multi-granularity frequency domain decomposition structure to perform energy concentration analysis on the sub-band energy distribution sequence. The server calculates the energy proportion and energy stability of each frequency band within a continuous time window, locates the target frequency band number with the highest energy concentration, and feeds back the target frequency band number to the frequency band partitioning parameter register of the wavelet packet transform layer. Upon receiving the target frequency band number, the frequency band partitioning parameter register triggers the wavelet packet transform layer to adjust the partitioning granularity parameters for the next round of frequency band partitioning based on the target frequency band number. For frequency bands with highly concentrated and rapidly changing energy, the server increases the partitioning granularity near that frequency band; for frequency bands with dispersed energy and weak changes, the server decreases the partitioning granularity of the corresponding frequency band. Through this feedback adjustment, the server can avoid the problem of insufficient identification of the main cycle of the warehousing environment due to fixed frequency band partitioning.

[0064] The server inputs the regenerated sub-band energy distribution sequence, after adjusting the frequency band division granularity parameters, into the preset main frequency band tracking layer. The main frequency band tracking layer invokes a continuous time window correlation analysis structure to track the frequency band migration trajectory of the sub-band energy distribution sequences in adjacent time windows. The server compares the positional changes, energy variation amplitudes, and frequency band correlations of the main energy frequency band in adjacent time windows to determine the path of change in the environmental cycle's main frequency band, whether it migrates from the low frequency band to the mid frequency band, from the mid frequency band to the high frequency band, or remains stable, generating a main frequency band migration trajectory sequence. For example, after a warehouse transitions from a low-load nighttime state to a concentrated daytime operation state, the frequency of temperature and humidity regulation significantly increases, and the server records the process of the main frequency band moving to a higher frequency band in the main frequency band migration trajectory sequence.

[0065] The server inputs the main frequency band migration trajectory sequence into a preset frequency band drift compensation layer and calls the frequency band phase correction structure in the frequency band drift compensation layer to perform phase continuity compensation processing on the frequency band jump points in the main frequency band migration trajectory sequence. The server identifies the phase break positions caused by the main frequency band switching between adjacent time windows, and performs phase correction on the frequency band jump points according to the frequency band phase difference before and after the switching, the time anchor interval, and the energy continuity relationship, generating a frequency band drift-compensated main frequency band phase sequence. The main frequency band phase sequence can maintain the continuity of phase expression when the environmental periodic frequency drifts, so that the subsequent periodic phase alignment process will not misidentify normal frequency band migration as abnormal environmental changes.

[0066] The server then replaces the original input of the environmental periodic basis component in the pre-constructed periodic phase alignment network with the frequency-drift-compensated dominant frequency band phase sequence. The server performs phase mapping, phase continuity, and phase state-space encoding on the frequency-drift-compensated dominant frequency band phase sequence through the periodic phase alignment network, projecting the dominant frequency band phase sequence onto the periodic phase space to generate a frequency-adaptively corrected environmental periodic phase feature trajectory. Compared to the environmental periodic phase feature trajectory directly generated based on the original environmental periodic basis component, the frequency-adaptively corrected environmental periodic phase feature trajectory can more accurately reflect the environmental periodic dominant frequency drift caused by changes in chiller operating frequency, chiller door opening and closing rhythm, and shift changes.

[0067] Finally, the server replaces the original environmental cycle phase feature trajectory with the frequency-adaptively corrected environmental cycle phase feature trajectory and inputs it into the pre-constructed time-series large model for joint modeling of cross-channel time dependencies. In subsequent joint modeling, the server combines the cargo turnover trend slope feature trajectory and the external event pulse decay feature trajectory, using the frequency-adaptively corrected environmental cycle information to predict the set of retention distribution parameters within future time windows. Through this processing, the server can maintain stable cycle characteristics when the warehousing environment cycle frequency dynamically changes with operational status, improving the adaptability of warehousing timeliness prediction results to different operational periods and different environmental adjustment states.

[0068] In this embodiment of the invention, the method further includes: The trend inflection point pre-identification processing is performed on the cargo trend base component. The cargo trend base component is input into a preset trend curvature analysis structure. The second-order difference convolution kernel in the trend curvature analysis structure is called to perform discrete curvature estimation processing on the cargo trend base component, generating the trend curvature sequence of the cargo trend base component at each time anchor point. The trend curvature sequence is input into a preset inflection point candidate positioning layer. The curvature zero-crossing detection structure in the inflection point candidate positioning layer is called to perform sign change boundary search on the trend curvature sequence. The time anchor points where the curvature sign flips are extracted as the trend inflection point candidate position set. The set of candidate trend inflection points is input into a preset inflection point confidence verification layer. The neighborhood trend consistency test structure in the inflection point confidence verification layer is called to perform a consistency comparison analysis of the trend direction in the neighborhood before and after each candidate trend inflection point. False candidate inflection points whose neighborhood trend direction consistency does not meet the screening conditions are eliminated to obtain a set of confirmed trend inflection point locations. The set of confirmed trend inflection point locations is input into a preset inflection point type classification layer. The slope comparison structure before and after the inflection point in the inflection point type classification layer is called to calculate the slope change magnitude of the forward and backward trend slopes of each confirmed trend inflection point location. Based on the slope change magnitude, the confirmed trend inflection point locations are divided into trend acceleration inflection point category, trend deceleration inflection point category, and trend reversal inflection point category. The set of confirmed trend inflection point locations and the inflection point category corresponding to each confirmed trend inflection point location are input into a preset inflection point-driven trend reshaping layer. The segmented trend reconstruction structure in the inflection point-driven trend reshaping layer is called to perform segmented trend reconstruction processing on the cargo trend base component with the confirmed trend inflection point location as the segment boundary, generating a segmented trend base component containing inflection point semantic tags. The segmented trend base component containing inflection point semantic tags is used to replace the cargo trend base component and input into the pre-built trend parsing network for trend slope space projection.

