Intelligent analysis method and system based on wms system and ai large model

Abstract

The application provides an intelligent analysis method and system based on a WMS system and an AI large model, relates to the field of warehouse management, and comprises the following steps: converting WMS business data into description text, identifying and labeling abnormalities by using an AI model, and constructing a causal knowledge graph associated with risk weights. Based on the graph, a candidate decision scheme is generated by multi-path reasoning, and the feasibility and risk exposure of the candidate decision scheme are evaluated to select the optimal scheme. Finally, the scheme is simulated for risk propagation and a blocking measure is generated. The application realizes the automatic identification, quantitative evaluation and intelligent decision of warehouse operation risks, and improves the timeliness and accuracy of risk response.

Description

Technical Field

[0001] This invention relates to warehouse management technology, and more particularly to an intelligent analysis method and system based on a WMS system and an AI large model. Background Technology

[0002] In the field of warehousing and logistics management, Warehouse Management Systems (WMS) have become a core tool for modern warehousing operations, responsible for collecting and recording business data across the entire process of receiving, storing, picking, and shipping. As enterprises accumulate increasingly rich amounts of data, how to extract valuable information from massive amounts of business data to optimize operations and prevent risks has become a focus of industry attention. Currently, the conventional practices in this field mainly rely on rule-based systems and traditional statistical analysis models. Specifically, a common approach is for operations personnel or data analysts to monitor key performance indicators (KPIs) generated in the WMS (such as inventory turnover, order fulfillment rate, and equipment failure alarms) by writing predefined rules or thresholds. When a KPI reaches the boundary set by the rule, the system triggers a corresponding alert. Another approach is to use historical business data to build statistical models or simple machine learning models, such as time series analysis for predicting inventory demand, or classification models for identifying common order anomaly patterns. These methods aim to transform data into actionable insights to assist managers in decision-making.

[0003] However, the aforementioned existing practices have significant limitations. On the one hand, systems based on predefined rules lack sufficient flexibility. Warehouse operations are complex and ever-changing, with numerous types of anomalies and constantly emerging new patterns. Fixed rule sets struggle to comprehensively cover all potential risks, especially complex anomalies involving multiple intertwined factors and intricate causal relationships. This leads to either a large number of false alarms, increasing the operational burden, or the omission of important hidden risks until they develop into actual operational problems. On the other hand, traditional statistical analysis or shallow machine learning models have limited processing capabilities. These models are typically designed for single, structured indicators, lacking the ability to understand the semantics of unstructured business contexts and failing to correlate different stages and types of anomalies holistically. Their analytical results are often isolated and one-sided, failing to reveal the deep causal chains and diffusion paths behind anomalies, thus failing to provide sufficient support for overall risk assessment and systematic decision-making, resulting in insufficient effectiveness and robustness of decision-making solutions. Summary of the Invention

[0004] This invention provides an intelligent analysis method and system based on a WMS system and an AI large model, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an intelligent analysis method based on a WMS system and a large AI model, comprising:

[0006] Obtain business data from the warehouse management system and convert the business data into warehouse operation description text;

[0007] The warehouse operation description text is input into a pre-trained artificial intelligence language model to identify abnormal segments, and the abnormal segments are labeled with abnormality type labels and impact scope labels to generate labeled text.

[0008] Extract causal relationships from the annotated text, calculate risk weight values ​​based on anomaly type labels, and construct a risk knowledge graph by associating the risk weight values ​​with causal relationships.

[0009] The risk knowledge graph is input into a trained artificial intelligence language model for multi-path reasoning to generate multiple candidate decision schemes. The feasibility score of each candidate decision scheme under the constraint set and the risk exposure in the risk knowledge graph are calculated. The candidate decision scheme with the lowest feasibility score and the lowest risk exposure is selected as the decision scheme.

[0010] The decision-making scheme is simulated to propagate along the causal relationship path in the risk knowledge graph. The propagation intensity is calculated based on the risk weight value. Propagation nodes with propagation intensity exceeding the preset propagation threshold are identified. Blocking measures are generated for the propagation nodes. The decision-making scheme and risk blocking measures are sent to the warehouse management system for execution.

[0011] Retrieving business data from the warehouse management system and converting it into warehouse operation description text includes:

[0012] Extract business data from the database of the warehouse management system. The business data includes inventory status data, order flow data, equipment operation data, and personnel operation data.

[0013] Construct a business entity dictionary to map the identification information in the business data to business entities;

[0014] A business entity sequence is generated based on the business entities in the business entity dictionary. Numerical and status information in the business data are extracted as attribute information. The attribute information is associated with the corresponding business entities in the business entity sequence to form a business entity sequence with attribute annotations.

[0015] The business entities in the business entity sequence are sorted according to their timestamps, the business relationships between adjacent business entities are identified, and connecting words representing business actions are generated based on the business relationships. The connecting words are then inserted between adjacent business entities to generate warehouse operation description text.

[0016] The warehouse operation description text is input into a pre-trained artificial intelligence language model to identify abnormal segments. The abnormal segments are then labeled with anomaly type and impact scope tags, generating labeled text including:

[0017] The text describing warehouse operations is input into a pre-trained artificial intelligence language model for semantic encoding, generating a vector to be analyzed.

[0018] Construct a normal operation vector library, calculate the semantic distance between the vector to be analyzed and each normal operation vector in the normal operation vector library, obtain the minimum semantic distance, and identify text fragments in the warehouse operation description text whose minimum semantic distance exceeds the distance threshold as abnormal fragments;

[0019] The business entity type and business action type in the abnormal fragment are extracted by a pre-trained artificial intelligence language model, an abnormal classification rule base is constructed, and the abnormal type label is determined by matching the business entity type and business action type in the abnormal classification rule base and labeled to the abnormal fragment.

[0020] Extract the timestamp and location information of the abnormal fragment in the warehouse operation description text, parse the preceding and subsequent business nodes associated with the timestamp, calculate the business link length based on the preceding business node, subsequent business node and location information, generate the impact range label according to the business link length, and mark it to the abnormal fragment.

[0021] The exception fragments labeled with exception type and scope of impact are integrated into the warehouse operation description text to generate labeled text.

[0022] The training process of the artificial intelligence language model includes:

[0023] A dual-tower contrastive learning model is constructed as an artificial intelligence language model to be trained. The dual-tower contrastive learning model includes a text encoding tower and a knowledge encoding tower. The text encoding tower is used to encode the text describing warehouse operations to generate text vectors, and the knowledge encoding tower is used to encode the knowledge in the warehouse domain to generate knowledge vectors.

[0024] Collect historical business data from the warehouse management system, extract historical abnormal events and handling solutions from the historical business data, construct a training sample library, and divide the historical business data in the training sample library into a training set and a validation set according to a preset ratio.

