Intelligent supply chain management system and method based on artificial intelligence and big data

Abstract

The invention discloses an intelligent supply chain management system and method based on artificial intelligence and big data, and belongs to the technical field of supply chain management and artificial intelligence, and the method comprises the steps: obtaining a state data sequence of a supply chain object, extracting abnormal features, and forming an abnormal feature data sequence, obtaining a supply chain environment and operation parameter time sequence aligned in time and space; and jointly inputting the abnormal feature data sequence and the supply chain environment and operation parameter time sequence into a pre-trained multi-modal deep learning model for fusion analysis, and outputting one or more key supply chain parameters causing the abnormal state and quantized abnormal fluctuation information thereof, accurately associating the key parameters with the specific physical position or visual form of the abnormal state on the supply chain object, and finally generating an association map; according to the invention, full-link closed loop from data perception, intelligent analysis to root cause visualization is realized, and the intelligent level and fault processing efficiency of supply chain management are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of supply chain management and artificial intelligence, and in particular to a smart supply chain management system and method based on artificial intelligence and big data. BACKGROUND

[0002] With the increasing globalization and complexity of supply chains, abnormal events in each link of the supply chain, such as logistics delays, product damage, and parameter deviations, can easily trigger a chain reaction, leading to increased costs and decreased efficiency. Therefore, accurate tracing and rapid response to supply chain abnormalities have become a core requirement of the industry. In existing supply chain management methods, some rely on manual inspection or single sensor data monitoring, which has the problems of lagging abnormal detection and high missed detection rate. Although some technologies attempt to analyze abnormalities through data collection, they lack unique identity management of supply chain objects, leading to disconnection between data and physical objects. In addition, existing technologies mostly process state data or environmental parameters separately and do not achieve the synchronization of both in time and space, making it difficult to build a complete abnormal cause-and-effect chain. Furthermore, traditional models mostly use single feature extraction and analysis methods, which cannot effectively integrate multi-source heterogeneous data, leading to ambiguous abnormal root cause positioning, difficulty in identifying key influencing parameters and their fluctuation characteristics, and inability to accurately associate key parameters with specific locations and manifestations of abnormal occurrences, thus failing to provide accurate basis for subsequent optimization. Therefore, there is an urgent need for a smart supply chain management method that integrates unique identification, time-space synchronized data collection, and multi-modal intelligent analysis to address the technical deficiencies of existing technologies, such as difficulty in data tracing, inaccurate correlation, and ambiguous abnormal positioning, and to improve the intelligence and reliability of supply chain management. SUMMARY

[0003] To address the shortcomings of existing technologies, the present application provides a smart supply chain management system and method based on artificial intelligence and big data. The system acquires state data sequences of supply chain objects, extracts abnormal features, forms abnormal feature data sequences, and acquires time series of supply chain environmental and operational parameters that are aligned in time and space. The abnormal feature data sequences and the time series of supply chain environmental and operational parameters are jointly input into a pre-trained multi-modal deep learning model for fusion analysis, which outputs one or more key supply chain parameters that cause the abnormal state and their quantified abnormal fluctuation information. The key parameters are accurately associated with the specific physical location or visual form of the abnormal state on the supply chain object, and finally an association map is generated. The present application realizes a full-link closed loop from data perception, intelligent analysis to root cause visualization, and improves the intelligence level and fault handling efficiency of supply chain management.

[0004] To achieve the above-mentioned purposes, the present application provides the following technical solutions: A smart supply chain management method based on artificial intelligence and big data, comprising: Obtain the state data sequence of the supply chain object with a unique identity at the key node, and extract the abnormal features to form an abnormal feature data sequence; Obtain the supply chain environment and operation parameter time sequence synchronized with the state data sequence in space and time; Input the abnormal feature data sequence and the supply chain environment and operation parameter time sequence into a trained multi-modal deep learning model; Through the multi-modal deep learning model, output at least one key supply chain parameter causing the abnormal state and its abnormal fluctuation information, and accurately associate the key supply chain parameter with the specific location or manifestation of the abnormal state on the supply chain object.

[0005] Specifically, the state data sequence of the supply chain object with a unique identity at the key node includes: Define a supply chain object, and assign a globally unique digital identifier to each supply chain object; the supply chain object is an entity logistics unit, including raw materials, work-in-process, finished products or transportation containers; the digital identifier is bound to the RFID tag, two-dimensional code or digital twin of the physical object; Periodically collect physical state data of the supply chain object through an array of Internet of Things sensors deployed at the key node; the physical state data includes one or more of temperature, humidity, vibration amplitude, spatial coordinates, and packaging integrity image; the key node includes but is not limited to raw material storage platform, key process on production line, finished product warehouse stacking position or transportation tool compartment; Associate the timestamp-ordered physical state data stream with the digital identifier to form the state data sequence.

[0006] Specifically, the process of extracting abnormal features includes: Input the state data sequence with a unique identifier into a pre-trained time series transformation network for multi-scale feature extraction to obtain a time series feature map; Input the time series feature map into a pre-trained autoencoder, calculate the reconstruction error after passing through the autoencoder, identify the time series interval with reconstruction error exceeding the preset error threshold, and cut it out from the time series feature map to form a candidate abnormal feature segment; the encoder part of the autoencoder maps the time series feature map into a reference embedding space representing the normal mode, and the decoder part can reconstruct the time series feature map in the normal mode from the reference embedding space with high precision; Input the abnormal time series segment into a pre-trained conditional variational autoencoder, which learns and samples the abnormal mode in the latent space with the abnormal type label as the condition, and outputs an enhanced abnormal embedding vector; All enhanced anomalous feature fragments are sorted according to their original timestamps and integrated into the anomalous feature data sequence.

[0007] Specifically, the supply chain environment and operational parameter time series are obtained through the supply chain data bus; the supply chain data bus retrieves data from the event-driven supply chain data lake based on the timestamps and location information of the state data series; the data lake aggregates multi-source data from the manufacturing execution system, warehouse management system, transportation management system, and environmental sensors in real time.

[0008] Specifically, the trained multimodal deep learning model includes a cross-modal attention fusion module and a causal reasoning module; The cross-modal attention fusion module is used to receive the synchronized and aligned anomaly feature data sequence and the supply chain environment and operation parameter time series; through the cross-modal attention fusion module, it calculates the attention weight of each parameter in the supply chain environment and operation parameter time series to each anomaly feature in the anomaly feature data sequence; based on the attention weight, it generates a joint feature representation that integrates the interaction information of anomaly features and environmental operation parameters. The causal reasoning module is used to receive the joint feature representation output by the cross-modal attention fusion module; based on the joint feature representation, it uses a gradient-based causal discovery algorithm to identify a subset of parameters that have a potential causal relationship with the abnormal feature data sequence from the time series of the supply chain environment and operation parameters; and outputs the subset of parameters as candidate key supply chain parameters that lead to the abnormal state.

[0009] Specifically, the gradient-based causal discovery algorithm includes: constructing a structural equation model, using the joint feature representation output by the cross-modal attention fusion module as the endogenous variable of the structural equation model, and using all parameters in the time series of the supply chain environment and operational parameters as the exogenous variables of the structural equation model; using a gradient-based continuous optimization method to solve the sparsity constraint of the directed acyclic graph adjacency matrix in the structural equation model to identify the causal direction strength between exogenous and endogenous variables; and including parameters whose causal direction strength exceeds a preset significance threshold into the candidate key supply chain parameter subset.