[0069] In this embodiment of the invention, for example, after obtaining the cargo trend base component, the server further performs trend inflection point pre-identification processing on the cargo trend base component, so that the positions where cargo turnover pressure accelerates, decelerates, or reverses can be marked in advance before entering the trend analysis network. The server inputs the cargo trend base component into a preset trend curvature analysis structure and calls the second-order difference convolution kernel in the trend curvature analysis structure to perform discrete curvature estimation processing on the cargo trend base component. The server sequentially reads the cargo trend values ​​corresponding to each time anchor point along the warehousing time axis, such as the quantity of goods waiting to be put on the shelf, the quantity of goods waiting to be picked, the quantity waiting for review, and the quantity of goods waiting to be loaded out of the warehouse. Based on the trend change amplitude between adjacent time anchor points, the server calculates the curvature response value of each time anchor point, generating a trend curvature sequence of the cargo trend base component at each time anchor point. The trend curvature sequence is used to characterize the changes in cargo turnover trend from flat to steep, from steep to flat, or from rising to falling.

[0070] The server inputs the trend curvature sequence into a preset inflection point candidate positioning layer and calls the curvature zero-crossing detection structure in the inflection point candidate positioning layer to perform a sign change boundary search on the trend curvature sequence. The server detects locations where the curvature sign changes from positive to negative, from negative to positive, or approaches zero within continuous time anchor points, and extracts the time anchor points where the curvature sign flips as a set of trend inflection point candidate locations. For example, after a concentrated arrival of inbound vehicles in the morning, the number of goods to be shelved changes from a slow increase to a rapid increase; the server extracts inflection point candidates at this trend acceleration location. After the outbound peak ends, the number of picking tasks changes from a continuous increase to a gradual decrease; the server extracts inflection point candidates at this trend reversal location.

[0071] The server inputs the set of candidate trend inflection points into a preset inflection point confidence verification layer and calls the neighborhood trend consistency check structure in the inflection point confidence verification layer to perform a consistency comparison analysis of the trend direction in the neighborhood before and after each candidate trend inflection point. The server calculates the forward trend direction in several time anchor points before the candidate position and the backward trend direction in several time anchor points after the candidate position, and determines whether there is a stable trend change relationship between the neighborhood before and after. For curvature jumps caused only by single scan anomalies, temporary data supplementation, or short-term noise, the server identifies insufficient consistency of the trend direction between the neighborhood before and after and removes the corresponding false inflection point candidate positions; for candidate positions with stable changes in the trend direction before and after, the server retains the corresponding time anchor points to obtain a set of confirmed trend inflection point positions.

[0072] The server inputs the set of confirmed trend inflection point locations into a preset inflection point type classification layer, and calls the slope comparison structure before and after the inflection point in the inflection point type classification layer to calculate the magnitude of the slope change for the forward and backward trend slopes of each confirmed trend inflection point location. Based on the magnitude and direction of the slope change, the server classifies the confirmed trend inflection point locations into three categories: trend acceleration inflection point, trend deceleration inflection point, and trend reversal inflection point. For example, when the quantity to be put on shelves changes from a slow increase to a rapid increase, the server marks it as a trend acceleration inflection point; when the outbound task processing speed decreases but still continues to grow, the server marks it as a trend deceleration inflection point; when the picking task volume changes from an increase to a decrease, the server marks it as a trend reversal inflection point.

[0073] Finally, the server inputs the confirmed trend inflection point location set and the inflection point category corresponding to each confirmed trend inflection point location into a preset inflection point-driven trend reshaping layer. The inflection point-driven trend reshaping layer invokes a segmented trend reconstruction structure, using the confirmed trend inflection point locations as segment boundaries, to perform segmented trend reconstruction processing on the cargo trend base components. Corresponding inflection point semantic tags are added to each segment boundary, generating segmented trend base components containing inflection point semantic tags. The server replaces the original cargo trend base components with these segmented trend base components containing inflection point semantic tags and inputs them into the pre-constructed trend parsing network for trend slope space projection. This ensures that the subsequently generated cargo turnover trend slope feature trajectory can not only express the trend direction and intensity but also the time position and type of key changes in cargo turnover pressure, thereby improving the predictive ability of the time-series large model for future warehouse congestion risk inflection points.

[0074] In this embodiment of the invention, the method further includes: The external pulse excitation component is modeled for pulse event chain propagation. The external pulse excitation component is input into a preset event chain relationship mining layer. The pulse time interval statistics structure in the event chain relationship mining layer is called to perform interval distribution statistics on the time interval between every two adjacent pulse events in the external pulse excitation component, and generate a pulse event interval time distribution histogram. The interval threshold segmentation structure in the event chain mining layer is invoked to perform valley threshold search on the pulse event interval time distribution histogram, and the segmentation interval threshold used to distinguish between chained pulse events and independent pulse events is extracted. The external pulse excitation component and the segmentation interval threshold are input into a preset chain event chain construction layer. The interval comparison link structure in the chain event chain construction layer is called to link adjacent pulse events with a time interval less than the segmentation interval threshold into a pulse event chain chain, thereby generating a set of pulse event chain chains. The set of pulse event chain chains is input into a preset chain propagation graph construction layer. The event type node mapping structure in the chain propagation graph construction layer is called to map each pulse event in the pulse event chain chain to a propagation graph node according to the event source type. Directed propagation connection edges are established between propagation graph nodes with chain relationships to generate a pulse event chain propagation topology graph. The pulse event chain propagation topology is input into a preset propagation path strength evaluation layer. The path frequency weighting structure in the propagation path strength evaluation layer is called to perform weighted fusion processing on the occurrence frequency and propagation delay time of each propagation path in the pulse event chain propagation topology to generate propagation path strength parameters. The pulse event chain propagation topology diagram carrying the propagation path strength parameters is input into a preset pulse chain attenuation modulation layer. The chain effect superposition structure in the pulse chain attenuation modulation layer is called to perform chain superposition correction on the attenuation curves of pulse events belonging to the same pulse event chain in the external pulse excitation component, thereby generating the chain-modulated external pulse excitation component. The external pulse excitation component after interlocking modulation is replaced by the external pulse excitation component and input into the pre-constructed pulse response analysis network for pulse intensity attenuation spatial projection.

