Work order processing method, system, and program product
By using a knowledge graph for work order processing to automate the processing of multimodal work order requests, the problem of selection difficulties and inefficiency caused by the complexity of enterprise information systems has been solved, and an efficient and intelligent work order processing workflow has been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU ROBAM APPLIANCES CO LTD
- Filing Date
- 2026-01-13
- Publication Date
- 2026-05-29
AI Technical Summary
The complexity of enterprise information systems makes the work order selection process difficult, resulting in a high error rate for manual selection, long processing cycles, low efficiency in form filling, and long time for new employees to become familiar with the process. Existing systems lack intelligent interaction and adaptive learning mechanisms.
Based on a pre-built knowledge graph of work order processing, the system automatically identifies the target business system and functional nodes through semantic recognition of multimodal work order requests, and sequentially calls the business nodes to process the work order requests, thereby achieving an automated process.
It improved work order processing efficiency, reduced human error, shortened processing cycles, reduced knowledge dependence, and increased the onboarding speed for new employees and system adaptability.
Smart Images
Figure CN122114582A_ABST
Abstract
Description
Technical Field
[0001] This application generally relates to the field of artificial intelligence technology, and in particular to a work order processing method, system, and program product. Background Technology
[0002] Currently, enterprise information systems are complex, so when employees fill out work orders, they need to manually determine which system the problem belongs to. Different systems cover multiple processes, which makes the work order selection process difficult. In addition, most forms have redundant high-frequency fields, resulting in low form filling efficiency.
[0003] Secondly, if the initial selection of a work order is incorrect, it needs to be reassigned, which also leads to a longer processing cycle. In addition, the existing work order system has a serious knowledge dependency problem. New employees need a long time to become familiar with some of the process operations. If key personnel leave, the processing efficiency of certain processes will decrease. Summary of the Invention
[0004] In view of the above-mentioned defects or deficiencies in the existing technology, it is desirable to provide a work order processing method, system and program product. The method determines the target functional node that matches the multimodal work order request based on a pre-created work order processing knowledge graph, so as to sequentially call each business node under the target functional node to process the multimodal work order request, thereby realizing the automated processing of multimodal work order requests and improving the processing efficiency of work order requests.
[0005] In a first aspect, the present invention provides a work order processing method, the method comprising: Receive multimodal work order requests, perform semantic recognition on the multimodal work order requests, and determine the target business system for processing the multimodal work order requests and the intent of the multimodal work order requests; Based on the work order processing knowledge graph, multiple candidate functional nodes corresponding to the target business system are identified, and the target functional node that matches the intent of the multimodal work order request is determined from the multiple candidate functional nodes; wherein, the multiple candidate functional nodes are slave nodes of the system node corresponding to the target business system in the work order processing knowledge graph; Based on the flow order of each business node under the target function node in the work order processing knowledge graph, each business node is called sequentially to process multimodal work order requests; business nodes include: business processing nodes and / or data nodes.
[0006] In one possible implementation, when the intent of a multimodal work order request matches multiple target functional nodes, based on the flow order of each business node under the target functional node in the work order processing knowledge graph, each business node is called sequentially, including: For each target functional node, the confidence level of the target functional node is determined based on the degree of matching between the description information of the target functional node and the intent. Based on the flow order of business nodes under the target function node with the highest confidence in the work order knowledge graph, the business nodes under the target function node with the highest confidence are called sequentially.
[0007] In one possible implementation, the method also includes: When the confidence level of the target functional node with the highest confidence level is less than the preset threshold, supplementary parameters for the multimodal work order request are generated based on the processing requirements of the business nodes under the target functional node with the highest confidence level. Prompt the user to ask for confirmation; the confirmation message is used to prompt the user to enter additional parameters.
[0008] In one possible implementation, the creation process of the work order processing knowledge graph includes: Based on the historical logs generated by multiple business systems when processing multiple historical multimodal work order requests, multiple functional nodes associated with each business system are identified. For each functional node, at least one business node is determined based on historical logs when the functional node processes the corresponding historical multimodal work order request, and the flow order of business nodes under the functional node is formed based on the calling order of at least one business node; wherein, business nodes include business processing nodes and / or data nodes. Create a work order processing knowledge graph based on the flow order of business nodes under multiple functional nodes.