[0025] Historical business data from the training set is input into the AI ​​language model to be trained for pre-training. The model parameters are optimized by maximizing the mutual information between the text vector and the knowledge vector to obtain the pre-trained model.

[0026] Construct a multi-task training objective function, which includes a loss function for identifying causal relationships, a loss function for classifying anomaly types, and a loss function for assessing risk levels;

[0027] The validation set data is input into the pre-trained model, the loss value of the multi-task training objective function is calculated, the model parameters are optimized through backpropagation, and when the loss value is less than the preset convergence threshold, the trained artificial intelligence language model is output.

[0028] Parsing and annotating text to extract causal relationships, calculating risk weight values ​​based on anomaly type labels, and constructing a risk knowledge graph by associating the risk weight values ​​with causal relationships include:

[0029] Extract part-of-speech and syntactic features from the annotated text, identify event trigger words and event entities in the annotated text, determine the combination relationship between event entities, and generate an event entity sequence;

[0030] Extract timestamp information from the event entity sequence, form event entity pairs by combining event entities with temporal correlation, calculate the semantic similarity of the event entity pairs, and obtain the semantic features of the event entity pairs;

[0031] The causal association strength between event entity pairs is calculated based on the semantic features of the event entity pairs. Event entity pairs whose causal association strength exceeds a preset strength threshold are filtered out. The filtered event entity pairs are determined as causal relationship pairs, and causal relationship triples containing causal association strength are generated.

[0032] Extract anomaly type labels from the labeled text, calculate the semantic representation of the anomaly type labels, and convert the semantic representation into risk weight values;

[0033] The event entities in the causal triples are constructed as graph nodes, and the causal association strength is constructed as the weight of the graph edge to generate an initial causal graph.

[0034] Risk weight values ​​are assigned to the corresponding abnormal event nodes in the initial causal graph, and risk propagation attenuation values ​​are calculated as the propagation attributes of the graph edges to generate a risk knowledge graph.

[0035] The risk knowledge graph is input into a trained AI language model for multi-path reasoning to generate multiple candidate decision schemes. The feasibility score of each candidate decision scheme under the constraint set and its risk exposure in the risk knowledge graph are calculated. The candidate decision scheme with the lowest risk exposure and the highest feasibility score is selected as the final decision scheme.

[0036] Extract node event attributes and edge risk attributes from the risk knowledge graph, construct graph feature vectors, and generate graph sequence representations by arranging them in chronological order;

[0037] Input the graph sequence representation into the trained artificial intelligence language model, identify risk propagation links based on the strength of causal association, calculate the risk level of the link based on the risk propagation attenuation value, determine the critical risk path, and generate a risk avoidance path containing risk avoidance nodes.

[0038] Extract risk avoidance nodes from the risk avoidance path, match the corresponding countermeasures for the risk avoidance nodes, combine them in chronological order to generate a sequence of countermeasures, and construct candidate decision schemes;

[0039] Extract business constraint rules to construct a constraint set, calculate the execution cost of countermeasures based on the constraint set, and transform the execution cost into a feasibility score for candidate decision solutions;

[0040] Calculate the local risk value of the response node, calculate the risk propagation value based on the risk propagation attenuation value, and combine the local risk value and the risk propagation value to form the risk exposure of the candidate decision scheme;

[0041] Candidate decision-making schemes with feasibility scores greater than a preset score threshold and the lowest risk exposure are selected as the decision-making schemes.

[0042] The decision-making scheme is simulated to propagate along the causal relationship path in the risk knowledge graph. The propagation strength is calculated based on the risk weight value. Propagation nodes with propagation strength exceeding a preset propagation threshold are identified. Blocking measures are generated for the propagation nodes. The decision-making scheme and risk blocking measures are sent to the warehouse management system for execution, including:

[0043] Extract the sequence of operation instructions from the decision-making scheme, parse the sequence of operation instructions to form an event feature vector, match the event node corresponding to the event feature vector in the risk knowledge graph, and obtain the risk weight value of the event node;

[0044] The event nodes are input into the risk knowledge graph for path expansion. Connected event nodes are obtained based on causal relationship paths. The propagation attenuation coefficient of the causal relationship paths is extracted to construct a propagation impact path graph.

[0045] The risk weight value is used as the propagation energy value. The propagation energy value is transmitted along the causal path in the propagation influence path diagram. The propagation intensity is obtained by multiplying the propagation attenuation coefficient by the propagation energy value.

[0046] The propagation intensity is compared with a preset propagation threshold, and event nodes that exceed the preset propagation threshold are selected as propagation nodes;

[0047] Extract the node connection relationships between propagation nodes, obtain the risk prevention and control rules corresponding to the propagation nodes from the risk knowledge graph, and generate blocking measures based on the node connection relationships to organize the risk prevention and control rules.

[0048] The decision-making schemes and blocking measures are sorted according to their propagation intensity to generate an execution queue. The execution queue is then converted into system call instructions and sent to the warehouse management system for execution.

[0049] A second aspect of this invention provides an intelligent analysis system based on a WMS system and an AI large model, comprising:

[0050] The data acquisition unit is used to acquire business data from the warehouse management system and convert the business data into warehouse operation description text.

[0051] An anomaly identification unit is used to input the warehouse operation description text into a pre-trained artificial intelligence language model to identify abnormal segments, label the abnormal segments with anomaly type labels and impact scope labels, and generate labeled text.

[0052] The graph construction unit is used to parse the labeled text to extract causal relationships, calculate risk weight values ​​based on anomaly type labels, and construct a risk knowledge graph by associating the risk weight values ​​with causal relationships.

[0053] The decision generation unit is used to input the risk knowledge graph into the trained artificial intelligence language model for multi-path reasoning, generate multiple candidate decision schemes, calculate the feasibility score of each candidate decision scheme under the constraint set and the risk exposure in the risk knowledge graph, and select the candidate decision scheme with the lowest feasibility score and the lowest risk exposure as the decision scheme.

[0054] The risk blocking unit is used to simulate the propagation of the decision plan along the causal relationship path in the risk knowledge graph, calculate the propagation intensity based on the risk weight value, identify propagation nodes whose propagation intensity exceeds the preset propagation threshold, generate blocking measures for the propagation nodes, and send the decision plan and risk blocking measures to the warehouse management system for execution.