[0075] In this embodiment of the invention, for example, after obtaining the external pulse excitation component, the server further performs pulse event chain propagation modeling processing on the external pulse excitation component, so that the continuous triggering relationship between multiple external impact events can be identified and used to correct the impact intensity of external events. The server inputs the external pulse excitation component into a preset event chain relationship mining layer, and calls the pulse time interval statistics structure in the event chain relationship mining layer to perform interval distribution statistics on the time interval between every two adjacent pulse events in the external pulse excitation component. The server sequentially reads the trigger times of pulse events such as rainstorm warnings, road congestion, temporary traffic control, vehicle arrival delays, and regional delivery interruptions along the warehouse time axis, calculates the interval duration between adjacent pulse events, and generates a pulse event interval time distribution histogram. The pulse event interval time distribution histogram can reflect whether external events occur densely and continuously or occur relatively independently of each other.

[0076] The server invokes the interval threshold segmentation structure in the event chaining mining layer to perform valley threshold search on the histogram of the pulse event interval time distribution. The server identifies low-frequency valley locations between short-interval event clusters and long-interval event clusters in the interval time distribution and extracts the time interval corresponding to these valley locations as the segmentation interval threshold. This segmentation interval threshold is used to distinguish between chained pulse events and independent pulse events. For example, if road flooding, main road congestion, and vehicle delays occur consecutively within a short period after a rainstorm warning is issued, the server identifies these short-interval events as potentially having a chain propagation relationship; for holiday order fluctuations and isolated road construction events occurring at longer intervals, the server identifies them as independent pulse events.

[0077] The server inputs the external pulse excitation component and the segmentation interval threshold into a preset chain event chain construction layer, and calls the interval comparison link structure in the chain event chain construction layer to link adjacent pulse events with a time interval less than the segmentation interval threshold into a pulse event chain, generating a set of pulse event chain chains. During this process, the server retains the chain direction according to the order in which the events are triggered. For example, it links "rainstorm warning - road congestion - vehicle arrival delay - increased outbound queue" into a single pulse event chain, so that the impact of external events on warehousing timeliness is no longer considered an isolated shock, but rather expressed as a continuous shock process with a propagation order.

[0078] The server inputs the set of pulse event chains into a preset chain propagation graph construction layer. The chain propagation graph construction layer invokes an event type node mapping structure to map each pulse event in each chain to a propagation graph node according to its event source type, such as weather nodes, transportation nodes, carrier nodes, and off-site arrival nodes. Directed propagation connections are established between propagation graph nodes with chain relationships, generating a pulse event chain propagation topology graph. The server then inputs this pulse event chain propagation topology graph into a preset propagation path strength evaluation layer and invokes a path frequency weighting structure to perform weighted fusion processing on the occurrence frequency and propagation delay time of each propagation path, generating propagation path strength parameters. Paths with higher occurrence frequency and shorter propagation delays are assigned higher propagation path strength, indicating that this type of external event chain has a stronger and more direct impact on warehousing timeliness.

[0079] The server further inputs the pulse event chain propagation topology map carrying the propagation path strength parameters into a preset pulse chain attenuation modulation layer. The pulse chain attenuation modulation layer invokes a chain effect superposition structure to perform chain superposition correction on the attenuation curves of multiple pulse events belonging to the same pulse event chain in the external pulse excitation component. The server increases the comprehensive impact intensity of continuous impact events based on the propagation path strength parameters and adjusts the superposition position of the attenuation curves according to the propagation delay between each pulse event, generating a chain-modulated external pulse excitation component. Finally, the server replaces the original external pulse excitation component with the chain-modulated external pulse excitation component and inputs it into the pre-constructed pulse response analysis network for pulse intensity attenuation spatial projection. This allows the subsequently generated external event pulse attenuation feature trajectory to simultaneously express the continuous impact caused by a single external event impact and the chain propagation of multiple external events, thereby improving the ability of warehouse timeliness prediction results to characterize complex external disturbance scenarios.

[0080] Please refer to the following: Figure 2 , Figure 2 A schematic diagram of an inter-channel causal lag-constrained attention structure provided in an embodiment of the present invention. In this embodiment, the method further includes: A cross-channel causal attribution structure is constructed within the pre-built time-series large model. The environmental cycle phase encoding vector sequence, the cargo trend slope encoding vector sequence, and the external pulse decay encoding vector sequence are input into the Granger causality test layer in the cross-channel causal attribution structure. The Granger causality test layer is called to perform Granger causality relationship testing on the environmental cycle phase encoding vector sequence, the cargo trend slope encoding vector sequence, and the external pulse decay encoding vector sequence between each pair of channels, generating an inter-channel causality matrix. The inter-channel causal relationship matrix is ​​input into the causal lag time estimation layer in the cross-channel causal attribution structure. The mutual information delay maximization structure in the causal lag time estimation layer is called to perform optimal causal lag time estimation processing on each channel pair with causal relationship in the inter-channel causal relationship matrix, thereby generating the inter-channel causal lag time matrix. The inter-channel causal relationship matrix and the inter-channel causal lag time matrix are input into the causal attention mask generation layer in the cross-channel causal attribution structure. The causal attention mask generation layer is then called to generate a causal constraint attention mask matrix for the multi-head cross-attention structure based on the inter-channel causal relationship matrix and the inter-channel causal lag time matrix. The causal constraint attention mask matrix is ​​injected into the cross-channel time dependency interaction layer of the pre-built temporal large model. Causal prior constraints are applied to the attention weight calculation process in the multi-head cross-attention structure of the cross-channel time dependency interaction layer, so that the multi-head cross-attention structure only allocates attention weights between channel pairs that have causal relationships and satisfy causal lag time.