[0009] In one possible implementation, semantic recognition is performed on the multimodal work order request to determine the target business system for processing the multimodal work order request, including: Obtain the health assessment score of each business system, the semantic matching score between multimodal work order requests and each business system, and the attribute information of multimodal work order requests and each business system; Based on the weighted results of health assessment, semantic matching, and attribute information, the target business system for processing multimodal work order requests is determined.
[0010] In one possible implementation, the method also includes: For each historical multimodal work order request, determine the actual functional node that will ultimately process the historical multimodal work order request; Based on the number of all historical multimodal work order requests and the number of work order requests in all historical multimodal work order requests whose actual functional nodes differ from the target functional nodes determined based on the work order knowledge graph, update the weight values of health assessment degree, semantic matching degree, and attribute information.
[0011] In one possible implementation, the method also includes: When a business node under the target function node experiences an abnormal operation, update the flow order of the business nodes under the target function node.
[0012] Secondly, a work order processing system is provided, which includes a work order processing platform and multiple business systems that interface with the work order processing platform. The work order processing platform is used to receive multimodal work order requests, perform semantic recognition on the multimodal work order requests, and determine the target business system and the intent of the multimodal work order requests from multiple business systems; based on the work order processing knowledge graph, it determines multiple candidate functional nodes corresponding to the target business system, and selects the target functional node that matches the intent of the multimodal work order request from the multiple candidate functional nodes; wherein, the multiple candidate functional nodes are slave nodes of the system nodes corresponding to the target business system in the work order processing knowledge graph; The work order processing platform is also used to process multimodal work order requests by calling business nodes under the target function nodes associated with the target business system, based on the work order processing knowledge graph.
[0013] In one possible implementation, the work order processing platform includes a multimodal input module. The multimodal input module is used to receive multimodal work order requests in at least one of the following forms: voice, text, and image. The multimodal input module is also used to obtain the input features of the multimodal work order request based on the input format of the multimodal work order request, and to fuse the input features according to the text association characteristics.
[0014] Thirdly, a computer program product is provided, which includes instructions that, when executed, perform the method described in any one of the first aspects.
[0015] Fourthly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the method described in any one of the first aspects above.
[0016] Fifthly, a computer-readable storage medium is provided having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method described in any one of the first aspects above.
[0017] Compared to existing technologies where manual judgment in work order processing systems and the heavy knowledge dependency of these systems lead to low efficiency in processing various work orders, the work order processing method, system, and program products provided in this application, on the one hand, can automatically determine the target business system for processing multimodal work order requests based on the semantic recognition results of the requests, thereby shortening the time spent determining the system to which the multimodal work order requests belong. On the other hand, based on a pre-built work order processing knowledge graph, the target functional nodes that match the multimodal work order requests are further identified among the functional nodes associated with the target business system. The business nodes under the target functional nodes are then sequentially called to process the multimodal work order requests, thereby achieving an automated processing flow for multimodal work order requests and improving processing efficiency. Furthermore, the construction of the work order processing knowledge graph also avoids the knowledge dependency problem of the work order system, facilitating work order processing operations for different users.
[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0019] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the architecture of the work order processing system 10 provided in the embodiments of this application; Figure 2 This is a schematic diagram of the front-end software interface of the work order processing platform 11 provided in this application embodiment; Figure 3 This is a flowchart illustrating a work order processing method provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of the work order processing knowledge graph provided in the embodiments of this application; Figure 5 This is a schematic diagram illustrating a call to a business node under a target function node provided in an embodiment of this application; Figure 6 This is a schematic diagram of the module distribution of the work order processing platform 11 provided in this application embodiment; Figure 7 This is a schematic diagram of the update based on the adaptive learning module 114 provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of a computer device according to an embodiment of this application. Detailed Implementation
[0020] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.
[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present application will now be described in detail with reference to the accompanying drawings and embodiments. Furthermore, the term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The terms "first" and "second," etc., in the specification and claims of the embodiments of this application are used to distinguish different objects, not to describe a specific order of objects.
[0022] With the continuous development of information technology, enterprises face highly complex information challenges. Currently, there are more than 50 independent enterprise information systems deployed, including OA, BI, PLM, ERP, MES, SRM, EHR, contract management, CRM, electronic archives, and travel systems. These independent systems have close business connections but lack intelligent interaction channels.