[0010] Specifically, the output includes at least one key supply chain parameter and its abnormal fluctuation information that led to the abnormal state, including: The contribution metric analysis is performed on the candidate key supply chain parameters generated by the internal inference of the multimodal deep learning model, and the candidate key supply chain parameters with a contribution higher than the preset contribution threshold are screened out and preliminarily identified as key supply chain parameters. scanning and locating the most significant critical fluctuation interval before and after the abnormal state on the time series of the key supply chain parameters by using a time series saliency detection algorithm; quantitative analysis of the key fluctuation interval by combining the Bayesian change point detection method, calculating the fluctuation amplitude, fluctuation duration and deviation degree from the reference value of each key supply chain parameter in the key fluctuation interval, which together constitute the initial abnormal fluctuation information; Based on the initial abnormal fluctuation information, the adversarial samples of the key supply chain parameters in the abnormal state are generated by using the adversarial generative network, and are input into the multi-modal deep learning model. By comparing the output difference of the multi-modal deep learning model on the original data and the adversarial samples, the robustness of the causal relationship between the key supply chain parameters and the abnormal state is verified; the original data is the original abnormal feature data sequence and the supply chain environment and operation parameter time series; The verified key supply chain parameters and their final confirmed abnormal fluctuation information are integrated and output as the root cause analysis result.

[0011] Specifically, the accurate association of the key supply chain parameters and the specific location or form of the abnormal state on the supply chain object includes: Obtaining multi-modal image data of the surface or internal structure of the supply chain object when the abnormality occurs provided by a high-resolution scanning device or an image recognition subsystem; In the multi-modal deep learning model, an internal association mapping subnetwork is used to spatially align and time-stamp match the determined key supply chain parameters and their abnormal fluctuation information with the annotated physical defect locations or performance degradation manifestations in the multi-modal image data; Based on the matching results of the association mapping subnetwork, an association map is generated, which directly points each key supply chain parameter and its abnormal fluctuation information to the corresponding physical location or visualized abnormal manifestation form on the supply chain object in a visualized manner.

[0012] Specifically, the method further includes a root cause tracing, and the specific steps include: Based on the association map, the complete data history of the key supply chain parameters is traced back; Using a graph neural network to analyze the upstream event chain that causes the abnormal fluctuation of the key supply chain parameters, and locating the original operation instruction, environmental mutation event or device failure node that initially triggered the parameter fluctuation.

[0013] The intelligent supply chain management system based on artificial intelligence and big data includes a data acquisition module, a fusion and association module, and an abnormality analysis module. The data collection module is used for collecting state data of the supply chain object and performing preliminary processing. The fusion and association module is used for spatiotemporal alignment of the abnormal feature data and the context environment and the operation parameters, and construction of a complete analysis data set. The abnormal analysis module is used for interpretation, association and visualization of the output result of the multi-modal deep learning model.

[0014] Compared with the prior art, the beneficial effects of the present application are: The present application proposes a smart supply chain management method based on artificial intelligence and big data, which realizes accurate and automatic root cause positioning of supply chain abnormalities by fusing multi-source heterogeneous data and using a multi-modal deep learning model for deep analysis. The method can intelligently identify the key parameters and specific abnormal fluctuation patterns that cause abnormalities from complex supply chain environments and operation parameters, accurately associate abstract data abnormalities with specific physical locations or forms on the supply chain object, and thus break through the bottleneck of traditional methods that can only alarm but cannot locate.

[0015] The present application proposes a smart supply chain management method based on artificial intelligence and big data, which improves the intelligent level and decision-making efficiency of supply chain management. Through data-driven methods, abnormal diagnosis is transformed from experience-dependent passive response to accurate and visual active insight, which not only can quickly guide maintenance and intervention, but also can provide quantitative basis for process optimization, effectively reducing operating costs and business risks. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A schematic diagram of the present application based on artificial intelligence and big data smart supply chain management method is shown. Figure 2 A principle flowchart of the present application based on artificial intelligence and big data smart supply chain management method is shown. Figure 3 A system architecture diagram of the present application based on artificial intelligence and big data smart supply chain management system is shown. DETAILED DESCRIPTION

[0017] Example 1 Please refer to Figures 1-2 The present application provides an embodiment: a smart supply chain management method based on artificial intelligence and big data, which comprises S1-S4, including the following steps: S1: obtaining state data sequences of supply chain objects with unique identity at key nodes, and extracting abnormal features to form abnormal feature data sequences; S2: obtaining supply chain environment and operation parameter time series that are synchronized with the state data sequences in space and time; The supply chain environment and operation parameter time series are obtained through a supply chain data bus; the supply chain data bus retrieves from an event-driven supply chain data lake based on the time stamp and location information of the state data sequence; the data lake real-time converges multi-source data of a manufacturing execution system, a warehouse management system, a transportation management system and an environment sensor.

[0018] S3: inputting the abnormal feature data sequence and the supply chain environment and operation parameter time series into a trained multi-modal deep learning model; S4: outputting at least one key supply chain parameter causing the abnormal state and its abnormal fluctuation information through the multi-modal deep learning model, and accurately associating the key supply chain parameter with the specific location or manifestation of the abnormal state on the supply chain object.

[0019] The method further comprises a root cause tracing, and the specific steps comprise: Based on the association graph, the complete data history of the key supply chain parameter is traced back; Further, based on the accurate correspondence between the key supply chain parameter and the physical defect established by the association graph, the system starts deep backtracking analysis of the complete data history of the key supply chain parameter. Specifically, the backtracking process first reversely locates one or more key supply chain parameters that have strong causal association with the specific abnormal physical location and abnormal manifestation indicated in the association graph; according to the unique code of the key supply chain parameter and its logical level in the supply chain, all historical data records of the key supply chain parameter within a specified time window are automatically traced back and called, and the time window needs to completely cover the entire period from the normal state of the parameter to the occurrence of the abnormal evolution, thereby forming a parameter historical data set for deep analysis; after obtaining the parameter historical data set, the system performs multi-dimensional evolution trajectory analysis, which aims to reconstruct the whole process of each key parameter gradually deviating from the normal baseline state until finally causing the abnormality. Specifically, the system identifies whether there is a periodic fluctuation rule or a long-term trend change in the parameter historical data, and detects whether there is a small, continuous deviation or mutation point before the occurrence of the abnormality. At the same time, the analysis process associates different supply chain links and operation scenarios of the parameter, and evaluates the specific influence of external operation events or environmental changes on the parameter trajectory.

[0020] Further, the system will integrate and cross-compare the historical evolution trajectories of multiple key parameters, aiming to reveal the chain reaction and coupling effect between parameters, identify whether the initial abnormality of a parameter triggers the deviation of non-self-associated parameters, thus forming a cascading failure path leading to the final physical defect. Finally, all analysis results are integrated into a complete parameter life history report, which not only describes the historical behavior pattern of key parameters in detail, but also accurately locates the initial time point and initial cause of abnormality, and clearly outlines the complete link of abnormality propagation and amplification in the supply chain system.

[0021] Using graph neural networks to analyze the upstream event chain that leads to abnormal fluctuations in the key supply chain parameters, locate the original operation instruction, environmental mutation event, or device failure node that initially triggered the parameter fluctuations.