[0081] In this embodiment of the invention, the Granger causality test layer uses the channel coding vector sequence as the test object. The server denotes the environmental cycle channel, cargo trend channel, and external pulse channel as different channels, and constructs the historical coding state of the source channel and the current coding state of the target channel for any two channels. The server first constructs a first prediction residual based on the historical coding state of the target channel itself, and then constructs a second prediction residual based on the historical coding state of the target channel itself and the historical coding state of the source channel. When the second prediction residual is less than the first prediction residual, and the decrease in the residual between the two exceeds a preset causal significance threshold, the server determines that there is a Granger causal relationship between the source channel and the target channel; otherwise, the server determines that there is no Granger causal relationship between the source channel and the target channel. The server generates an inter-channel causality matrix based on the determination results of all channel pairs, where the matrix elements represent whether there is a causal influence direction from the corresponding source channel to the target channel.

[0082] After generating the inter-channel causal relationship matrix, the server performs causal lag time estimation on channel pairs with causal relationships. Within a preset candidate lag time range, the server calculates the mutual information value between the historical encoding state of the source channel and the subsequent encoding state of the target channel, and determines the candidate lag time with the largest mutual information value as the causal lag time of the corresponding channel pair. For channel pairs without causal relationships, the server sets the corresponding causal lag time to a null value or a preset invalid flag. In this way, the server can obtain an inter-channel causal lag time matrix that simultaneously includes causal direction and impact delay. , , .in, This represents an element of the causal relationship matrix indicating that source channel i points to target channel j. This indicates the prediction residual obtained solely based on the historical coding state of the target channel itself. This represents the prediction residual obtained based on the historical coding states of both the target channel and the source channel. This indicates a preset threshold for causal significance. This represents the causal lag time between channel i and channel j. Indicates the maximum candidate lag time. Indicates the time lag of source channel i The mutual information value between the target channel j and the target channel j. This represents the elements of the causal lag time matrix between channels.

[0083] In this embodiment of the invention, the causal constraint attention mask matrix is ​​used to restrict channel interactions that lack causal relationships or do not meet the causal lag time in a multi-head cross-attention structure. For source channel i and target channel j, when the inter-channel causal relationship matrix indicates that there is no causal relationship between source channel i and target channel j, the server sets the attention mask value of the corresponding channel pair to a suppression value in the causal constraint attention mask matrix; when there is a causal relationship between source channel i and target channel j, but the time difference between the source channel time step s and the target channel time step t does not meet the causal lag time recorded in the causal lag time matrix, the server also sets the corresponding attention mask value to a suppression value; when there is a causal relationship between source channel i and target channel j, and the time difference between the source channel time step s and the target channel time step t meets the causal lag time, the server sets the corresponding attention mask value to an allow value.

[0084] When calculating attention weights in a multi-head cross-attention structure, the server superimposes the causal constraint attention mask matrix onto the attention scoring matrix. For positions set to suppression values, the server reduces the attention score to near-invalid, resulting in a near-zero attention weight after normalization. For positions set to allow values, the server retains the original attention score, ensuring that the multi-head cross-attention structure only performs effective information interaction between channel pairs that satisfy causal direction and causal lag constraints. Through this method, the server can reduce the interference of irrelevant synchronization fluctuations and future information leakage on warehouse delivery time prediction results. , .in, This represents the causal constraint attention mask value of source channel i at time step s pointing to target channel j at time step t. Elements representing the causal relationships between channels This represents the elements of the inter-channel causal lag time matrix. Indicates the allowable range of lag time deviation. This represents the attention weight of the corresponding channel pair and time step pair in the h-th attention head. This represents the query vector corresponding to target channel j at time step t. Let d represent the key vector corresponding to source channel i at time step s, and d represent the dimension of the attention vector.

[0085] In this embodiment of the invention, for example, the server constructs a cross-channel causal attribution structure within a pre-built time-series large model to constrain the influence direction and lag relationships between the environmental cycle channel, the cargo trend channel, and the external pulse channel. After completing channel encoding of the three types of feature trajectories in the time-series large model, the server obtains the environmental cycle phase encoding vector sequence, the cargo trend slope encoding vector sequence, and the external pulse decay encoding vector sequence, and inputs these three encoding vector sequences into the Granger causality test layer in the cross-channel causal attribution structure. The server calls the Granger causality test layer to perform Granger causality relationship testing on the three types of encoding vector sequences for each pair of channels, determining whether the historical state of one channel can improve the explanatory power of the future state of another channel. For example, if the server detects frequent warehouse door openings and chiller compensation operation in the environmental cycle phase encoding vector sequence, which can explain the decrease in picking efficiency and the accumulation of outbound tasks in the subsequent cargo trend slope encoding vector sequence, then a causal relationship is marked at the position where the environmental cycle channel points to the cargo trend channel. Similarly, if the server detects the residual impact of traffic control in the external pulse decay encoding vector sequence, which can explain the increase in outbound waiting in the subsequent cargo trend channel, then a causal relationship is marked at the position where the external pulse channel points to the cargo trend channel. The server generates a causal relationship matrix between channels based on all pairwise test results.