[0023] Therefore, when employees fill out work order requests, they need to manually determine the system to which the work order belongs. Each information system corresponds to a large number of different types of processing procedures (for example, the OA system covers more than 1,000 procedures such as contracts, procurement, and expenses, while the EHR system includes nearly 100 procedures such as leave and attendance). This makes it difficult to manually select the procedure. Moreover, if the system to which the work order belongs is initially selected incorrectly, it needs to be transferred to other information systems for processing, which leads to a longer processing cycle for work order requests.
[0024] Secondly, the high-frequency fields in most existing forms are redundant, resulting in low form filling efficiency; and the application of various systems has serious knowledge dependence problems, which means that new employees need a long time to become familiar with some process operations, and if employees in key positions leave, the processing efficiency of specific processes will also decline accordingly.
[0025] To address the aforementioned issues, the mainstream solutions for enterprise IT work order processing include work order routing systems based on rule engines and work order recommendation systems based on simple natural language processing (NLP).
[0026] However, the aforementioned systems still suffer from several problems: insufficient semantic understanding (i.e., inability to distinguish the actual differences between similar terms, for example, "contract approval" may involve OA legal processes or CRM customer contract modules), lack of system topology awareness (i.e., the rule engine cannot dynamically identify the dependencies between different information systems), lack of adaptive learning mechanisms (i.e., when a new system is added or system policies change, rules need to be manually modified), and lack of multimodal interaction (i.e., voice-based work orders need to be manually transcribed and forms filled out).
[0027] Based on this, this application provides an intelligent work order processing system. This system can receive work order requests through a multimodal fusion input interface and automatically match and call the corresponding business nodes to process the work order requests based on a pre-built work order processing knowledge graph, thereby improving the efficiency of work order processing and thus improving the overall business operation efficiency of the enterprise.
[0028] In one possible implementation, Figure 1 This is a schematic diagram of the architecture of the work order processing system 10 provided in the embodiments of this application, as shown below. Figure 1 As shown, the work order processing system 10 may include a work order processing platform 11 and multiple business systems 12 that are connected to the work order processing platform 11; wherein, each business system 12 corresponds to the independent information systems already deployed by the enterprise, such as OA, BI, PLM, ERP, MES, SRM, EHR, contract management, CRM, electronic archives, travel system and other business systems.
[0029] In the specific implementation, refer to Figure 1 The work order processing platform 11 can first determine the target business system for processing the multimodal work order request based on the semantic recognition result of the received multimodal work order request, and then determine the target functional node of the target business system that matches the intent of the multimodal work order request, thereby calling the various business nodes under the functional node to process the multimodal work order request.
[0030] It should be noted that the intent of a multimodal work order request can be understood as the business event content that the multimodal work order request wants to implement, and each functional node corresponding to business system 12 can be understood as a slave node of business system 12 when implementing a certain business function; wherein, the functional node corresponding to business system 12 can be determined based on a pre-built work order processing knowledge graph.
[0031] For example, the work order processing platform 11 can obtain the corresponding form of work order request based on the user's operation type on the front-end software interface, wherein the user operation type specifically includes text input operation, voice input operation and image upload operation.
[0032] Correspondingly, the work order processing platform 11 includes a multimodal input module 111, which is used to receive work order requests in voice, text, or image formats according to the user's operation type.
[0033] For example, Figure 2 This is a schematic diagram of the front-end software interface of the work order processing platform 11 provided in this application embodiment, such as... Figure 2 As shown, the front-end software interface includes a text input area 21, an image upload area 22, and a voice input control 23, enabling users to perform text input operations in the text input area 21, image upload operations in the image upload area 22, and voice input operations by clicking the voice input control 23.
[0034] For example, after receiving a work order request from a user, the multimodal input module 111 can obtain the input features of the work order request based on the input format of the work order request.
[0035] In one example, the multimodal input module 111 integrates a multilingual speech input model, Whisper-large-v3, to receive work order requests in speech form and obtain the corresponding voiceprint features.
[0036] Specifically, the multimodal input module 111 can also apply a noise suppression algorithm (e.g., a real-time noise suppression algorithm based on deep learning, RNNoise algorithm) to filter environmental noise (e.g., background noise) in the voice work order request in real time to improve the accuracy of speech recognition.