[0055] A third aspect of the present invention provides an electronic device, comprising:

[0056] processor;

[0057] Memory used to store processor-executable instructions;

[0058] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0059] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0060] In this embodiment, structured business data in the warehouse management system can be automatically converted into natural language descriptions of warehouse operations, thus generating unstructured text data. Utilizing a pre-trained AI language model to intelligently analyze the descriptive text, abnormal segments can be efficiently identified, and the anomaly type and impact range can be automatically labeled, generating structured labeled text, significantly improving the accuracy and automation level of anomaly identification. The optimal decision-making scheme with high feasibility and lowest risk exposure is automatically selected. The selected decision-making scheme is simulated along a causal path, dynamically calculating the risk propagation intensity and accurately identifying propagation nodes where the propagation intensity exceeds a preset threshold. Corresponding risk mitigation measures are automatically generated for these key nodes, and the decision-making scheme and mitigation measures are sent to the warehouse management system for execution. This allows for the pre-assessment and control of potential risk spread before decision execution, improving the robustness and security of warehouse operations. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the intelligent analysis method based on a WMS system and a large AI model, according to an embodiment of the present invention.

[0062] Figure 2 This is a flowchart illustrating the decision-making, execution, and blocking process of the risk knowledge graph in an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0065] Figure 1 This is a flowchart illustrating the intelligent analysis method based on a WMS system and a large AI model according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] Obtain business data from the warehouse management system and convert the business data into warehouse operation description text;

[0067] The warehouse operation description text is input into a pre-trained artificial intelligence language model to identify abnormal segments, and the abnormal segments are labeled with abnormality type labels and impact scope labels to generate labeled text.

[0068] Extract causal relationships from the annotated text, calculate risk weight values ​​based on anomaly type labels, and construct a risk knowledge graph by associating the risk weight values ​​with causal relationships.

[0069] The risk knowledge graph is input into a trained artificial intelligence language model for multi-path reasoning to generate multiple candidate decision schemes. The feasibility score of each candidate decision scheme under the constraint set and the risk exposure in the risk knowledge graph are calculated. The candidate decision scheme with the lowest feasibility score and the lowest risk exposure is selected as the decision scheme.

[0070] The decision-making scheme is simulated to propagate along the causal relationship path in the risk knowledge graph. The propagation intensity is calculated based on the risk weight value. Propagation nodes with propagation intensity exceeding the preset propagation threshold are identified. Blocking measures are generated for the propagation nodes. The decision-making scheme and risk blocking measures are sent to the warehouse management system for execution.

[0071] Retrieving business data from the warehouse management system and converting it into warehouse operation description text includes:

[0072] Extract business data from the database of the warehouse management system. The business data includes inventory status data, order flow data, equipment operation data, and personnel operation data.

[0073] Construct a business entity dictionary to map the identification information in the business data to business entities;

[0074] A business entity sequence is generated based on the business entities in the business entity dictionary. Numerical and status information in the business data are extracted as attribute information. The attribute information is associated with the corresponding business entities in the business entity sequence to form a business entity sequence with attribute annotations.

[0075] The business entities in the business entity sequence are sorted according to their timestamps, the business relationships between adjacent business entities are identified, and connecting words representing business actions are generated based on the business relationships. The connecting words are then inserted between adjacent business entities to generate warehouse operation description text.

[0076] Business data is extracted from the warehouse management system database according to data table type. Inventory status data includes shelf number, product category identifier, storage quantity, and occupied capacity ratio, obtained by querying the relationship fields between the main inventory table and the storage location table. Order flow data includes order creation timestamp, picking start timestamp, review completion timestamp, outbound timestamp, and order status code, obtained from the order lifecycle table. Equipment operation data is extracted from the equipment monitoring table, including forklift equipment identifier, AGV vehicle identifier, conveyor belt identifier, and their corresponding operating status values, fault codes, and number of operations. Personnel operation data includes employee ID, work area code, work type code, and work start and end timestamps.

[0077] When constructing the business entity dictionary, shelf numbers are mapped to "shelf entities," order numbers to "order entities," equipment identifiers to "equipment entities," and employee numbers to "personnel entities." This mapping process is completed through a predefined entity type table, which stores the correspondence between business data field names and entity types. When the shelf number "A03-12" is retrieved from the main inventory table, the entity type table is queried based on the field name "shelf number," and it is mapped to a shelf entity.

[0078] A sequence of business entities is generated based on the timestamp field of the business data. Inventory quantity and occupied floor space ratio are extracted as attribute information for shelf entities, order status codes as attribute information for order entities, and equipment operating status values ​​and fault codes as attribute information for equipment entities. Attribute information is associated with corresponding entities in key-value pairs, forming attribute-labeled sequences such as "Shelf entity A03-12 [Inventory quantity: 320][Floor space ratio: 0.85]" and "Order entity WH20240315001 [Status: Picked]".

[0079] The business entity sequence is sorted in ascending order by timestamp to identify business relationships between adjacent entities. When the difference between the work timestamp of the personnel entity and the picking start timestamp of the order entity is less than 5 seconds, a "picking execution" relationship is determined. When the outbound timestamp of the order entity and the work start timestamp of the equipment entity are close and the equipment type is forklift, a "loading and transportation" relationship is determined. Based on the relationship type, the corresponding action word is selected from the connector word library. The connector word "picking in progress" corresponding to the "picking execution" relationship is inserted between the personnel entity and the order entity, and the connector word "transported by forklift" corresponding to the "loading and transportation" relationship is inserted between the order entity and the equipment entity. Finally, a warehouse operation description text is generated: "Personnel entity WK0025 is picking order entity WH20240315001 [Status: In Progress], the order entity completed outbound transportation by forklift to equipment entity FK007." This text retains timestamp information and attribute annotations, providing structured text input for subsequent anomaly identification.

[0080] The warehouse operation description text is input into a pre-trained artificial intelligence language model to identify abnormal segments. The abnormal segments are then labeled with anomaly type and impact scope tags, generating labeled text including:

[0081] The text describing warehouse operations is input into a pre-trained artificial intelligence language model for semantic encoding, generating a vector to be analyzed.

[0082] Construct a normal operation vector library, calculate the semantic distance between the vector to be analyzed and each normal operation vector in the normal operation vector library, obtain the minimum semantic distance, and identify text fragments in the warehouse operation description text whose minimum semantic distance exceeds the distance threshold as abnormal fragments;

[0083] The business entity type and business action type in the abnormal fragment are extracted by a pre-trained artificial intelligence language model, an abnormal classification rule base is constructed, and the abnormal type label is determined by matching the business entity type and business action type in the abnormal classification rule base and labeled to the abnormal fragment.

[0084] Extract the timestamp and location information of the abnormal fragment in the warehouse operation description text, parse the preceding and subsequent business nodes associated with the timestamp, calculate the business link length based on the preceding business node, subsequent business node and location information, generate the impact range label according to the business link length, and mark it to the abnormal fragment.

[0085] The exception fragments labeled with exception type and scope of impact are integrated into the warehouse operation description text to generate labeled text.