[0086] The server inputs the inter-channel causal relationship matrix into the causal lag time estimation layer of the cross-channel causal attribution structure. The causal lag time estimation layer invokes the mutual information delay maximization structure to perform optimal causal lag time estimation for each channel pair with a causal relationship in the inter-channel causal relationship matrix. The server calculates the information correlation strength between the historical encoding state of the source channel and the subsequent encoding state of the target channel within multiple candidate lag times, and selects the lag time with a higher level of mutual information as the optimal causal lag time for the corresponding channel pair. For example, if the server identifies that external traffic congestion events typically cause vehicle arrival delays after several time anchor points, further leading to a backlog of outbound tasks, then this is recorded as the lag time corresponding to the causal relationship between the external pulse channel and the cargo trend channel; if the server identifies that the impact of environmental temperature fluctuations on the verification efficiency of cold chain goods has a short lag, then this is recorded as a shorter lag time between the environmental cycle channel and the cargo trend channel. Based on this, the server generates the inter-channel causal lag time matrix.

[0087] The server inputs the inter-channel causal relationship matrix and the inter-channel causal lag time matrix into the causal attention mask generation layer in the cross-channel causal attribution structure. The causal attention mask generation layer generates a causal constraint attention mask matrix for the multi-head cross-attention structure based on whether a causal relationship exists between channels and whether the corresponding causal lag times satisfy the temporal constraints. For channel pairs without a causal relationship, the server sets a suppression flag in the causal constraint attention mask matrix to restrict the multi-head cross-attention structure from allocating attention weights between these channel pairs. For channel pairs with a causal relationship but whose time steps do not satisfy the causal lag relationship, the server also sets a suppression flag to prevent the model from using future states or irrelevant synchronization fluctuations for current predictions. For channel pairs with a causal relationship and whose time steps satisfy the lag constraints, the server sets an allow flag to enable effective interaction between the corresponding channels.

[0088] Finally, the server injects the causal constraint attention mask matrix into the cross-channel time dependency interaction layer of the large-scale time series model, and applies causal prior constraints to the attention weight calculation process in the multi-head cross-attention structure of the cross-channel time dependency interaction layer. When calculating the attention weights between the environmental periodic phase encoding vector sequence, the cargo trend slope encoding vector sequence, and the external pulse decay encoding vector sequence, the multi-head cross-attention structure only assigns weights between channel pairs that have a causal relationship and satisfy the causal lag time. Through the above processing, the server ensures that the cross-channel joint modeling process of the large-scale time series model no longer relies solely on statistical correlation, but is further constrained by the causal direction and time lag relationship between channels. This reduces interference from irrelevant channels and spurious correlations, improving the predictive stability and interpretability of the lingering distribution parameter set within the future time window.

[0089] In this embodiment of the invention, the method further includes: The distribution pattern perturbation analysis is performed on the retention distribution parameter set in the storage timeliness prediction result. The retention distribution parameter set is input into the preset distribution pattern analysis layer. The multi-moment statistical structure in the distribution pattern analysis layer is called to extract the statistical moments from the first to the fourth moment of the retention distribution parameter set, generating the distribution expectation moment parameter, distribution variance moment parameter, distribution skewness moment parameter and distribution kurtosis moment parameter. The expected moment parameter, variance moment parameter, skewness moment parameter, and kurtosis moment parameter are input into a preset distribution morphology mutation detection layer. The continuous time window moment parameter difference structure in the distribution morphology mutation detection layer is called to perform difference operations on the expected moment parameter, variance moment parameter, skewness moment parameter, and kurtosis moment parameter between adjacent time windows to generate a mutation intensity sequence. The mutation intensity sequence is input into a preset mutation pattern recognition layer. The mutation template matching structure in the mutation pattern recognition layer is called to perform template matching processing between the mutation intensity sequence and the preset warehouse abnormal event mutation template library. The abnormal mutation pattern fragments that are successfully matched in the mutation intensity sequence and the corresponding abnormal event type labels are identified. The abnormal mutation pattern fragment and the corresponding abnormal event type are labeled and input into the preset abnormal source inference layer. The multi-source signal backtracking structure in the abnormal source inference layer is called to retrieve the abnormal fluctuation fragments of the environmental periodic phase feature trajectory, the cargo turnover trend slope feature trajectory and the external event pulse decay feature trajectory in the corresponding time interval according to the time positioning of the abnormal mutation pattern fragment, and generate an abnormal fluctuation fragment set. The abnormal fluctuation fragment set and the abnormal event type label are input into a preset source attribution chain generation layer. The causal chain reasoning structure in the source attribution chain generation layer is called to perform cross-channel attribution logic chain construction processing on the abnormal fluctuation fragment set, generating an attribution source chain pointing from the abnormal event type label to the multi-source signal abnormal fluctuation fragment. The attribution source chain is then attached to the retention distribution parameter set of the corresponding time step in the warehouse timeliness prediction result.

[0090] Please refer to the following: Figure 3 , Figure 3 This is a schematic diagram of the abnormal mutation identification and attribution tracing structure provided in an embodiment of the present invention. In this embodiment, the mutation intensity sequence is obtained by weighting the changes in the expected moment parameter, variance moment parameter, skewness moment parameter, and kurtosis moment parameter between adjacent time windows. The server sets the moment parameter weights according to the sensitivity of different moment parameters to warehouse anomalies, and performs a weighted summation after normalizing the changes in each moment parameter to obtain the mutation intensity value corresponding to each time step. The mutation intensity value is used to characterize the degree of mutation in the stagnant distribution pattern of the current time step relative to adjacent time steps.

[0091] The warehouse anomaly event mutation template library stores mutation templates corresponding to multiple anomaly event types. Each mutation template includes template peak amplitude, template duration, template rise slope, template fall slope, and template long-tail retention segment. The server uses a sliding window to extract candidate mutation fragments from the mutation intensity sequence and calculates the template distance between each candidate mutation fragment and the mutation template corresponding to each anomaly event type. When the minimum template distance is less than the template matching threshold for the corresponding anomaly event type, the server identifies the candidate mutation fragment as a successfully matched anomaly mutation pattern fragment and uses the anomaly event type corresponding to the minimum template distance as the anomaly event type label for the anomaly mutation pattern fragment. When all template distances are not less than the corresponding template matching threshold, the server marks the candidate mutation fragment as an unidentified mutation fragment and does not generate an anomaly event type label. , , .in, This represents the mutation intensity value at time step t. , , , These represent the abrupt change weights of the distribution's expectation, variance, skewness, and kurtosis, respectively. This represents the mutation template corresponding to the m-th type of abnormal event. The template length represents the template length of the m-th type of mutation, and DTW represents the dynamic time warp distance. This indicates the template number of the most matching exception event. This indicates the corresponding template matching threshold. This indicates whether the candidate mutation fragments starting from the q-th time step are successfully matched.