[0022] Further, using graph neural networks to analyze the upstream event chain that leads to abnormal fluctuations in the key supply chain parameters, locate the original operation instruction, environmental mutation event, or device failure node that initially triggered the parameter fluctuations, including: (1) Construct a supply chain knowledge graph as the basis for graph neural network analysis, where the supply chain knowledge graph is a global network constructed based on the existing supply chain digital twin model or historical data of the enterprise. The nodes in the supply chain knowledge graph are precisely defined as various entity types, including but not limited to specific operation instruction nodes, environmental state nodes, device unit nodes, and known key supply chain parameter nodes. At the same time, the edges in the supply chain knowledge graph represent the complex relationships between these entities, such as an edge representing an operation instruction triggering a device action, which in turn affects an environmental parameter, which ultimately affects the key supply chain parameter being analyzed. Therefore, this supply chain knowledge graph is a dynamic model that records the changes in the state of each node and the evolution of relationship strength in the time dimension; (2) Starting from the key supply chain parameter node that has occurred an abnormality, initiate a reverse, multi-hop traceability reasoning in the graph neural network. Adjacent nodes in the supply chain knowledge graph exchange feature information through neural networks: each upstream node transmits a signal to its downstream node, which integrates the state of the upstream node itself, the type and strength of its relationship with the downstream node, and the time sequence of the event. Through this layer-by-layer information transmission and aggregation, the graph neural network calculates a quantitative influence score for each causal path that may point to the abnormal parameter, thereby evaluating the contribution of all upstream nodes to the target abnormal state. (3) Based on the influence score calculated by the graph neural network, a few key supply chain parameter nodes with the highest influence score are selected from a large number of upstream nodes, which are determined as the root nodes causing abnormal fluctuations, and the root nodes are accurately analyzed and attributed: if the root node corresponds to an operation instruction, the complete content, time point of issuance, identity of the executor and deviation degree from the standard operating procedure of the instruction are traced back, if the root node corresponds to an environmental state, the mutation amplitude and duration thereof are quantified, and if the root node corresponds to equipment failure, the maintenance record, real-time sensor reading and performance degradation history thereof are associated; (4) Based on the attributed root nodes and their connection relationship in the supply chain knowledge graph, the system reconstructs a clear upstream event chain in time reverse order or causal order, which describes how the initial root node triggers a chain reaction through the relationship network defined in the supply chain knowledge graph and finally conducts to the target key supply chain parameter to cause its abnormality.

[0023] In summary, the present application allocates a globally unique digital identifier to a raw material, a finished product and the like physical logistics unit, periodically collects physical state data such as temperature, humidity and vibration amplitude through an array of Internet of Things sensors deployed at key nodes such as raw material storage, production process and transportation compartment, forms a state data sequence in association with the identifier according to the time stamp; extracts multi-scale features of the state data through a time series transformation network, identifies abnormal time series intervals with reconstruction error exceeding a threshold value using an autoencoder, and then enhances the abnormal pattern through a conditional variational autoencoder, forms an abnormal feature data sequence in time stamp order; based on the time stamp and location information of the state data, retrieves the time series of the supply chain environment and operation parameters that are spatiotemporally synchronized from the data lake that aggregates multi-system data through the supply chain data bus; inputs the two sequences into the trained model, first calculates the attention weights of the parameters and the abnormal features through the cross-modal attention fusion module to generate a joint feature representation; then filters out candidate key parameters that have a causal relationship with the abnormality through the causal reasoning module, quantifies the contribution, locates the fluctuation interval, and verifies the adversarial samples to determine the final key supply chain parameter and abnormal fluctuation information; combined with high-resolution image data, the specific location / performance of the key parameter and the abnormality such as physical defects and performance degradation are spatially and temporally aligned through an association mapping subnetwork to generate a visual association graph; finally, the upstream event chain of the key parameter is traced back by the graph neural network to locate the operation instruction, environmental mutation or equipment failure that initially caused the abnormality.

[0024] Exemplarily, the food enterprise implements the management method for the transportation link of cold-chain finished products, specifically including: assigning a unique digital identifier to each batch of quick-frozen seafood, binding an RFID tag, wherein the RFID tag records batch information, destination, etc.; setting the compartment of the transport refrigerated vehicle as a key node, deploying temperature sensors, vibration sensors, and high-definition cameras, collecting the temperature in the compartment every 5 minutes, collecting the vibration amplitude during transportation every 1 minute, and collecting the integrity image of the seafood package every 10 minutes, and associating the time stamp with the RFID tag to form a transportation state data sequence of the batch of seafood; inputting the state data sequence into a pre-trained time series transformation network to obtain multi-scale time series feature maps of temperature, vibration, and package image, reconstructing the feature maps through an autoencoder, finding that the temperature time series data reconstruction error before and after transportation to the midway station is far beyond the preset threshold, and intercepting the interval as a candidate abnormal feature segment, and inputting the candidate abnormal feature segment into a conditional variational autoencoder, wherein the conditional variational autoencoder takes temperature anomaly as a label, outputs an enhanced temperature anomaly embedding vector, and integrates the abnormal feature data sequence according to the time stamp; the supply chain data bus retrieves synchronous environment and operation parameters from the data lake based on the time stamp of the temperature anomaly, such as the 8th to 9th hour of transportation, and the location of the refrigerated vehicle, such as the vicinity of the midway station, including the operating power of the refrigeration equipment of the refrigerated vehicle, the number of times the vehicle door is opened and closed, the ambient temperature of the midway station, the operation record of the driver, etc., to form a supply chain environment and operation parameter time sequence; inputting the abnormal feature data sequence and the supply chain environment and operation parameter time sequence into a multi-modal deep learning model, calculating the attention weight of the refrigeration equipment operating power and the vehicle door opening and closing times on the temperature anomaly through the cross-modal attention fusion module to obtain the highest, generating a joint feature representation that fuses the interaction information of the two, and constructing a structural equation model through a causal reasoning module, taking the joint feature representation as an endogenous variable and all operation / environment parameters as exogenous variables, solving the adjacency matrix through gradient optimization, finding that the causal direction strength of the refrigeration equipment operating power far exceeds the threshold, and including the candidate key parameter, further quantitative analysis shows that the fluctuation amplitude of the candidate key parameter in the abnormal period reaches 30% of the normal benchmark value, and the duration is 45 minutes, and the power normal adversarial sample is generated through the adversarial generation network, the multi-modal deep learning model no longer outputs the temperature anomaly, verifies the robustness of the causal relationship between the two, and finally determines that the refrigeration equipment operating power anomaly is a key supply chain parameter; obtaining the package image during the abnormal period through the high-definition camera in the refrigerated vehicle, and labeling the specific position of the seafood package damage, such as the upper goods in the front compartment, through a high-resolution image recognition subsystem; the spatial alignment between the refrigeration equipment operating power anomaly and the damage position is performed through an association mapping subnetwork, such as the upper goods in the front compartment being closest to the refrigeration air outlet and the temperature fluctuation affecting the most directly, and the time stamp matching, such as the power anomaly period being consistent with the appearance time of the package damage, generating an association graph to show the corresponding relationship between the power anomaly and the front upper package damage;Finally, through the graph neural network, the running history data of the refrigeration equipment is traced back to find the upstream event chain of the power anomaly: when the intermediate station unloads, the driver does not turn off the power of the refrigeration equipment but keeps the container door open for a long time, causing the temperature in the container to rise sharply, the load of the refrigeration equipment to increase suddenly, and the filter element to be blocked due to long-term non-replacement, so that the normal operation power cannot be maintained, and finally the local temperature fluctuation is too large, causing the packaging of the seafood to be damaged.