[0086] After obtaining the warehouse operation description text, an encoder is used to process the text. Specifically, the warehouse operation description text is segmented into several semantic units, each containing a complete business description fragment. A pre-trained BERT (Bidirectional Encoder Representations from Transformers) type language model is used to vectorize each semantic unit. The model's encoding layer maps the text into a 768-dimensional vector representation, which can capture the deep semantic features of the text and serves as the vector to be analyzed.

[0087] The normal operation vector library is constructed based on historical normal operation data. Warehouse operation texts with no abnormal records in the past six months are selected and processed using the same semantic encoding process to generate a vector set. This vector library contains at least 5000 normal operation vector samples, covering various routine business scenarios such as inbound, outbound, inventory, and transfer. The cosine distance between the vector to be analyzed and each vector in the normal operation vector library is calculated, and the minimum value among all distance values ​​is recorded. When the minimum semantic distance exceeds the set distance threshold of 0.35, it indicates that the semantic unit deviates significantly from the normal operation pattern, and the corresponding text segment is marked as an abnormal segment.

[0088] Fine-grained analysis is performed on the identified abnormal segments. Named entity recognition technology is used to extract business entities from the segments, including key entities such as product numbers, storage location numbers, equipment identifiers, and employee numbers. These entities are categorized into four business entity types: materials, spaces, equipment, and personnel. Simultaneously, action verbs such as "delay," "missing," "error," and "timeout" are extracted from the segments and mapped to abnormal action types. An anomaly classification rule base predefines combination rules between business entity types and abnormal action types; for example, "material + missing" corresponds to "inventory discrepancy," and "equipment + fault" corresponds to "equipment anomaly." Based on the extracted entity types and action types, a matching query is performed in the rule base to determine the anomaly type tag for the abnormal segment, such as "picking timeout," "inventory discrepancy," or "path blockage," and the tag is appended to the abnormal segment in a structured tagging format.

[0089] Determining the scope of impact relies on analysis across time and space dimensions. Timestamp information, recording the specific moment the anomaly occurred, is extracted from the anomaly fragment. The business process sequence of the warehouse operation description text is parsed to locate the timestamp's position in the business chain. The nearest business node before the timestamp is identified as the preceding business node, and the nearest business node after the timestamp is identified as the following business node. Simultaneously, the storage location number or work area code is extracted from the anomaly fragment as location information. The number of business steps involved from the preceding to the following business node is calculated. Combined with the location information, the affected storage area range is determined, and the product of the number of business steps and the number of storage areas is taken as the business link length. When the business link length is less than 3, a "local impact" label is generated; when the length is between 3 and 8, a "regional impact" label is generated; and when the length exceeds 8, a "global impact" label is generated. This impact scope label is also annotated to the corresponding anomaly fragment in a structured format.

[0090] After annotation, all anomalous fragments, along with their accompanying anomalous type and impact scope labels, are re-embedded into their corresponding positions within the original warehouse operation description text. This maintains the original text order and semantic coherence, forming an annotated text containing complete annotation information. This annotated text preserves the integrity of the original business description while adding structured anomalous information, providing a clear data foundation for subsequent causal relationship extraction and risk analysis.

[0091] The training process of the artificial intelligence language model includes:

[0092] A dual-tower contrastive learning model is constructed as an artificial intelligence language model to be trained. The dual-tower contrastive learning model includes a text encoding tower and a knowledge encoding tower. The text encoding tower is used to encode the text describing warehouse operations to generate text vectors, and the knowledge encoding tower is used to encode the knowledge in the warehouse domain to generate knowledge vectors.

[0093] Collect historical business data from the warehouse management system, extract historical abnormal events and handling solutions from the historical business data, construct a training sample library, and divide the historical business data in the training sample library into a training set and a validation set according to a preset ratio.

[0094] Historical business data from the training set is input into the AI ​​language model to be trained for pre-training. The model parameters are optimized by maximizing the mutual information between the text vector and the knowledge vector to obtain the pre-trained model.

[0095] Construct a multi-task training objective function, which includes a loss function for identifying causal relationships, a loss function for classifying anomaly types, and a loss function for assessing risk levels;

[0096] The validation set data is input into the pre-trained model, the loss value of the multi-task training objective function is calculated, the model parameters are optimized through backpropagation, and when the loss value is less than the preset convergence threshold, the trained artificial intelligence language model is output.

[0097] During the model training phase, a dual-tower contrastive learning architecture is first constructed as the core training framework. The text encoding tower employs a multi-layer Transformer structure, containing 12 encoding layers, each with 768-dimensional hidden states and 12 attention heads, used to map warehouse operation description text into fixed-dimensional semantic vectors. The knowledge encoding tower employs a graph neural network structure, containing 4 graph convolutional layers, each with 512-dimensional features, used to structurally encode professional knowledge in the warehousing domain, mapping knowledge entities and their relationships into a knowledge vector space. The output vector dimension of both encoding towers is set to 256 dimensions to facilitate subsequent similarity calculations.

[0098] Business records from the past three years were collected from the historical database of the warehouse management system. Event data containing anomaly identifiers was extracted. Each data entry included the anomaly occurrence time, anomaly type description, affected operational processes, handling measures taken, and results. The collected 80,000 historical anomaly events were divided in a 7:3 ratio, with 56,000 used as the training set and 24,000 as the validation set. The training set data was preprocessed, converting structured fields into natural language descriptions. For example, "Outbound delay, wave number W20231205, delay duration 120 minutes, reason: insufficient picking personnel" was transformed into "Outbound operation of wave W20231205 was delayed by 120 minutes due to insufficient picking personnel."

[0099] In the pre-training phase, a contrastive learning strategy is employed to optimize model parameters. Text descriptions from the training set are input into a text encoding tower to obtain text vectors, and corresponding knowledge entries are input into a knowledge encoding tower to obtain knowledge vectors. The cosine similarity between the two types of vectors is calculated. Positive and negative sample pairs are constructed: positive pairs consist of text and knowledge vectors for the same anomalous event, while negative pairs are combinations of vectors from different anomalous events. The InfoNCE loss function is used to maximize the similarity of positive sample pairs while minimizing the similarity of negative sample pairs, with the optimization objective being to align the text and knowledge representations in the vector space. The pre-training process uses a batch size of 64 and a learning rate of 0.0001, and the pre-trained model is obtained after 50 iterations.