[0092] In this embodiment of the invention, the causal chain inference structure constructs an attribution chain based on the occurrence time, intensity, channel type, and causal lag relationship between abnormal fluctuation segments. The server represents each abnormal fluctuation segment as a candidate attribution node, which includes the channel to which the abnormal fluctuation segment belongs, the start and end times of the abnormal fluctuation segment, the intensity of the abnormal fluctuation, and the direction of the abnormal fluctuation. The server uses the abnormal mutation pattern segment corresponding to the abnormal event type as the target node, and selects candidate attribution nodes from the set of abnormal fluctuation segments that are temporally earlier than the target node and satisfy the causal lag time constraint.

[0093] The server calculates an attribution correlation score based on the temporal matching degree between candidate attribution nodes and the target node, the strength of causal relationship between channels, the similarity of fluctuation intensity, and the suitability of abnormal event types. When the attribution correlation score is greater than a preset attribution threshold, the server establishes an attribution edge between the candidate attribution node and the target node. For multiple candidate attribution nodes, the server selects at least one candidate attribution node as the direct cause node according to the attribution correlation score from high to low, and continues to backtrack to retrieve the preceding abnormal fluctuation segment corresponding to the direct cause node, forming a multi-level attribution chain. The server arranges the multi-level attribution chain according to temporal order and causal direction, generating an attribution tracing chain pointing from the abnormal event type label to the abnormal fluctuation segment of the multi-source signal. ,.in, This represents the attribution association score between candidate attribution node a and target node b. This indicates the strength of the causal relationship between the channel to which candidate attribution node a belongs and the channel to which target node b belongs. This represents the time interval between candidate attribution node a and target node b. This indicates the causal lag time between corresponding channels. Indicates the time decay coefficient. This indicates the score for adapting to different event types. This indicates the score matching the intensity of abnormal fluctuations. , , , Indicates the attribution weight parameter. Indicates the attribution threshold. Indicates whether to establish an attribution edge between node a and node b.

[0094] In this embodiment of the invention, for example, after generating the storage timeliness prediction result, the server further performs distribution morphology perturbation analysis on the storage timeliness prediction result's retention distribution parameter set to identify possible abnormal retention mutations within future time windows and form traceable attribution information. The server first inputs the retention distribution parameter set corresponding to each future time step into a preset distribution morphology analysis layer, and calls the multi-moment statistical structure in the distribution morphology analysis layer to perform statistical moment extraction processing on the retention distribution parameter set from first to fourth moments. The server calculates the distribution expectation moment parameter based on the retention distribution parameter set, which represents the central level of the goods' retention time at that time step; calculates the distribution variance moment parameter, which represents the dispersion of retention time among different goods or different storage areas; calculates the distribution skewness moment parameter, which represents whether the retention time shifts towards longer retention; and calculates the distribution kurtosis moment parameter, which represents whether the retention distribution exhibits peaking or long-tailing. Through this processing, the server can determine from the perspective of probability distribution whether the storage timeliness is only changing slowly overall, or whether there is an abnormal concentration of delays in certain goods or time periods.

[0095] The server then inputs the expected moment parameter, variance moment parameter, skewness moment parameter, and kurtosis moment parameter into a preset distribution morphology mutation detection layer. It then calls the continuous time window moment parameter difference structure in the distribution morphology mutation detection layer to perform difference operations on various moment parameters between adjacent time windows, generating a mutation intensity sequence. For example, if the server identifies a sudden increase in the expected value of the retention distribution, a significant increase in variance, a shift in skewness towards longer retention periods, and an increase in kurtosis during continuous prediction time steps, it indicates a potential risk of abnormally prolonged retention times for a small number of goods in the future. The server integrates the changes in these multiple moment parameters into a mutation intensity value and arranges them sequentially along future time windows to form a mutation intensity sequence.

[0096] The server inputs the mutation intensity sequence into a preset mutation pattern recognition layer and calls the mutation template matching structure in the mutation pattern recognition layer to perform template matching processing between the mutation intensity sequence and a preset warehouse anomaly event mutation template library. The warehouse anomaly event mutation template library stores mutation pattern templates corresponding to anomalies such as concentrated congestion at the inbound warehouse, queuing at the outbound platform, cold chain environment anomalies, residual external transportation delays, and in-warehouse transfer conflicts. Based on the peak position, duration, rise rate, and fall rate of the mutation intensity sequence, the server identifies successfully matched anomalous mutation pattern segments and generates corresponding anomaly event type labels for these segments. For example, if the mutation intensity rapidly increases before the outbound peak and persists for multiple time steps, the server labels this segment as an outbound queuing anomaly; if the mutation intensity highly overlaps with the residual time of external pulse decay, the server labels this segment as an external transportation delay impact.

[0097] The server further inputs the abnormal mutation pattern fragments and corresponding abnormal event type labels into a preset abnormal source inference layer. The abnormal source inference layer calls a multi-source signal backtracking structure, and based on the time positioning of the abnormal mutation pattern fragments in a future time window, backtracks and retrieves abnormal fluctuation fragments of the environmental cycle phase feature trajectory, the cargo turnover trend slope feature trajectory, and the external event pulse decay feature trajectory within the corresponding time interval and the preceding influence interval. The server searches for abnormal advancement of temperature and humidity phases, frequent opening of warehouse doors, or abnormal refrigeration compensation cycles in the environmental channel; it searches for rapid increases in the amount of goods to be picked, slowed release of outbound tasks, or occupancy of channels by transfer tasks in the cargo trend channel; and it searches for the residual decay states of external events such as rainstorms, road congestion, and traffic control in the external pulse channel, generating a set of abnormal fluctuation fragments.