[0037] For example, the multimodal input module 111 can receive and recognize voice work order requests through the following code: # Lightweight domain adaptation fine-tuning of the Whisper-large-v3 model using LoRA (Low-Rank Adaptation) technique #Objective: To optimize speech recognition accuracy in industrial environments while avoiding the high computational cost of full-parameter fine-tuning. #Configure LoRA fine-tuning parameters and define adapter structure peft_config = LoraConfig( task_type="SPEECH_RECOGNITION", # Specifies the task type as speech recognition r = 8, # Sets the rank of the low-rank matrix, controlling the number of adaptation parameters (balancing effect and computational efficiency). target_modules=["q_proj", "v_proj"] # Specifies the model modules to be adapted (query and value projection matrix) ) # Load the pre-trained Whisper-large-v3 base model (OpenAI open-source model) This model performs well in general speech recognition tasks, but needs optimization for industrial terminology and ambient noise. model=WhisperForConditionalGeneration.from_pretrained("openai / whisper- large - v3") # Convert the original model to a PEFT (Parameter-Efficient Fine-Tuning) model #Inject trainable parameters only into the modules specified in target_modules, while keeping most of the original model's parameters frozen. model = get_peft_model(model, peft_config) In another example, the multimodal input module 111 also integrates a multi-level word segmenter based on business domains, which is used to identify the exclusive terms of each business system in the voice work order request (e.g., PLM material code change, BOM maintenance) and obtain the corresponding text semantic features.
[0038] Specifically, the multi-layer word segmenter includes a base layer, a domain layer, and a dynamic disambiguation layer. The base layer performs word segmentation based on the general scenario of whole word masking (BERT-wwm) to identify the specific terms of different business systems. The domain layer is based on the injection of system-specific dictionaries (e.g., PLM dictionary: material code -> / PLM / part_code). The dynamic disambiguation layer distinguishes homonyms through a contextual attention mechanism to perform contextual segmentation of multiple content segments.
[0039] In another example, the multimodal input module 111 also integrates a multimodal pre-trained model LayoutLMv3 for document classes, so as to receive work order requests in the form of images and obtain the corresponding visual features.
[0040] Specifically, the aforementioned multimodal pre-trained model can process scanned documents or screenshots input by the user to extract key fields (e.g., contract number, device serial number) from the scanned documents or screenshots.
[0041] For example, after the multimodal input module 111 obtains the input features of the work order request in the above manner, it can fuse the input features according to the text association characteristics to form the feature vector corresponding to the work order request.
[0042] Specifically, the multimodal input module 111 can concatenate the obtained multidimensional features according to the context features to construct the feature vector corresponding to the work order request; among them, the multidimensional features may also include environmental features, IoT device status and other features.
[0043] In this embodiment, by extracting and fusing features of multimodal work order requests through the multimodal input module 111, unified normalization processing of heterogeneous inputs of the work order processing platform 11 can be achieved.
[0044] Figure 3 This is a flowchart illustrating a work order processing method provided in an embodiment of this application. This work order processing method can be implemented based on the aforementioned work order processing system 10, such as... Figure 3 As shown, the method includes the following steps: Step S301: Receive a multimodal work order request, perform semantic recognition on the multimodal work order request, and determine the target business system for processing the multimodal work order request and the intent of the multimodal work order request.
[0045] Compared to the existing technology that manually selects the business system corresponding to a work order request, this embodiment automatically determines the target business system for processing the multimodal work order request based on the semantic recognition result of the multimodal work order request (i.e., the business system that processes the multimodal work order request is, for example, an OA system or a travel system), so as to reduce the error rate that exists when manually selecting a business system and shorten the time spent selecting the target business system.
[0046] It should be noted that, based on the different forms of work order requests received by the multimodal input module 111, the multimodal work order requests that undergo semantic recognition in this embodiment can be understood as feature vectors after feature extraction and fusion by the multimodal input module 111. That is, the multimodal work order requests here are the work order request results after the multimodal input module 111 performs unified normalization processing on heterogeneous inputs.
[0047] For example, when a user inputs a work order request in text form on the aforementioned front-end software interface, the multimodal work order request with semantic recognition here refers to the work order request result after fusing the text semantic features extracted by the multimodal input module 111 with environmental features, IoT device status, and other features; when a user inputs a work order request in both voice and text form on the aforementioned front-end software interface, the multimodal work order request with semantic recognition here refers to the work order request result after fusing the voiceprint features, text semantic features, environmental features, IoT device status, and other features extracted by the multimodal input module 111.
[0048] In one possible implementation, the target business system for processing the multimodal work order request can be determined among multiple business systems 12 based on the semantic matching degree between the multimodal work order request and each business system 12.