[0100] A multi-task learning framework is constructed during the fine-tuning phase. The causal relationship identification task employs a sequence labeling loss function, using conditional random fields to calculate the probability distribution of the labeled sequences, with the optimization objective being to maximize the log-likelihood of the true causal label sequences. The anomaly type classification task uses a cross-entropy loss function, comparing the model's output classification probabilities with the true anomaly type labels to calculate the classification error. The risk level assessment task uses a mean squared error loss function, regressing the predicted risk score against the expert-labeled risk level. The loss functions of the three sub-tasks are combined using a weighted summation method, with weight coefficients set to 0.4, 0.3, and 0.3 respectively, forming a unified multi-task loss function.

[0101] The pre-trained model is fine-tuned using validation set data, with the total loss of the multi-task loss function calculated in each iteration. The AdamW optimizer is used to update model parameters, with a learning rate of 0.00005 and a gradient clipping threshold of 1.0 to prevent gradient explosion. Model convergence is determined when the loss value decreases by less than 0.001 over five consecutive iterations, and the overall evaluation metric on the validation set no longer improves. The converged model parameters are then fixed, and a fully trained model usable in real-world business scenarios is output. The entire training process utilizes mixed-precision computing to accelerate training, and completing the entire training process on a server with four GPUs takes approximately 18 hours.

[0102] Parsing and annotating text to extract causal relationships, calculating risk weight values ​​based on anomaly type labels, and constructing a risk knowledge graph by associating the risk weight values ​​with causal relationships include:

[0103] Extract part-of-speech and syntactic features from the annotated text, identify event trigger words and event entities in the annotated text, determine the combination relationship between event entities, and generate an event entity sequence;

[0104] Extract timestamp information from the event entity sequence, form event entity pairs by combining event entities with temporal correlation, calculate the semantic similarity of the event entity pairs, and obtain the semantic features of the event entity pairs;

[0105] The causal association strength between event entity pairs is calculated based on the semantic features of the event entity pairs. Event entity pairs whose causal association strength exceeds a preset strength threshold are filtered out. The filtered event entity pairs are determined as causal relationship pairs, and causal relationship triples containing causal association strength are generated.

[0106] Extract anomaly type labels from the labeled text, calculate the semantic representation of the anomaly type labels, and convert the semantic representation into risk weight values;

[0107] The event entities in the causal triples are constructed as graph nodes, and the causal association strength is constructed as the weight of the graph edge to generate an initial causal graph.

[0108] Risk weight values ​​are assigned to the corresponding abnormal event nodes in the initial causal graph, and risk propagation attenuation values ​​are calculated as the propagation attributes of the graph edges to generate a risk knowledge graph.

[0109] In the process of extracting causal relationships from annotated text and constructing a risk knowledge graph, natural language processing techniques are used to perform part-of-speech tagging and syntactic analysis on the annotated text. Dependency parsing trees are used to identify verbal trigger words, which typically include words indicating causal relationships such as "lead to," "cause," "incident," and "affect." Simultaneously, noun phrases associated with the trigger words are extracted as event entities, such as specific events in warehouse operations like "inventory backlog," "decreased picking efficiency," and "excessive occupancy rate." The subject-verb-object and modifier relationships between event entities are analyzed to determine which entities are causal events and which are result events, thereby generating an ordered sequence containing multiple event entities.

[0110] For a sequence of event entities, the corresponding time information, including the event occurrence time and duration, is extracted from the timestamp field of the original business data. A temporal sequence is determined, pairing earlier-occurring events with later-occurring events to form candidate event entity pairs. A pre-trained word vector model is used to convert the event entities into high-dimensional semantic vectors. The cosine similarity between the two entity vectors in each event entity pair is calculated; this similarity value reflects the degree of association between the two events in the semantic space and serves as the semantic feature of the event entity pair.

[0111] Based on the semantic features of event entity pairs, the causal association strength is calculated by combining the co-occurrence frequency of event entities in the labeled text and the dependency distance in the context window. Specifically, semantic similarity, temporal correlation, and context distance are weighted and fused, with weighting coefficients determined based on statistical analysis of historical warehousing data. A preset strength threshold of 0.6 is set, and event entity pairs with causal association strength exceeding this threshold are filtered and retained. For the filtered event entity pairs, triples are formed according to the structure of cause event-relationship type-result event, where the relationship type is marked as "causal association," and the calculated causal association strength is appended as the attribute value of the triple.

[0112] For anomaly type tags in the labeled text, such as categories like "inventory anomaly," "efficiency anomaly," and "equipment anomaly," tag embedding technology is used to map each anomaly type to a fixed-dimensional semantic representation vector. Based on historical business loss statistics caused by anomaly events, a mapping relationship between anomaly types and risk levels is established. A linear transformation converts the semantic representation vector into risk weight values ​​between 0 and 1, where high-risk anomaly types correspond to weight values ​​close to 1, and low-risk anomaly types correspond to weight values ​​close to 0.

[0113] Each event entity in the causal triple is created as a node in the graph, with node attributes including event name, event type, and timestamp. Directed edges are established between causally related event nodes, with the edges pointing from the cause event to the result event, and the edge weight is assigned the corresponding causal relationship strength value. A graph database storage structure is used to organize the relationships between nodes and edges, forming an initial causal graph.

[0114] Based on the initial causal graph, event nodes marked as anomalous in the annotated text are retrieved, and the risk weight values ​​corresponding to these anomalous event nodes are assigned as the risk attributes of the nodes. For each directed edge in the graph, the attenuation coefficient of risk during propagation is calculated according to the propagation characteristics of causal links. The calculation of the attenuation coefficient comprehensively considers the strength of causal association and the graph distance between nodes; the greater the distance, the greater the attenuation. The calculated attenuation values ​​are stored as the propagation attributes of the edges, ultimately forming a complete risk knowledge graph containing node risk weights and edge propagation attributes.

[0115] The risk knowledge graph is input into a trained AI language model for multi-path reasoning to generate multiple candidate decision schemes. The feasibility score of each candidate decision scheme under the constraint set and its risk exposure in the risk knowledge graph are calculated. The candidate decision scheme with the lowest risk exposure and the highest feasibility score is selected as the final decision scheme.

[0116] Extract node event attributes and edge risk attributes from the risk knowledge graph, construct graph feature vectors, and generate graph sequence representations by arranging them in chronological order;

[0117] Input the graph sequence representation into the trained artificial intelligence language model, identify risk propagation links based on the strength of causal association, calculate the risk level of the link based on the risk propagation attenuation value, determine the critical risk path, and generate a risk avoidance path containing risk avoidance nodes.

[0118] Extract risk avoidance nodes from the risk avoidance path, match the corresponding countermeasures for the risk avoidance nodes, combine them in chronological order to generate a sequence of countermeasures, and construct candidate decision schemes;

[0119] Extract business constraint rules to construct a constraint set, calculate the execution cost of countermeasures based on the constraint set, and transform the execution cost into a feasibility score for candidate decision solutions;

[0120] Calculate the local risk value of the response node, calculate the risk propagation value based on the risk propagation attenuation value, and combine the local risk value and the risk propagation value to form the risk exposure of the candidate decision scheme;

[0121] Candidate decision-making schemes with feasibility scores greater than a preset score threshold and the lowest risk exposure are selected as the decision-making schemes.