[0025] The state data sequence of the supply chain object with a unique identity at a key node is obtained, including: A1: defining a supply chain object and assigning a globally unique digital identifier to each supply chain object; the supply chain object is an entity logistics unit, including raw materials, work-in-process, finished products or transportation containers; the digital identifier is bound to the RFID tag, two-dimensional code or digital twin of the physical object; A2: periodically collect physical state data of the supply chain object through an array of Internet of Things sensors deployed at key nodes; the physical state data includes one or more of temperature, humidity, vibration amplitude, spatial coordinates, and packaging integrity images; the key nodes include but are not limited to raw material storage platforms, key processes on production lines, finished product warehouse stacking positions, or transportation tool compartments; A3: associate the time-stamped physical state data stream with the digital identifier to form the state data sequence.

[0026] The process of extracting abnormal features includes: B1: input the state data sequence with a unique identifier into a pre-trained time series transformation network for multi-scale feature extraction to obtain a time series feature map; Further, the specific steps of B1 include: (1) According to the unique identity of the supply chain object, retrieve its complete state data sequence from the database, and pre-process the state data sequence to form a standardized state data sequence, the pre-processing includes using the interpolation method of before and after time data to fill in the missing values, and linearly transforming the value range of all data points in the sequence to map it to the interval of zero to one, wherein the interpolation method and linear transformation are prior art contents in the art and are not the inventive scheme of the present application, and will not be described here. (2) the normalized state data sequence is segmented in a multi-scale sliding window mechanism, wherein the multi-scale sliding window mechanism operates multiple windows with different time lengths in parallel, for example, a window focuses on short-term fluctuations and has a length of only a few hours to capture sudden abnormal signals; another window focuses on medium-term trends and has a length of several days to perceive slow performance degradation; a third window focuses on long-term cycles and has a length of several weeks to identify seasonal or periodic patterns; the windows slide on the time axis in an overlapping manner and respectively cut out multiple short sequence segments with different time granularities from the normalized state data sequence; (3) the short sequence segments from different scales are simultaneously input into a pre-trained time series transformation network for feature encoding, and after the encoding process, each short sequence segment is converted into a multi-dimensional feature vector; the core of the time series transformation network is composed of multiple layers of stacked time convolution modules and a self-attention mechanism, wherein the time convolution module serves as a local feature detector and contains multiple convolution kernels with different dilation rates, which can efficiently extract short-range, medium-range and long-range local dependencies and patterns in the sequence, such as sharp peaks, gentle slopes or periodic oscillations, within their respective receptive fields; the self-attention mechanism globally scans the entire short sequence segment and dynamically evaluates the importance of data points at different times to the overall semantics of the current segment by calculating the correlation weight between any two time points, thereby capturing remote dependencies even if they are far apart but have strong causal or logical correlations; (4) for each independent scale, the feature vectors generated by all sliding windows at this scale are arranged in the original time order to form a feature trajectory representing the evolution process at this scale, and then the feature trajectories from short-term, medium-term and long-term scales are organically combined through a feature fusion layer, such as channel splicing followed by one-dimensional convolution dimension reduction, and finally the fused feature set containing multi-scale information is restructured into a two-dimensional matrix, i.e. a time series feature map, where the rows represent the time dimension and the columns represent different feature channels.

[0027] B2: input the time series feature map into a pre-trained autoencoder, calculate the reconstruction error after passing through the autoencoder, identify the time intervals with reconstruction error exceeding the preset error threshold, and cut them out from the time series feature map to form candidate abnormal feature segments; the encoder part of the autoencoder maps the time series feature map into a reference embedding space representing normal patterns, and the decoder part can reconstruct the time series feature map under normal patterns from the reference embedding space with high precision; Further, the specific steps of B2 include: (1) input the time series feature map to be analyzed into a pre-trained autoencoder, the encoder of the autoencoder compresses and maps it to a reference embedding space representing normal patterns to generate an embedding vector, and the decoder receives the embedding vector and reconstructs a time series feature map with the same size as the input based on the understanding of the normal patterns by the embedding vector; (2) at each time point and each feature channel, point-by-point calculate the difference between the input time series feature map and the reconstructed time series feature map output by the decoder to obtain a reconstruction error, wherein the reconstruction error is the square of the difference in the corresponding point value; (3) along the time dimension, aggregate the reconstruction errors of all feature channels to obtain a reconstruction error curve varying with time, the reconstruction error curve reflects the difficulty of the autoencoder in understanding or reproducing the input data at each time point, and the peak of the reconstruction error curve indicates the time point where the abnormal pattern may exist; (4) compare the reconstruction error curve with an error threshold set in advance by a plurality of normal data, identify all time series intervals whose reconstruction errors continuously and significantly exceed the preset error threshold, and form a high error interval; (5) according to the time stamp corresponding to the identified high error interval, cut out the corresponding data segment from the input time series feature map, and each cut-out segment encapsulates an independent event of abnormal pattern occurring in a specific time period, and is named as a candidate abnormal feature segment.

[0028] B3: input the abnormal time series segment into a pre-trained conditional variational autoencoder, the conditional variational autoencoder learns and samples the abnormal pattern in the latent space conditioned on the abnormal type label, and outputs an enhanced abnormal embedding vector, wherein the conditional variational autoencoder is prior art in the art and is not the inventive scheme of the present application, and will not be described here; Further, the specific steps of B3 include: (1) use the labeled historical abnormal time series segment and its corresponding abnormal type label to pre-train a conditional variational autoencoder, in this process, the encoder learns to map each input abnormal time series segment to a probability distribution in the latent space, which is defined by a mean vector and a variance vector, at the same time, the conditional variational autoencoder injects the abnormal type label as condition information into the encoding and decoding process, so as to learn to orderly organize different categories of abnormal patterns in different regions of the latent space, after training, the conditional variational autoencoder has a structured latent space conditioned on the abnormal type; For example, when a time series segment labeled as a periodic oscillation anomaly is input, the label is converted into a condition vector, which is input into the encoder along with the data of the time series segment to guide the learning of the distribution characteristics specific to this type of anomaly. Similarly, during decoding, the condition vector is combined with the latent variable sampled from the distribution to input into the decoder, which is instructed to generate a data pattern consistent with the periodic oscillation anomaly. In this way, the conditional variational autoencoder learns to map different types of anomaly patterns in an ordered and non-interfering manner to different sub-regions of the latent space.

[0029] (2) input a new candidate abnormal time series segment into the pre-trained conditional variational autoencoder, and the encoder of the conditional variational autoencoder analyzes the candidate abnormal time series segment and calculates a corresponding probability distribution in the constructed conditional latent space for the candidate abnormal time series segment; (3) randomly sample multiple times from the obtained probability distribution to generate latent space vectors with subtle morphological differences, combine the latent space vectors with one or more most likely abnormal type condition vectors to form latent variable-condition pairs, and input the latent variable-condition pairs into the decoder, which plays the role of a generator based on the latent variable-condition pairs to reconstruct a new time series pattern; (4) map the new time series pattern generated by the decoder back to the latent space through the encoder to obtain a set of new and expanded embedding vectors, forming enhanced abnormal embedding vectors.

[0030] B5: sort all enhanced abnormal feature segments according to their original timestamps and integrate them into the abnormal feature data sequence.

[0031] The trained multi-modal deep learning model comprises a cross-modal attention fusion module and a causal reasoning module; The cross-modal attention fusion module is configured to receive the synchronized and aligned abnormal feature data sequence and the supply chain environment and operation parameter time series, calculate the attention weight of each parameter in the supply chain environment and operation parameter time series on each abnormal feature in the abnormal feature data sequence through the cross-modal attention fusion module, and generate a joint feature representation that integrates the interaction information between abnormal features and environment operation parameters based on the attention weight. Further, the joint feature representation is realized by weighted summation of the interaction information between abnormal features and environment operation parameters, wherein the weight is the attention weight.