[0049] For example, the target business system can be determined based on the weighted results of three dimensions: the health assessment degree of each business system 12, the semantic matching degree between the multimodal work order request and each business system 12, and the attribute information of the multimodal work order request (i.e., the weight values of the three dimensions are not the same); for example, the business system 12 with the highest weighted result corresponds to the target business system.
[0050] In one example, the health assessment of each business system 12 can be determined based on the API running status and data timeliness status of each business system 12.
[0051] For example, the following code can be used to filter out normal business systems from multiple business systems 12 whose API running status is normal: MATCH(t:Ticket{id:$ticket_id})-[r]->(s:System) WHEREr.api_status="healthy"; # Filter business systems whose API status is normal. Subsequently, based on the weighted result of the API operation status and data timeliness status of each normal business system, the health assessment degree of each normal business system is determined. Specifically, the corresponding health assessment degree can be obtained through the following code: RETURNs.name, 0.7 * api_sla_score + 0.3 * data_freshness AS kg_score; # The health assessment score is the sum of the API running status with a weight of 0.7 and the data freshness status with a weight of 0.3. In another example, the semantic matching degree between multimodal work order requests and each business system 12 can be determined based on the functional category of the business processed by each business system 12.
[0052] For example, when a multimodal work order request is "PDA cannot scan location A12-3", the following code can be used to obtain three business systems 12 with a high semantic matching degree to this multimodal work order request and their corresponding semantic matching values: { "input":"PDA cannot scan storage location A12-3", "top3_candidates":[ {"system":"MES","score":0.91}, {"system":"SRM","score":0.87}, {"system":"IT Operations and Maintenance","score":0.32}]} In another example, the attribute information of a multimodal work order request may include the work order request type, the work order request time period, the target system CPU utilization, and the department information of the user (i.e., the initiator) who initiated the work order request; the attribute information of each business system 12 may include the first success rate, the first failure rate, and the timeout status when each business system 12 processes historical multimodal work order requests.
[0053] Based on this, the weighted result of each business system 12 can be calculated using the following formula: alpha*nlp_model.predict(ticket,system)+beta*kg_query(system,ticket.context)+gamma*rl_agent.get_reward(system,ticket) Among them, nlp_model.predict(ticket,system) is used to represent the semantic matching degree between the business system and the multimodal work order request, alpha is used to represent the weight value of the semantic matching degree, kg_query(system,ticket.context) is used to represent the health assessment degree of the business system, beta is used to represent the weight value of the health assessment degree, rl_agent.get_reward(system,ticket) is used to represent the attribute information of the multimodal work order request, and gamma is used to represent the weight value of the attribute information.
[0054] The embodiments of this application adopt a hybrid decision-making model architecture. By using the weighted scoring of the three dimensions of semantic matching degree, health assessment degree and attribute information, the target business system for processing multimodal work order requests can be accurately judged.
[0055] Step S302: Based on the work order processing knowledge graph, determine multiple candidate functional nodes corresponding to the target business system, and determine the target functional node that matches the intent of the multimodal work order request from the multiple candidate functional nodes.
[0056] For example, the work order processing knowledge graph contains multiple business system nodes (i.e., each business system node corresponds to a business system), and each business system node corresponds to at least one functional node (i.e., each functional node corresponds to the node that can perform the function of the upper-level business system); based on this, the above-mentioned candidate functional nodes are the slave nodes of the system node corresponding to the target business system in the work order processing knowledge graph.
[0057] Correspondingly, Figure 4 This is a schematic diagram of the structure of the work order processing knowledge graph provided in the embodiments of this application, such as... Figure 4 As shown, the main node of the work order processing knowledge graph can correspond to the aforementioned work order processing platform 11. The first-level slave nodes under the main node can be the corresponding nodes of multiple business systems that are connected to the work order processing platform 11. The second-level slave nodes under the first-level slave nodes can be the functional nodes that each business system can implement. That is, each functional node corresponds to a function that the upper-level business system can implement.
[0058] Correspondingly, the slave nodes under each functional node are specifically at least one business node that needs to be called when implementing the corresponding function of that functional node. The business node may include the business processing node and / or data node to be called.
[0059] For example, based on the above-mentioned work order processing knowledge graph, the cosine similarity between the description information of each candidate functional node (e.g., text information used to describe the function) and the intent of the multimodal work order can be determined according to the semantic matching degree between the description information of each candidate functional node and the intent of the multimodal work order request, so as to determine the target functional node of the multimodal work order request based on the cosine similarity matrix.