[0098] Finally, the server inputs the set of abnormal fluctuation segments and the abnormal event type labels into a preset attribution chain generation layer, and calls the causal chain reasoning structure in the attribution chain generation layer to construct a cross-channel attribution logic chain for the set of abnormal fluctuation segments. Based on the temporal sequence of the abnormal fluctuation segments, the causal lag relationship between channels, and the abnormal event type labels, the server generates an attribution chain pointing from the abnormal event type labels to the abnormal fluctuation segments of multi-source signals. For example, the server associates "outbound queuing anomaly" with the attribution chain of "enhanced residual impact of external road congestion - delayed vehicle arrival - concentrated release of outbound tasks - rapid increase in cargo trend slope - long-tailed distribution of stagnation." The server attaches the attribution chain to the stagnation distribution parameter set at the corresponding time step in the warehousing timeliness prediction result, enabling the warehousing timeliness prediction result to not only output future stagnation risks but also simultaneously explain the triggering source, propagation path, and affected objects of the abnormal risks, providing an interpretable basis for the warehousing management system to conduct anomaly warnings, scheduling interventions, and responsibility identification.

[0099] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned method for accurate prediction of warehouse delivery timeliness by fusing a large-scale time-series model with multi-source heterogeneous data. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0100] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method for accurately predicting warehouse delivery timeliness by integrating a large-scale time-series model with multi-source heterogeneous data, characterized in that, include: Acquire the environmental status time series reported by warehouse sensor acquisition devices, the cargo turnover event stream reported by cargo entry and exit scanning devices, and the external impact event records pushed by external event publishing platforms; According to the preset timestamp mapping relationship, the environmental collection timestamp in the environmental state time series, the cargo event timestamp in the cargo turnover event stream, and the external event timestamp in the external influence event record are mapped to time anchors on the warehousing time axis, and a hybrid data stream with a unified time anchor mark is generated based on the time anchors. The pre-built temporal decomposition network is invoked to decouple the hybrid data stream in time series, generating environmental periodic basis components, cargo trend basis components, and external impulse excitation components; The environmental periodic base component, the cargo trend base component, and the external pulse excitation component are projected onto the periodic phase space, the trend slope space, and the pulse intensity decay space, respectively, to generate the environmental periodic phase feature trajectory, the cargo turnover trend slope feature trajectory, and the external event pulse decay feature trajectory. The environmental cycle phase characteristic trajectory, the cargo turnover trend slope characteristic trajectory, and the external event pulse decay characteristic trajectory are input into a pre-constructed time series large model to perform cross-channel time dependency joint modeling, generating warehousing timeliness prediction results. The warehousing timeliness prediction results include the set of retention distribution parameters corresponding to each time step within the future time window.

2. The method according to claim 1, characterized in that, The step involves mapping environmental acquisition timestamps in the environmental state time series, cargo event timestamps in the cargo turnover event stream, and external event timestamps in the external impact event records to time anchors on the warehousing time axis based on a preset timestamp mapping relationship, and generating a hybrid data stream with unified timestamp markers based on these time anchors, including: The environmental data collection timestamp, the cargo event timestamp, and the external event timestamp are extracted respectively, and mapped to environmental time anchors, cargo time anchors, and external time anchors on the storage time axis using the preset timestamp mapping relationship. Using each time anchor point on the warehousing timeline as an alignment reference, the hit environmental state records, cargo turnover event records, or external impact event records are marked to the corresponding time anchor point; When the time anchor point corresponds to a data record from a single source, compensation and filling are performed based on the same type of data records corresponding to adjacent time anchor points. When the time anchor point corresponds to data records from two or three different sources, the data records are merged according to the time anchor point order on the storage time axis; Each data record corresponding to a time anchor point is marked as a hybrid record unit, and all the hybrid record units are arranged sequentially along the storage time axis to generate the hybrid data stream with uniform time anchor point markings.

3. The method according to claim 1, characterized in that, The invocation of a pre-built time-series decomposition network decouples the hybrid data stream in time series, generating environmental periodic basis components, cargo trend basis components, and external impulse excitation components, including: The hybrid data stream with unified time anchor marks is input into the pre-constructed time series decomposition network, and environmental state signal subsequence, cargo turnover signal subsequence and external event signal subsequence are extracted according to the environmental dimension, cargo dimension and external event dimension, respectively. The environmental state signal subsequence is analyzed by periodic component analysis to obtain the environmental periodic basis component; The trend component analysis is performed on the cargo turnover signal subsequence to obtain the cargo trend base component; The external event signal subsequence is analyzed by pulse component analysis to obtain the external pulse excitation component.

4. The method according to claim 1, characterized in that, The environmental periodic basis components are projected onto the periodic phase space using a pre-constructed periodic phase alignment network to generate environmental periodic phase feature trajectories, including: The environmental periodic basis components are input into the pre-constructed periodic phase alignment network, and autocorrelation analysis and periodic peak detection are performed on the environmental periodic basis components to obtain the environmental basis frequency. Based on the environmental base frequency, the environmental periodic base component is phase-mapped and phase-continuum-processed to obtain a continuous phase angle sequence. Phase state space encoding is performed on the continuous phase angle sequence to obtain the environmental periodic phase feature trajectory arranged in time order in the periodic phase space.