[0122] The risk knowledge graph is traversed to extract node event attributes, including event type, occurrence time, duration, and impact scope. These attributes are then converted into multi-dimensional vector representations. All edge relationships in the graph are traversed to extract edge risk attributes, including risk weight, propagation probability, and propagation delay. These attributes are converted into relation vectors. The node event attribute vectors and edge risk attribute vectors are concatenated according to the graph topology to form a complete graph feature vector. Finally, the graph feature vectors are sorted according to the chronological order of occurrence of abnormal events in the business process to generate a graph sequence representation.

[0123] The generated graph sequence is input into a trained AI language model, which encodes the node relationships in the graph sequence based on an attention mechanism. The causal association strength between any two nodes in the graph is calculated; this strength is equal to the product of the propagation probabilities in the risk attributes of all edges along the connection path. Node pairs with causal association strengths greater than a preset threshold are identified, and their connection paths are marked as risk propagation links. For each risk propagation link, a risk propagation attenuation value is calculated based on the path length and propagation delay time; this attenuation value is negatively correlated with the path length. The risk weight values ​​of each edge along the propagation link are multiplied by the risk propagation attenuation value and then summed to obtain the link's risk level. The propagation link with the highest risk level is selected as the critical risk path. Risk avoidance nodes are inserted at the critical node positions of the critical risk path; these nodes are configured with the attribute of intervention measures capable of interrupting risk propagation, thus generating a risk avoidance path.

[0124] The process iterates through all marked risk avoidance nodes in the risk avoidance path, reading the event type attribute of each node. Matches are retrieved from a pre-defined response knowledge base, which stores the mapping between exception types and response measures. Matched response measures are extracted, including resource allocation, process adjustment, and equipment maintenance. These response measures are then sequentially combined according to the temporal position of the risk avoidance nodes in the path, forming a response measure sequence. This sequence is then encapsulated as a structured candidate decision plan, containing information such as the measure steps, execution order, and resource requirements.

[0125] Extract business constraint rules from the warehouse management system, including resource capacity limits, time window constraints, and operation sequence dependencies. After formalizing these rules, construct a constraint set. For each response measure in the candidate decision-making scheme, calculate the resource consumption, time cost, and manpower input required for its execution. Compare the execution cost of the response measure with the constraints in the constraint set, and statistically analyze the number and severity of constraint violations. Set a penalty coefficient based on the severity of constraint violations, and combine the execution cost with the penalty coefficient to calculate the comprehensive cost. Normalize the comprehensive cost, converting it into a feasibility score ranging from 0 to 1; a higher score indicates that the candidate decision-making scheme is easier to execute.

[0126] For each response node in the candidate decision scheme, the corresponding node position is located in the risk knowledge graph, and the node's risk weight value is read as the local risk value. All outgoing edges originating from this response node are traversed, obtaining the propagation probability of the edge and the risk weight value of the target node. The propagation probability is multiplied by the target node's risk weight value, and then multiplied by the previously calculated risk propagation attenuation value to obtain the risk propagation value of that path. The risk propagation values ​​of all outgoing edges of this node are summed to obtain the node's total risk propagation value. The local risk value of the response node and the total risk propagation value are weighted and summed according to a preset weight ratio to obtain the risk exposure of the candidate decision scheme.

[0127] Iterate through all candidate decision-making schemes and select those with feasibility scores greater than a preset threshold to enter the candidate set. Sort the schemes in the candidate set by risk exposure from low to high, and select the candidate decision-making scheme with the lowest risk exposure value. The selected candidate decision-making scheme is then determined as the final decision-making scheme for subsequent execution in the warehouse management system.

[0128] The decision-making scheme is simulated to propagate along the causal relationship path in the risk knowledge graph. The propagation strength is calculated based on the risk weight value. Propagation nodes with propagation strength exceeding a preset propagation threshold are identified. Blocking measures are generated for the propagation nodes. The decision-making scheme and risk blocking measures are sent to the warehouse management system for execution, including:

[0129] Extract the sequence of operation instructions from the decision-making scheme, parse the sequence of operation instructions to form an event feature vector, match the event node corresponding to the event feature vector in the risk knowledge graph, and obtain the risk weight value of the event node;

[0130] The event nodes are input into the risk knowledge graph for path expansion. Connected event nodes are obtained based on causal relationship paths. The propagation attenuation coefficient of the causal relationship paths is extracted to construct a propagation impact path graph.

[0131] The risk weight value is used as the propagation energy value. The propagation energy value is transmitted along the causal path in the propagation influence path diagram. The propagation intensity is obtained by multiplying the propagation attenuation coefficient by the propagation energy value.

[0132] The propagation intensity is compared with a preset propagation threshold, and event nodes that exceed the preset propagation threshold are selected as propagation nodes;

[0133] Extract the node connection relationships between propagation nodes, obtain the risk prevention and control rules corresponding to the propagation nodes from the risk knowledge graph, and generate blocking measures based on the node connection relationships to organize the risk prevention and control rules.

[0134] The decision-making schemes and blocking measures are sorted according to their propagation intensity to generate an execution queue. The execution queue is then converted into system call instructions and sent to the warehouse management system for execution.

[0135] In the risk propagation simulation phase, operational instruction sequences are extracted from the data structure of the decision-making scheme. Each operational instruction includes an operational type identifier, an associated business object code, and an execution timestamp. Semantic parsing is performed on the operational instruction sequences, converting the operational type into vector dimension values, mapping the business object code to a predefined entity space index, and normalizing the timestamp into a time feature component, combining these to form a multi-dimensional event feature vector. Vector similarity matching is performed on the node set of the risk knowledge graph. Cosine distance is used to measure the proximity between the event feature vector and the graph node features. The node with the highest similarity is selected as the matching event node, and the stored risk weight value is read from the attribute fields of that node.

[0136] Starting with the matched event node, traverse the directed edges marked as causal relationships in the risk knowledge graph to obtain the set of downstream event nodes directly connected to that node. Extract the attribute parameters of each causal relationship edge, including a propagation attenuation coefficient reflecting the degree of weakening of causal strength; this coefficient ranges from 0 to 1. Construct a directed graph structure of event nodes and their connected causal paths, with nodes storing event identifiers and edge weights recording the propagation attenuation coefficients, forming a data structure for the propagation impact path graph.