[0060] For example, the cosine similarity between the description information of each candidate function node and the intent of the multimodal work order request can be determined using the following code: def forward(self, text, system_desc): # Intent Tower Forward Propagation: Processing Multimodal Work Order Requests from Input # text: The text content of the work order request (e.g., "Production equipment malfunctions and needs repair") # Output the hidden state (aggregated semantic vector) at index [1] to obtain the hidden state at position [CLS]. intent_embed = self.intent_tower(text)[1] # System Tower Forward Propagation: Descriptive Information of Candidate Functional Nodes # system_desc: Function description text (e.g., "The MES system is responsible for production execution management") system_embed = self.system_tower(system_desc)[1] # Calculate the cosine similarity between the intent vector and the system vector # Return value range [-1, 1], the higher the value, the better the work order intent matches the system function. return cosine_similarity(intent_embed, system_embed) Correspondingly, Figure 5 This is a schematic diagram illustrating a call to a business node under a target function node provided in an embodiment of this application, such as... Figure 5 As shown, regarding the determination of the target functional node, firstly, the semantic intent of the multimodal work order request input by the user is identified using an intent coding tower (e.g., a sequence classification model based on Chinese BERT). At the same time, the descriptive information of each candidate functional node is summarized using a system coding tower (e.g., a custom-configured BERT model) to determine the cosine similarity between the descriptive information of each candidate functional node and the intent of the multimodal work order request, thereby constructing a similarity matrix. Then, the target functional node of the multimodal work order request is determined based on the similarity matrix.
[0061] Specifically, the semantic intent recognition of the aforementioned intent encoding tower and the information aggregation of the system encoding tower can be implemented using the following code: class DualTowerModel(nn.Module): def __init__(self): super(DualTowerModel, self).__init__() # Intent Coding Tower: A Sequence Classification Model Based on Chinese BERT # Used to extract semantic intent features of work order requests (such as "reimbursement application", "equipment repair request", etc.) self.intent_tower = BertForSequenceClassification .from_pretrained("bert-base-chinese") # System Functional Encoding Tower: Using a custom-configured BERT model # SystemAwareConfig contains domain-specific parameters to adapt enterprise system function description text. # Used for coding functional descriptions and interface documents for various systems (OA, ERP, MES, etc.) self.system_tower=BertModel(config=SystemAwareConfig()) This application's embodiments employ a dual-tower model structure to construct a hierarchical semantic understanding engine. Specifically, it utilizes a system encoding tower that focuses on encoding the functional characteristics of each system to establish a system function vector representation, and an intent encoding tower that focuses on understanding the semantic intent of the work order content to identify the user's real needs, thereby achieving joint recognition of intent and system.
[0062] Step S303: Based on the flow order of each business node under the target function node in the work order processing knowledge graph, call each business node in sequence to process the multimodal work order request.
[0063] For example, when the target functional node corresponding to the multimodal work order request is determined, the work order processing platform 11 can, based on... Figure 4 The work order processing knowledge graph shown sequentially calls at least one business node corresponding to the target function node under the target business system to process multimodal work order requests.
[0064] For example, during the actual invocation process of each business node under the above target function node, if a business node malfunctions, the flow order of each business node under the target function node is updated; for example, the malfunctioning business node can be replaced with another business node that can satisfy the current invocation relationship, or the weight value (i.e., routing weight) of the flow order of each business node under the target function node can be adjusted.
[0065] In another embodiment of this application, a specific implementation method for invoking the business node under the target function node is also provided.
[0066] In one possible implementation, when the intent of a multimodal work order request matches multiple target function nodes among multiple candidate function nodes, the confidence level of each target function node can be further determined based on the matching degree between the description information of each target function node and the intent. Then, based on the flow order of the business nodes under the target function node with the highest confidence level in the work order knowledge graph, the business nodes under the target function node with the highest confidence level are called sequentially.
[0067] For example, refer to Figure 5When the cosine similarity in the similarity matrix is sorted in descending order, the candidate functional nodes corresponding to the top three cosine similarities in the descending order can be determined as the target functional nodes. Based on the confidence levels of the three target functional nodes, the business node under the target functional node with the highest confidence level is called.
[0068] In one possible implementation, different invocation methods for business nodes under the target functional node can be formed based on the confidence level corresponding to the target functional node with the highest confidence level.