5. The method according to claim 1, characterized in that, The cargo trend base components are projected onto the trend slope space using a pre-constructed trend parsing network to generate a cargo turnover trend slope feature trajectory, including: The cargo trend base component is input into the pre-constructed trend parsing network, and the cargo trend base component is divided into trend windows to obtain multiple trend analysis window segments. For each of the aforementioned trend analysis window segments, a local trend slope estimate is performed to obtain the local trend slope parameter; The local trend slope parameter is quantized in direction and labeled in intensity to obtain trend direction category label and trend intensity label; Connect the trend direction category markers and trend intensity markers corresponding to adjacent trend analysis window segments in chronological order to generate the cargo turnover trend slope feature trajectory.

6. The method according to claim 1, characterized in that, The external pulse excitation component is projected onto the pulse intensity decay space using a pre-constructed impulse response analysis network to generate the external event pulse decay feature trajectory, including: The external pulse excitation component is input into the pre-constructed pulse response analysis network to locate the pulse start boundary of the external pulse excitation component, thereby obtaining the pulse start timestamp and pulse start amplitude parameters. Based on the pulse start timestamp, the external pulse excitation component is fitted with an attenuation envelope to obtain a pulse attenuation envelope sequence. The pulse decay rate parameter and pulse decay steady-state parameter are extracted from the pulse decay envelope sequence. The pulse initiation amplitude parameter, the pulse decay rate parameter, and the pulse decay steady-state parameter are mapped to the pulse intensity decay space, and the decay state is deduced along the time step direction to generate the pulse decay characteristic trajectory of the external event.

7. The method according to claim 1, characterized in that, The process involves inputting the environmental cycle phase characteristic trajectory, the cargo turnover trend slope characteristic trajectory, and the external event pulse decay characteristic trajectory into a pre-constructed time-series large model for joint modeling of cross-channel time dependencies, generating warehousing timeliness prediction results, including: Channel encoding is performed on the environmental cycle phase feature trajectory, the cargo turnover trend slope feature trajectory, and the external event pulse decay feature trajectory to obtain the environmental cycle phase encoding vector sequence, the cargo trend slope encoding vector sequence, and the external pulse decay encoding vector sequence, respectively. The environmental periodic phase encoding vector sequence, the cargo trend slope encoding vector sequence, and the external pulse attenuation encoding vector sequence are subjected to cross-channel time dependency interaction processing to obtain an interactive encoding vector sequence. The interactive coding vector sequence is subjected to temporal state evolution processing to obtain a temporal state vector sequence; The set of retention distribution parameters corresponding to each time step in the future time window is generated based on the time-series state vector sequence, and the set of retention distribution parameters is aggregated to obtain the storage timeliness prediction result.

8. The method according to claim 1, characterized in that, The method further includes: A cross-channel causal attribution structure is constructed within the pre-built temporal large model; The environmental cycle phase encoding vector sequence, the cargo trend slope encoding vector sequence, and the external pulse attenuation encoding vector sequence are input into the cross-channel causal attribution structure to perform Granger causality test processing between each pair of channels, and generate an inter-channel causality matrix. Based on the inter-channel causal relationship matrix, causal lag time estimation is performed on the channel pairs with causal relationships to generate an inter-channel causal lag time matrix. A causal constraint attention mask matrix is ​​generated based on the inter-channel causal relationship matrix and the inter-channel causal lag time matrix. The causal constraint attention mask matrix is ​​injected into the cross-channel time dependency interaction layer of the pre-built temporal large model to apply causal prior constraints to the attention weight calculation process in the multi-head cross-attention structure.

9. The method according to claim 1, characterized in that, The method further includes: The distribution pattern perturbation analysis is performed on the retention distribution parameter set in the storage timeliness prediction result. The retention distribution parameter set is input into the preset distribution pattern analysis layer. The multi-moment statistical structure in the distribution pattern analysis layer is called to extract the statistical moments from the first to the fourth moment of the retention distribution parameter set, generating the distribution expectation moment parameter, distribution variance moment parameter, distribution skewness moment parameter and distribution kurtosis moment parameter. The expected moment parameter, variance moment parameter, skewness moment parameter, and kurtosis moment parameter are input into a preset distribution morphology mutation detection layer. The continuous time window moment parameter difference structure in the distribution morphology mutation detection layer is called to perform difference operations on the expected moment parameter, variance moment parameter, skewness moment parameter, and kurtosis moment parameter between adjacent time windows to generate a mutation intensity sequence. The mutation intensity sequence is input into a preset mutation pattern recognition layer. The mutation template matching structure in the mutation pattern recognition layer is called to perform template matching processing between the mutation intensity sequence and the preset warehouse abnormal event mutation template library. The abnormal mutation pattern fragments that are successfully matched in the mutation intensity sequence and the corresponding abnormal event type labels are identified. The abnormal mutation pattern fragment and the corresponding abnormal event type are labeled and input into the preset abnormal source inference layer. The multi-source signal backtracking structure in the abnormal source inference layer is called to retrieve the abnormal fluctuation fragments of the environmental periodic phase feature trajectory, the cargo turnover trend slope feature trajectory and the external event pulse decay feature trajectory in the corresponding time interval according to the time positioning of the abnormal mutation pattern fragment, and generate an abnormal fluctuation fragment set. The abnormal fluctuation fragment set and the abnormal event type label are input into a preset source attribution chain generation layer. The causal chain reasoning structure in the source attribution chain generation layer is called to perform cross-channel attribution logic chain construction processing on the abnormal fluctuation fragment set, generating an attribution source chain pointing from the abnormal event type label to the multi-source signal abnormal fluctuation fragment. The attribution source chain is then attached to the retention distribution parameter set of the corresponding time step in the warehouse timeliness prediction result.

10. A warehouse timeliness prediction system that integrates a large-scale time-series model with multi-source heterogeneous data, characterized in that, Includes at least one service node; The service node includes a storage unit and a computing unit; the storage unit is used to store program code; the computing unit is used to run the program code to execute the warehouse timeliness accurate prediction method that integrates time-series large model and multi-source heterogeneous data as described in any one of claims 1 to 9.

Images