[0137] Energy transfer calculations are performed on the propagation impact path graph, setting the risk weight value of the starting event node as the initial propagation energy value. Following the topological order of the graph, energy is transferred from the starting node to downstream nodes along causal relationship edges. For each propagation path, the propagation energy value of the upstream node is multiplied by the propagation attenuation coefficient of the path edge, and the result is used as the propagation intensity value to the downstream node. When a node has multiple incoming edges, the energy values ​​transferred by all incoming edges are summed to obtain the total propagation intensity of that node.

[0138] The calculated propagation intensity of each node is compared numerically with a preset propagation threshold. This preset propagation threshold is pre-configured based on the risk tolerance of the warehousing operation, typically set as a specific percentage of the initial risk weight value. Nodes with propagation intensity values ​​greater than the preset threshold are selected and marked as propagation nodes requiring special attention; their node identifiers and propagation intensity values ​​are recorded.

[0139] The process involves extracting connection edges between nodes in the propagation node set and identifying node connection relationship types such as dependency, triggering, or concurrency relationships. Based on the event type identifier of each propagation node, the corresponding risk prevention and control rule is queried from the rule base of the risk knowledge graph. Risk prevention and control rules are stored in the form of condition-action pairs, containing triggering condition expressions and prevention and control action instructions. The execution order and combination method of the rules are determined based on the node connection relationships. For nodes with dependency relationships, a serial rule chain is used; for nodes with concurrent relationships, a parallel rule set is used. The organized rule sequence is then encapsulated into a structured blocking measure data package.

[0140] The operational instructions in the decision-making plan and the prevention and control actions in the containment measures are uniformly arranged into an execution task list. Tasks are sorted in descending order based on the propagation strength of their associated nodes, with tasks of higher propagation strength having higher priority and appearing at the front of the execution queue. For each task in the execution queue, according to the interface protocol specifications of the warehouse management system, the task parameters are converted into a system-recognizable call instruction format, including the interface address, request method, parameter key-value pairs, and authentication token. The converted system call instruction is sent to the execution module of the warehouse management system through the application programming interface, triggering corresponding business operations and risk control actions.

[0141] A second aspect of this invention provides an intelligent analysis system based on a WMS system and an AI large model, the system comprising:

[0142] The data acquisition unit is used to acquire business data from the warehouse management system and convert the business data into warehouse operation description text.

[0143] An anomaly identification unit is used to input the warehouse operation description text into a pre-trained artificial intelligence language model to identify abnormal segments, label the abnormal segments with anomaly type labels and impact scope labels, and generate labeled text.

[0144] The graph construction unit is used to parse the labeled text to extract causal relationships, calculate risk weight values ​​based on anomaly type labels, and construct a risk knowledge graph by associating the risk weight values ​​with causal relationships.

[0145] The decision generation unit is used to input the risk knowledge graph into the trained artificial intelligence language model for multi-path reasoning, generate multiple candidate decision schemes, calculate the feasibility score of each candidate decision scheme under the constraint set and the risk exposure in the risk knowledge graph, and select the candidate decision scheme with the lowest feasibility score and the lowest risk exposure as the decision scheme.

[0146] The risk blocking unit is used to simulate the propagation of the decision plan along the causal relationship path in the risk knowledge graph, calculate the propagation intensity based on the risk weight value, identify propagation nodes whose propagation intensity exceeds the preset propagation threshold, generate blocking measures for the propagation nodes, and send the decision plan and risk blocking measures to the warehouse management system for execution.

[0147] A third aspect of the present invention provides an electronic device, comprising:

[0148] processor;

[0149] Memory used to store processor-executable instructions;

[0150] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0151] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0152] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent analysis method based on WMS system and AI large model, characterized in that, include: Obtain business data from the warehouse management system and convert the business data into warehouse operation description text; The warehouse operation description text is input into a pre-trained artificial intelligence language model to identify abnormal segments, and the abnormal segments are labeled with abnormality type labels and impact scope labels to generate labeled text. Extract causal relationships from the annotated text, calculate risk weight values ​​based on anomaly type labels, and construct a risk knowledge graph by associating the risk weight values ​​with causal relationships. The risk knowledge graph is input into a trained artificial intelligence language model for multi-path reasoning to generate multiple candidate decision schemes. The feasibility score of each candidate decision scheme under the constraint set and the risk exposure in the risk knowledge graph are calculated. The candidate decision scheme with the lowest feasibility score and the lowest risk exposure is selected as the decision scheme. The decision-making scheme is simulated to propagate along the causal relationship path in the risk knowledge graph. The propagation intensity is calculated based on the risk weight value. Propagation nodes with propagation intensity exceeding the preset propagation threshold are identified. Blocking measures are generated for the propagation nodes. The decision-making scheme and risk blocking measures are sent to the warehouse management system for execution.

2. The method according to claim 1, characterized in that, Retrieving business data from the warehouse management system and converting it into warehouse operation description text includes: Extract business data from the database of the warehouse management system. The business data includes inventory status data, order flow data, equipment operation data, and personnel operation data. Construct a business entity dictionary to map the identification information in the business data to business entities; A business entity sequence is generated based on the business entities in the business entity dictionary. Numerical and status information in the business data are extracted as attribute information. The attribute information is associated with the corresponding business entities in the business entity sequence to form a business entity sequence with attribute annotations. The business entities in the business entity sequence are sorted according to their timestamps, the business relationships between adjacent business entities are identified, and connecting words representing business actions are generated based on the business relationships. The connecting words are then inserted between adjacent business entities to generate warehouse operation description text.

3. The method according to claim 1, characterized in that, The warehouse operation description text is input into a pre-trained artificial intelligence language model to identify abnormal segments. The abnormal segments are then labeled with anomaly type and impact scope tags, generating labeled text including: The text describing warehouse operations is input into a pre-trained artificial intelligence language model for semantic encoding, generating a vector to be analyzed. Construct a normal operation vector library, calculate the semantic distance between the vector to be analyzed and each normal operation vector in the normal operation vector library, obtain the minimum semantic distance, and identify text fragments in the warehouse operation description text whose minimum semantic distance exceeds the distance threshold as abnormal fragments; The business entity type and business action type in the abnormal fragment are extracted by a pre-trained artificial intelligence language model, an abnormal classification rule base is constructed, and the abnormal type label is determined by matching the business entity type and business action type in the abnormal classification rule base and labeled to the abnormal fragment. Extract the timestamp and location information of the abnormal fragment in the warehouse operation description text, parse the preceding and subsequent business nodes associated with the timestamp, calculate the business link length based on the preceding business node, subsequent business node and location information, generate the impact range label according to the business link length, and mark it to the abnormal fragment. The exception fragments labeled with exception type and scope of impact are integrated into the warehouse operation description text to generate labeled text.