[0032] Further, the cross-modal attention fusion module is configured to calculate the attention weight of each parameter in the supply chain environment and operation parameter time series on each abnormal feature in the abnormal feature data sequence, comprising: (1) The system defines the supply chain environment and operation parameter time series as a parameter mode, and defines the abnormal feature data sequence as an abnormal mode. The system uses two independent neural networks to convert the parameter mode into a parameter feature vector sequence and the abnormal mode into an abnormal feature vector sequence, respectively, to ensure that the parameter feature vector sequence and the abnormal feature vector sequence are strictly synchronized in the time dimension. The neural network is a prior art in the field and is not the inventive scheme of the present application, and will not be described here. (2) The system takes each vector in the abnormal feature vector sequence as a query vector and all vectors in the parameter feature vector sequence as key vectors. For each query vector, the system calculates the correlation score with all key vectors to evaluate the correlation between the parameter state at different times and the current abnormal feature. The correlation score is calculated by dot product, which measures the correlation between the abnormal state at each time and the parameter state before, at the time, and after the abnormal state in terms of numerical pattern and trend. The correlation score uses the Pearson correlation coefficient, and the calculation formula of the Pearson correlation coefficient is a prior art in the field and is not the inventive scheme of the present application, and will not be described here. (3) The original correlation score is normalized by the Softmax function to convert it into a probability distribution with a sum of 1. Each value obtained after normalization is the attention weight, which accurately quantifies the importance of each time point in the parameter sequence to the explanation of the current abnormality. The Softmax function is a prior art in the field and is not the inventive scheme of the present application, and will not be described here.

[0033] The causal reasoning module is configured to receive the joint feature representation output by the cross-modal attention fusion module, identify a parameter subset having a potential causal relationship with the abnormal feature data sequence from the supply chain environment and operation parameter time series based on the joint feature representation using a gradient-based causal discovery algorithm, and output the parameter subset as a candidate key supply chain parameter causing the abnormal state.

[0034] The gradient-based causal discovery algorithm includes: C1: Construct a structural equation model, taking the joint feature representation output by the cross-modal attention fusion module as an endogenous variable of the structural equation model, and taking all parameters in the supply chain environment and operation parameter time series as exogenous variables of the structural equation model. Further, the specific steps of C1 include: (1) Define the joint feature representation output by the cross-modal attention fusion module as an endogenous variable, and define each independent parameter in the supply chain environment and operation parameter time series as an exogenous variable. (2) Based on the variable definition, the system constructs a theoretical causal network diagram, which assumes that all exogenous variables and their interactions affect and explain the changes of endogenous variables, and the arrows in the diagram directly represent the causal paths of these assumptions; It should be noted that in the causal network diagram, each exogenous parameter is taken as a starting point, and one or more arrows are sent directly to the end point, i.e., the joint feature representation.

[0035] (3) The constructed causal network diagram, actual joint feature representation and exogenous variable data are input together, and a partial least squares method is used to fit the structural equation model, and through iterative optimization, a specific numerical weight, i.e., a path coefficient, is assigned to each path in the causal network diagram, thereby quantifying the direct and indirect influence strength of each exogenous variable on the endogenous variable, wherein the partial least squares method is a prior art content in the art and is not the inventive scheme of the present application, and will not be described here; (4) After the path coefficient calculation is completed, a goodness-of-fit test is performed to evaluate the goodness of fit of the structural equation model and the data, and after verification, the system analyzes the size and significance of all path coefficients, and by comparing the path coefficients, the system selects the key exogenous variables that have the most significant causal driving force on the endogenous variables from all exogenous variables, wherein the goodness-of-fit test is achieved by calculating the chi-square value, and the chi-square value calculation method is a prior art content in the art and is not the inventive scheme of the present application, and will not be described here; (5) The key exogenous variables and their associated path coefficients selected are taken as the core of the structural equation model and output, and thus the structural equation model is obtained.

[0036] C2: A gradient-based continuous optimization method is used to solve the sparse constraint of the directed acyclic graph adjacency matrix in the structural equation model to identify the causal direction strength between the exogenous variables and the endogenous variables. Further, the specific steps of C2 include: (1) A weighted adjacency matrix is constructed to represent the possible causal connections between all exogenous variables and endogenous variables, and in the weighted adjacency matrix, each directed edge is no longer represented by 0 or 1, but is assigned an initial weight value, thereby converting the discrete graph structure problem into a differentiable continuous optimization problem, and the initial weight value is a non-zero random number at initialization, representing an initial assumption of the causal direction strength of the edge; (2) define a composite loss function as the guiding target of the optimization process, which is composed of two parts: the first part is the data reconstruction loss, which measures the difference between the endogenous variable value predicted by the structural equation model based on the current weighted adjacency matrix representing the network structure and the actual observed joint feature representation, usually measured by mean square error; the second part is the sparsity constraint loss, which is represented by L1 norm regularization, which is to punish the sum of the absolute values of all non-zero weight values in the weighted adjacency matrix to drive the unimportant connection weight to zero, wherein the mean square error and the L1 norm regularization are prior art in the art and are not the creative scheme of the present application, and will not be described here; (3) the system enters an iterative optimization cycle: in each iteration, the gradient of the composite loss function with respect to each initial weight value in the weighted adjacency matrix is calculated by back propagation, and the optimizer is used to update all weight values in the adjacency matrix according to the calculated gradient. Through multiple iterations, the data reconstruction loss and the sparsity constraint loss are minimized, and the weight values in the weighted adjacency matrix are stabilized, wherein back propagation and optimizer are prior art in the art and are not the creative scheme of the present application, and will not be described here; (4) after optimization, the system interprets the final stable weighted adjacency matrix, most of the weight values of the edges in the stable weighted adjacency matrix have been compressed to near zero, and only a few edges remain with significant non-zero weights; (5) set a minimum threshold for binary processing, retain these significant non-zero weights, and confirm their absolute values as causal direction strength, thereby outputting a sparse directed acyclic graph with causal strength information, wherein binary processing is prior art in the art and is not the creative scheme of the present application, and will not be described here; (6) according to the generated directed acyclic graph, the key exogenous variables with the strongest driving force on the exogenous variables and their causal direction strength are revealed, and the causal discovery is completed.

[0037] C3: include the parameters with causal direction strength exceeding the preset significance threshold in the candidate key supply chain parameter subset, and the preset significance threshold is set to 0.05 in the present application.

[0038] The output at least one key supply chain parameter causing the abnormal state and its abnormal fluctuation information includes: D1: perform contribution quantification analysis on the candidate key supply chain parameters generated by the internal reasoning of the multi-modal deep learning model, and select the candidate key supply chain parameters with contribution higher than the preset contribution threshold, which are preliminarily confirmed as key supply chain parameters; Further, the specific steps of D1 include: (1) Extract the intermediate data generated in the internal reasoning process of the multi-modal deep learning model as the original evidence for contribution analysis, which mainly includes the attention weight distribution of the multi-modal deep learning model to each candidate parameter, the gradient information of the multi-modal deep learning model output relative to the input of each candidate parameter, and the activation value of the key network layer; (2) Based on the original evidence extracted in the first step, the system calculates the contribution score in multiple dimensions in parallel, including: calculating the attention contribution score based on the attention weight distribution, calculating the gradient contribution score based on the gradient information, and calculating the replacement contribution score by parameter replacement. These scores reflect the influence degree of each candidate parameter on model decision-making from different angles; Further, the contribution score based on attention weight is calculated by statistics of the weight mean and variance obtained by each candidate parameter on all attention heads; the contribution score based on gradient is calculated by calculating the gradient size of the multi-modal deep learning model output relative to the input of each candidate parameter, and the larger the gradient, the more critical the parameter is to the output result; the contribution score based on parameter replacement, the system temporarily replaces the time series of any candidate parameter with random noise or baseline value, then re-runs the multi-modal deep learning model and observes the degree of performance decline of the multi-modal deep learning model output, the more serious the performance decline, the more indispensable the parameter is.