4. The method according to claim 1, characterized in that, The training process of the artificial intelligence language model includes: A dual-tower contrastive learning model is constructed as an artificial intelligence language model to be trained. The dual-tower contrastive learning model includes a text encoding tower and a knowledge encoding tower. The text encoding tower is used to encode the text describing warehouse operations to generate text vectors, and the knowledge encoding tower is used to encode the knowledge in the warehouse domain to generate knowledge vectors. Collect historical business data from the warehouse management system, extract historical abnormal events and handling solutions from the historical business data, construct a training sample library, and divide the historical business data in the training sample library into a training set and a validation set according to a preset ratio. Historical business data from the training set is input into the AI ​​language model to be trained for pre-training. The model parameters are optimized by maximizing the mutual information between the text vector and the knowledge vector to obtain the pre-trained model. Construct a multi-task training objective function, which includes a loss function for identifying causal relationships, a loss function for classifying anomaly types, and a loss function for assessing risk levels; The validation set data is input into the pre-trained model, the loss value of the multi-task training objective function is calculated, the model parameters are optimized through backpropagation, and when the loss value is less than the preset convergence threshold, the trained artificial intelligence language model is output.

5. The method according to claim 1, characterized in that, Parsing and annotating text to extract causal relationships, calculating risk weight values ​​based on anomaly type labels, and constructing a risk knowledge graph by associating the risk weight values ​​with causal relationships include: Extract part-of-speech and syntactic features from the annotated text, identify event trigger words and event entities in the annotated text, determine the combination relationship between event entities, and generate an event entity sequence; Extract timestamp information from the event entity sequence, form event entity pairs by combining event entities with temporal correlation, calculate the semantic similarity of the event entity pairs, and obtain the semantic features of the event entity pairs; The causal association strength between event entity pairs is calculated based on the semantic features of the event entity pairs. Event entity pairs whose causal association strength exceeds a preset strength threshold are filtered out. The filtered event entity pairs are determined as causal relationship pairs, and causal relationship triples containing causal association strength are generated. Extract anomaly type labels from the labeled text, calculate the semantic representation of the anomaly type labels, and convert the semantic representation into risk weight values; The event entities in the causal triples are constructed as graph nodes, and the causal association strength is constructed as the weight of the graph edge to generate an initial causal graph. Risk weight values ​​are assigned to the corresponding abnormal event nodes in the initial causal graph, and risk propagation attenuation values ​​are calculated as the propagation attributes of the graph edges to generate a risk knowledge graph.

6. The method according to claim 1, characterized in that, The risk knowledge graph is input into a trained AI language model for multi-path reasoning to generate multiple candidate decision schemes. The feasibility score of each candidate decision scheme under the constraint set and its risk exposure in the risk knowledge graph are calculated. The candidate decision scheme with the lowest risk exposure and the highest feasibility score is selected as the final decision scheme. Extract node event attributes and edge risk attributes from the risk knowledge graph, construct graph feature vectors, and generate graph sequence representations by arranging them in chronological order; Input the graph sequence representation into the trained artificial intelligence language model, identify risk propagation links based on the strength of causal association, calculate the risk level of the link based on the risk propagation attenuation value, determine the critical risk path, and generate a risk avoidance path containing risk avoidance nodes. Extract risk avoidance nodes from the risk avoidance path, match the corresponding countermeasures for the risk avoidance nodes, combine them in chronological order to generate a sequence of countermeasures, and construct candidate decision schemes; Extract business constraint rules to construct a constraint set, calculate the execution cost of countermeasures based on the constraint set, and transform the execution cost into a feasibility score for candidate decision solutions; Calculate the local risk value of the response node, calculate the risk propagation value based on the risk propagation attenuation value, and combine the local risk value and the risk propagation value to form the risk exposure of the candidate decision scheme; Candidate decision-making schemes with feasibility scores greater than a preset score threshold and the lowest risk exposure are selected as the decision-making schemes.

7. The method according to claim 1, characterized in that, The decision-making scheme is simulated to propagate along the causal relationship path in the risk knowledge graph. The propagation strength is calculated based on the risk weight value. Propagation nodes with propagation strength exceeding a preset propagation threshold are identified. Blocking measures are generated for the propagation nodes. The decision-making scheme and risk blocking measures are sent to the warehouse management system for execution, including: Extract the sequence of operation instructions from the decision-making scheme, parse the sequence of operation instructions to form an event feature vector, match the event node corresponding to the event feature vector in the risk knowledge graph, and obtain the risk weight value of the event node; The event nodes are input into the risk knowledge graph for path expansion. Connected event nodes are obtained based on causal relationship paths. The propagation attenuation coefficient of the causal relationship paths is extracted to construct a propagation impact path graph. The risk weight value is used as the propagation energy value. The propagation energy value is transmitted along the causal path in the propagation influence path diagram. The propagation intensity is obtained by multiplying the propagation attenuation coefficient by the propagation energy value. The propagation intensity is compared with a preset propagation threshold, and event nodes that exceed the preset propagation threshold are selected as propagation nodes; Extract the node connection relationships between propagation nodes, obtain the risk prevention and control rules corresponding to the propagation nodes from the risk knowledge graph, and generate blocking measures based on the node connection relationships to organize the risk prevention and control rules. The decision-making schemes and blocking measures are sorted according to their propagation intensity to generate an execution queue. The execution queue is then converted into system call instructions and sent to the warehouse management system for execution.

8. An intelligent analysis system based on a WMS system and an AI large-scale model, used to implement the method of any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire business data from the warehouse management system and convert the business data into warehouse operation description text. An anomaly identification unit is used to input the warehouse operation description text into a pre-trained artificial intelligence language model to identify abnormal segments, label the abnormal segments with anomaly type labels and impact scope labels, and generate labeled text. The graph construction unit is used to parse the labeled text to extract causal relationships, calculate risk weight values ​​based on anomaly type labels, and construct a risk knowledge graph by associating the risk weight values ​​with causal relationships. The decision generation unit is used to input the risk knowledge graph into the trained artificial intelligence language model for multi-path reasoning, generate multiple candidate decision schemes, calculate the feasibility score of each candidate decision scheme under the constraint set and the risk exposure in the risk knowledge graph, and select the candidate decision scheme with the lowest feasibility score and the lowest risk exposure as the decision scheme. The risk blocking unit is used to simulate the propagation of the decision plan along the causal relationship path in the risk knowledge graph, calculate the propagation intensity based on the risk weight value, identify propagation nodes whose propagation intensity exceeds the preset propagation threshold, generate blocking measures for the propagation nodes, and send the decision plan and risk blocking measures to the warehouse management system for execution.

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

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

Images