[0039] (3) Normalize the attention contribution score, the gradient contribution score and the replacement contribution score respectively, and according to the pre-defined weight strategy, weight and sum the normalized contribution score to calculate a comprehensive contribution total score for each candidate parameter; Further, the pre-defined weight strategy is: assign a weight to each index, and weight and fuse these normalized scores to calculate a final and comprehensive contribution total score for each candidate key supply chain parameter.

[0040] (4) According to the statistical distribution of the comprehensive contribution total score of all candidate parameters, dynamically calculate a contribution threshold, then compare the comprehensive contribution total score of each candidate parameter with the contribution threshold, and automatically select all parameters whose scores are higher than the threshold, wherein the contribution threshold is dynamically determined based on the statistical distribution of the contribution total score of all candidate parameters, for example, set the contribution threshold to the median or upper quartile of the total score to ensure that the selected parameters are truly prominent; (5) Preliminarily confirm the selected parameters as key supply chain parameters, and generate a contribution analysis report, which records the comprehensive contribution total score of each key supply chain parameter and its score in each dimension, the contribution threshold used, and the elimination reason of the unselected parameters.

[0041] D2: Scanning and locating the most significant critical fluctuation interval before and after the abnormal state occurs on the time series of the key supply chain parameters using a time series significance detection algorithm; Further, the specific steps of D2 include: (1) Taking the time point of the abnormal state as the benchmark, expand the analysis window forward and backward from the complete time series data of the key supply chain parameters that have been preliminarily confirmed to ensure complete coverage of the cause accumulation, concentrated outbreak and continuous impact period of the anomaly; (2) Within the designated analysis window, use the probe to slide at a fixed step, at each sliding stop position, calculate the statistical distance between the internal data segment and the baseline data model representing the normal state, such as the dynamic time warping distance, and the statistical distance value is the initial significance score of the window position, thereby generating an initial significance score curve that changes over time, wherein the fixed step is one data point per move, and dynamic time warping is a prior art content in the field and is not the inventive scheme of the present application, which will not be described here; (3) Smooth the initial significance score curve to suppress noise and identify all local extreme points on the smoothed curve, especially the peak points, and mark the local extreme points as candidate key points, wherein the smoothing filter is a prior art content in the field and is not the inventive scheme of the present application, which will not be described here; (4) Set an adaptive threshold to filter out candidate key points whose peak values exceed the adaptive threshold, and expand to both sides from each filtered candidate key point as the center until the score drops to the peak value proportion threshold, forming a temporary critical fluctuation interval, and in the expansion process, if two adjacent temporary intervals overlap or are very close in time, it is determined whether they belong to the same fluctuation event, and they are merged into a unified critical fluctuation interval; the adaptive threshold is determined based on the statistical characteristics of the overall significance score, and in the present application, the peak value proportion threshold is set to [0.2, 0.5]; (5) For all candidate critical fluctuation intervals, rank them according to their peak significance score, duration and overall deviation, and confirm the top one or a few intervals as the critical fluctuation interval.

[0042] D3: Combine the Bayesian change point detection method to quantitatively analyze the key fluctuation interval, calculate the fluctuation amplitude, fluctuation duration and deviation from the baseline value of each key supply chain parameter in the key fluctuation interval, and jointly constitute the initial abnormal fluctuation information; Further, the specific steps of D3 include: (1) Taking the time stamp of the key fluctuation interval located by the time sequence saliency detection algorithm as the boundary, the data of the interval is cut out from the complete sequence of the key supply chain parameter, at the same time, a stable historical data before the abnormality occurs and far away from the interval is selected, the statistical characteristics are calculated, and the normal baseline value is defined; (2) The key fluctuation interval data cut out is input into the Bayesian change point model, the Bayesian change point model iteratively reasons through the Markov chain Monte Carlo sampling algorithm, and outputs the posterior probability of each time point becoming a change point, so as to identify the significant change points in which the statistical characteristics in the fluctuation interval suddenly change, such as the starting point, peak point and ending point of the fluctuation, wherein the Bayesian change point model and the Markov chain Monte Carlo sampling algorithm are prior art contents in the field, and are not the inventive scheme of the present application, and will not be described here; (3) Based on the identified significant change points and the defined normal baseline value, the system performs quantitative calculation of three core indexes: The first is the fluctuation amplitude: calculating the maximum absolute deviation of the parameter value in the key fluctuation interval relative to the normal baseline value; The second is the fluctuation duration: calculating the time difference between the first and last significant change points; The third is the deviation degree: calculating the overall measure of the difference between all data points in the key fluctuation interval and the normal baseline value; (4) The three quantitative indexes of fluctuation amplitude, fluctuation duration and deviation degree calculated are integrated and packaged to jointly constitute the initial abnormal fluctuation information of the key supply chain parameter.

[0043] D4: Based on the initial abnormal fluctuation information, an adversarial sample of the key supply chain parameter in an abnormal state is generated by using an adversarial generation network, and is input into the multi-modal deep learning model, and the robustness of the causal relationship between the key supply chain parameter and the abnormal state is verified by comparing the output difference of the multi-modal deep learning model on the original data and the adversarial sample; the original data is the original abnormal feature data sequence and the supply chain environment and operation parameter time sequence; Further, the specific steps of D4 include: (1) Taking the fluctuation mode described by the initial abnormal fluctuation information as the condition, driving the adversarial generation network to generate a small and targeted adversarial disturbance, and superimposing the adversarial disturbance on the corresponding key supply chain parameter segment of the original supply chain environment and operation parameter time sequence to create an adversarial sample, wherein the adversarial generation network is a prior art content in the field, and is not the inventive scheme of the present application, and will not be described here; (2) input the original data and the adversarial sample into the trained multi-modal deep learning model simultaneously for inference, and obtain the original output and the adversarial output of the model on the two groups of data respectively; the output content includes the abnormal type considered by the model, the confidence, the list of key supply chain parameters identified and the contribution degree, wherein the adversarial sample is the original abnormal feature data sequence and the supply chain environment and operation parameter time sequence that has been modified; (3) compare the difference between the original output and the adversarial output, and analyze the core focus: whether the contribution degree of the key supply chain parameter disturbed remains stable or improves, and whether the determination confidence of the multi-modal deep learning model on the abnormal type fluctuates significantly; If the adversarial output and the original output are highly consistent in the contribution degree of the key parameter and the abnormal type, it is proved that the causal relationship between the key supply chain parameter and the abnormal state is robust; If there is a significant deviation, it indicates that the initial association judgment is fragile and needs to be reevaluated.

[0044] D5: integrate the verified key supply chain parameters and the finally confirmed abnormal fluctuation information and output them as the results of root cause analysis.

[0045] The accurate association of the key supply chain parameter and the abnormal state on the specific position or manifestation of the supply chain object includes: E1: obtain the multi-modal image data of the surface or internal structure of the supply chain object when the abnormality occurs, which is provided by a high-resolution scanning device or an image recognition subsystem; E2: in the multi-modal deep learning model, an internal association mapping subnetwork is used to spatially align and time-stamp match the determined key supply chain parameters and abnormal fluctuation information with the annotated physical defect positions or performance degradation manifestations in the multi-modal image data; Further, the specific steps of E2 include: (1) the association mapping subnetwork receives two input modalities: one is the key supply chain parameter and its abnormal fluctuation information, which is encoded into a parameter feature vector through a fully connected network; the other is the annotated physical defect position or performance degradation manifestation in the multi-modal image data, which is encoded into a visual feature vector through a convolutional neural network; (2) based on the abnormal time window contained in the parameter feature vector and the time stamp metadata corresponding to the visual feature vector, the cross-attention weight of the timeline of the two is calculated, and the most synchronized parameter-image data pair in time is found out, and the time sequence alignment is completed, wherein the cross-attention is a prior art content in the field and is not the inventive scheme of the present application, and will not be described here; (3) input the time-aligned parameter feature vector as a query vector, the time-aligned visual feature vector as a key vector and a value vector into the cross-modal attention layer for calculation, and generate a spatial attention weight graph by calculating the similarity between the query vector and all key vectors, wherein the spatial attention weight graph indicates which spatial area of the visual feature the parameter feature is most active, and the similarity calculation process is prior art content in the art and is not the inventive scheme of the present application, and is not described here; (4) decode the spatial attention weight graph, integrate it with the timestamp, parameter and abnormal information, and generate a set of structured association tuples, each of which records a key supply chain parameter, its abnormal fluctuation information, timestamp and corresponding physical defect coordinates or performance degradation area.

[0046] E3: based on the matching result of the association mapping subnetwork, generate an association graph that directly points each key supply chain parameter and its abnormal fluctuation information to the corresponding physical location or visual abnormal manifestation of the supply chain object in a visual manner.

[0047] Further, the specific steps of E3 include: (1) based on the association tuples output by the association mapping subnetwork, create two types of nodes: key supply chain parameter nodes representing data anomalies and physical defect nodes representing the physical world, wherein the key supply chain parameter nodes store parameter name, fluctuation amplitude, etc. information, and the physical defect nodes store type, spatial coordinates, etc. information; (2) according to the mapping relationship in the association tuple, create a directed connection edge between the key supply chain parameter node and the physical defect node, with the key supply chain parameter node as the starting point of the edge and the physical defect node as the end point of the edge; (3) take the three-dimensional design drawing of the supply chain object as a spatial base map, and accurately anchor all physical defect nodes to the actual position on the spatial base map according to their internal coordinates, and at the same time, layout the key supply chain parameter nodes on the side of the base map, and according to the weight information of the connection edge, with different thicknesses, including: 1) load the three-dimensional design drawing of the supply chain object as a spatial base map, and establish a unified coordinate system, calibrate the coordinates stored in the physical defect nodes with the spatial base map, and at the same time, demarcate a logical and clear parameter panel area on the side of the spatial base map; 2) traverse all physical defect nodes, accurately anchor them to the calibrated spatial base map according to their coordinate data, and use differentiated icons to identify their defect types, and at the same time, orderly arrange all key supply chain parameter nodes in the parameter panel area, and can be grouped according to their system category or contribution; 3) Read all the data of the connected edges, map them to different visual widths and colors according to the associated weight strength encapsulated, the higher the weight, the thicker the edge and the more eye-catching the color, then the system draws these directed arrows with specific visual styles from each key supply chain parameter node in the parameter panel to all the physical defect nodes associated with it on the spatial base map; 4) Combine the spatial base map, the physical defect nodes with completed layout, the key supply chain parameter nodes and the drawn directed connection edges into a complete, non-interactive two-dimensional or three-dimensional image file, i.e. static correlation atlas.

[0048] (4) Inject interactive logic into the static correlation atlas to form an interactive visualization interface, in which the user clicks on any key supply chain parameter node to highlight all the connection edges and connected physical defect nodes derived from it, and dynamically display the detailed abnormal information of the parameter; conversely, clicking on any physical defect node will highlight all the connection edges and key supply chain parameter nodes pointing to it, and display the associated details.

[0049] Example 2: Please refer to Figure 3 The present application provides another embodiment: a smart supply chain management system based on artificial intelligence and big data, comprising: a data acquisition module, a fusion and correlation module, and an abnormality analysis module; The data acquisition module is used to collect the state data of the supply chain objects and perform preliminary processing; The fusion and correlation module is used to perform spatio-temporal alignment of the abnormal feature data with the context environment and operation parameters, and construct a complete analysis dataset; The abnormality analysis module is used to interpret, correlate and visualize the output results of the multi-modal deep learning model.

[0050] The data acquisition module includes a unique identifier management unit, a sensor acquisition unit and an abnormal feature extraction unit. The unique identifier management unit is used to assign a globally unique digital identifier to each supply chain object, and bind it with the RFID tag, two-dimensional code or digital twin of the physical object, to ensure accurate correspondence and full traceability of physical flow and information flow; The sensor acquisition unit is used to periodically or event-drivenly acquire the physical state data of the supply chain objects through the sensor array deployed at the key nodes; The abnormal feature extraction unit is used to automatically identify and extract features representing abnormal patterns from the state data sequence, and combine them into an abnormal feature data sequence.

[0051] The fusion and association module comprises a data access unit, a supply chain data lake unit, and a space-time synchronization retrieval unit. The data access unit receives environmental and operating parameter data from business systems such as manufacturing execution systems, warehouse management systems, and transportation management systems by constructing a data bus or middleware. The supply chain data lake unit provides a centralized data storage and management platform, and in an event-driven manner, real-time gathers and stores all multi-source data from IoT sensors and business systems. The space-time synchronization retrieval unit is used to extract environmental parameters and operating parameters that act on the same supply chain object within the same time period from the data lake by taking the timestamp and location information in the state data sequence as the retrieval condition, and forms a supply chain environmental and operating parameter time sequence.

[0052] The anomaly analysis module comprises a parameter confirmation unit, an abnormal fluctuation quantification unit, a space-time correlation mapping unit, and a result pushing unit. The parameter confirmation unit is used to analyze and sort the contribution degree of the candidate key parameters output by the multi-modal deep learning model, and filter out the parameters with the highest contribution degree to the abnormal occurrence as the final key supply chain parameters. The abnormal fluctuation quantification unit is used to format and quantitatively describe the abnormal fluctuation information of the key parameters, such as outputting the temperature fluctuation amplitude as +5℃, the duration as 30 minutes, and the deviation from the reference value as 20%. The space-time correlation mapping unit is used to accurately correlate and map the confirmed key supply chain parameters with the specific physical location and specific manifestation of the abnormal state on the supply chain object. The result pushing unit is used to present the final analysis results, including the abnormal object ID, location, manifestation, reason parameter, and fluctuation details to the user in an intuitive form, and triggers a work order system or an automatic control instruction.

[0053] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative and not limiting. Those skilled in the art can make changes, modifications, replacements, and variations to the above embodiments without departing from the purpose and scope of the present application, and these are all within the protection scope of the present application.

Claims

1. A smart supply chain management method based on artificial intelligence and big data, characterized in that: include: Obtain the status data sequence of supply chain objects with unique identifiers at key nodes, and extract abnormal features to form an abnormal feature data sequence; Acquire a time series of supply chain environment and operational parameters that are spatiotemporally synchronized with the state data sequence; The abnormal feature data sequence and the supply chain environment and operation parameter time series are input into a trained multimodal deep learning model; The multimodal deep learning model outputs at least one key supply chain parameter that causes the abnormal state and its abnormal fluctuation information, and accurately associates the key supply chain parameter with the specific location or manifestation of the abnormal state on the supply chain object.

2. The intelligent supply chain management method based on artificial intelligence and big data as described in claim 1, characterized in that, The acquisition of the state data sequence of supply chain objects with unique identifiers at key nodes includes: Define supply chain objects and assign a globally unique digital identifier to each supply chain object; the supply chain object is a physical logistics unit, including raw materials, work-in-process, finished products or transport containers; the digital identifier is bound to the RFID tag, QR code or digital twin of the physical object; The physical status data of the supply chain objects is periodically collected by an IoT sensor array deployed at key nodes. The physical status data includes one or more of the following: temperature, humidity, vibration amplitude, spatial coordinates, and packaging integrity images. The key nodes include, but are not limited to, raw material receiving platforms, key processes on the production line, finished product warehouse stacks, or transport vehicle bodies. The physical state data stream, sorted by timestamp, is associated with the digital identifier to form the state data sequence.

3. The intelligent supply chain management method based on artificial intelligence and big data as described in claim 2, characterized in that, The process of extracting abnormal features includes: The state data sequence with unique identifiers is input into a pre-trained temporal transformation network for multi-scale feature extraction to obtain a temporal feature map; The temporal feature map is input into a pre-trained autoencoder, the reconstruction error after passing through the autoencoder is calculated, the temporal intervals where the reconstruction error exceeds a preset error threshold are identified, and these intervals are extracted from the temporal feature map to form candidate abnormal feature segments; the encoder part of the autoencoder maps the temporal feature map into a reference embedding space representing the normal mode, and its decoder part can reconstruct the temporal feature map in the normal mode with high precision from the reference embedding space; The abnormal time sequence segment is input into a pre-trained conditional variational autoencoder. The conditional variational autoencoder learns and samples the abnormal pattern in the latent space with the abnormal type label as a condition, and outputs an enhanced abnormal embedding vector. All enhanced anomalous feature fragments are sorted according to their original timestamps and integrated into the anomalous feature data sequence.

4. The intelligent supply chain management method based on artificial intelligence and big data as described in claim 3, characterized in that, The supply chain environment and operational parameter time series are obtained through the supply chain data bus; the supply chain data bus retrieves data from the event-driven supply chain data lake based on the timestamp and location information of the state data sequence; the data lake aggregates multi-source data from the manufacturing execution system, warehouse management system, transportation management system, and environmental sensors in real time.

5. The intelligent supply chain management method based on artificial intelligence and big data as described in claim 4, characterized in that, The trained multimodal deep learning model includes a cross-modal attention fusion module and a causal reasoning module; The cross-modal attention fusion module is used to receive the synchronized and aligned anomaly feature data sequence and the supply chain environment and operation parameter time series; and to calculate the attention weight of each parameter in the supply chain environment and operation parameter time series to each anomaly feature in the anomaly feature data sequence through the cross-modal attention fusion module. Based on the attention weights, a joint feature representation that integrates abnormal features and environmental operation parameter interaction information is generated; The causal reasoning module is used to receive the joint feature representation output by the cross-modal attention fusion module; based on the joint feature representation, it uses a gradient-based causal discovery algorithm to identify a subset of parameters that have a potential causal relationship with the abnormal feature data sequence from the time series of the supply chain environment and operation parameters; and outputs the subset of parameters as candidate key supply chain parameters that lead to the abnormal state.

6. The intelligent supply chain management method based on artificial intelligence and big data as described in claim 5, characterized in that, The gradient-based causal discovery algorithm includes: constructing a structural equation model, using the joint feature representation output by the cross-modal attention fusion module as the endogenous variable of the structural equation model, and using all parameters in the time series of the supply chain environment and operational parameters as the exogenous variable of the structural equation model; using a gradient-based continuous optimization method to solve the sparsity constraint of the directed acyclic graph adjacency matrix in the structural equation model to identify the causal direction strength between exogenous and endogenous variables; and including parameters whose causal direction strength exceeds a preset significance threshold into the subset of candidate key supply chain parameters.

7. The intelligent supply chain management method based on artificial intelligence and big data as described in claim 6, characterized in that, The output includes at least one key supply chain parameter and its abnormal fluctuation information that led to this abnormal state, including: The contribution metric analysis is performed on the candidate key supply chain parameters generated by the internal inference of the multimodal deep learning model, and the candidate key supply chain parameters with a contribution higher than the preset contribution threshold are screened out and preliminarily identified as key supply chain parameters. The time-series significance detection algorithm is used to scan and locate the most significant key fluctuation intervals before and after the occurrence of abnormal states on the time series of the key supply chain parameters; By combining the Bayesian change point detection method, the key fluctuation range is quantitatively analyzed, and the fluctuation amplitude, fluctuation duration and deviation from the benchmark value of each key supply chain parameter in the key fluctuation range are calculated, which together constitute its initial abnormal fluctuation information. Based on the initial abnormal fluctuation information, an adversarial generative network is used to generate adversarial examples of the key supply chain parameters under abnormal conditions, and these examples are then input into the multimodal deep learning model. By comparing the output differences of the multimodal deep learning model on the original data and the adversarial examples, the robustness of the causal relationship between the key supply chain parameters and the abnormal conditions is verified. The original data consists of the original abnormal feature data sequence and the time series of supply chain environment and operational parameters. The verified key supply chain parameters and their finally confirmed abnormal fluctuation information will be integrated and output as the results of the root cause analysis.

8. The intelligent supply chain management method based on artificial intelligence and big data as described in claim 7, characterized in that, The precise association between key supply chain parameters and the specific location or manifestation of abnormal states on the supply chain object includes: Acquire multimodal image data of the surface or internal structure of the supply chain object when an anomaly occurs, provided by a high-resolution scanning device or image recognition subsystem; In the multimodal deep learning model, a built-in association mapping subnetwork is used to spatially align and time-stamp match the identified key supply chain parameters and their abnormal fluctuation information with the marked physical defect locations or performance degradation manifestations in the multimodal image data. Based on the matching results of the associated mapping subnetwork, an association graph is generated. The association graph visualizes each key supply chain parameter and its abnormal fluctuation information, directly pointing to the corresponding physical location or visualized abnormal behavior on the supply chain object.

9. The intelligent supply chain management method based on artificial intelligence and big data as described in claim 8, characterized in that, The method also includes a root cause analysis, the specific steps of which include: Based on the aforementioned correlation graph, the complete data history of the key supply chain parameters is traced back. By using graph neural networks to analyze the upstream event chain that caused the abnormal fluctuations in the key supply chain parameters, the original operating instructions, environmental mutation events, or equipment failure nodes that initially triggered the parameter fluctuations can be located.

10. A smart supply chain management system based on artificial intelligence and big data, used to implement the smart supply chain management method based on artificial intelligence and big data as described in any one of claims 1-9, characterized in that, include: Data acquisition module, fusion and correlation module, anomaly analysis module; The data acquisition module is used to collect status data of supply chain objects and perform preliminary processing; The fusion and association module is used to align abnormal feature data with the context environment and operation parameters in time and space to construct a complete analysis dataset; The anomaly analysis module is used to interpret, correlate, and visualize the output results of the multimodal deep learning model.

Images