[0069] For example, when the confidence level of the target functional node with the highest confidence level is greater than the first preset threshold, the business nodes under the target functional node are called sequentially based on the work order processing knowledge graph to generate a direct route; for example, the first preset threshold is 0.8.
[0070] For example, when a user voice-inputs a multimodal work order requesting "paternity leave from next Wednesday to Friday," the target functional node of this multimodal work order request can be determined to be the leave request functional node. If the confidence level of this leave request functional node is greater than 0.8, the following field mapping can be directly generated to complete the filling of the smart form: "EHR Leave Application Form": { "Leave Type": "Paternity Leave" "Start Date": "2025-07-23", "End Date": "2025-07-25", "Supporting Documents": "Automatically Linked Electronic Marriage Certificate File"} Specifically, a text parsing library (e.g., the Duckling library for parsing structured information such as dates, times, numbers, and durations) can be used to convert the time information in the multimodal work order request into a specific date (i.e., convert "next Wednesday to Friday" into a specific date). Then, based on the user's personal information (e.g., user ID), detailed information about the user (e.g., marital status, marriage certificate attachments, etc.) can be queried using SQL.
[0071] For example, when the confidence level of the target functional node with the highest confidence level is less than the first preset threshold but greater than the second preset threshold, supplementary parameters for the multimodal work order request can be generated based on the processing requirements of the business node under the target functional node with the highest confidence level, so as to prompt the user to ask for confirmation information. This follow-up confirmation information is used to prompt the user to input the corresponding supplementary parameters; for example, the second preset threshold is 0.5.
[0072] For example, when a user inputs a multimodal work order request "take leave next Monday" via voice, the target functional node of the multimodal work order request can be determined to be the leave request functional node. If the confidence level of the leave request functional node is greater than 0.5 and less than 0.8, the supplementary parameters of the multimodal work order request can be determined as leave type and leave duration based on the processing requirements of the business nodes under the leave request functional node.
[0073] Based on this, the system can prompt the user with follow-up confirmation information such as "Is this annual leave, sick leave, or some other type?" If the user enters a supplementary parameter indicating sick leave, the system can continue to prompt the user with follow-up confirmation information such as "Additional requirement: 'Please upload a medical certificate'". If the user enters a supplementary parameter indicating annual leave, the system can call the corresponding business node (e.g., the annual leave confirmation node) to automatically check the remaining annual leave days and prompt the user.
[0074] In another embodiment of this application, a specific method for creating a work order processing knowledge graph is also provided.
[0075] In one possible implementation, multiple functional nodes associated with each business system can be identified based on historical logs generated by multiple business systems when processing multiple historical multimodal work order requests. Then, for each functional node, at least one business node applied by the functional node when processing the corresponding historical multimodal work order request is determined according to the historical logs. The flow order of business nodes (i.e., including business processing nodes and / or data nodes) under the functional node is formed based on the calling order of at least one business node. Finally, a work order processing knowledge graph is created based on the flow order of business nodes under multiple functional nodes.
[0076] For example, for each functional node associated with each business system, historical log analysis via API calls or distributed tracing tools such as OpenTelemetry can be used to perform distributed tracing on historical logs to discover and identify the call chain between business nodes when processing a certain historical multimodal work order request (for example, the OA expense settlement process can trigger the ERP expense invoice data generation node, etc.), thereby forming the flow order of business nodes under each functional node.
[0077] For example, after obtaining the flow order of business nodes under each functional node, data standardization processing can be performed on each flow order to create a work order processing knowledge graph based on the standardized flow order; for example, a certain functional node in the work order processing knowledge graph can correspond to {"system": "OA", "interfaces": ["soap_createFlowApi,..."], "dependency": ["ERP->Expense Warehouse", "SRM->Purchasing Material Requirements"...]...}.
[0078] Optionally, after obtaining the flow order of business nodes under each functional node, the flow order can be supplemented with corresponding rules through manual inspection.
[0079] In one possible implementation, monitoring probes can be deployed at different business nodes based on the work order processing knowledge graph constructed above, so as to obtain the running status of each business node in real time. In this way, when a business node with abnormal operation occurs, the weight value of the flow order of the business nodes under the corresponding functional node in the graph can be updated in a timely manner (i.e., the routing weight).
[0080] For example, a Prometheus probe (specifically used to proactively detect the health status and availability of the target service) can be deployed to collect the API error rate of each business node in real time; or a JDBC probe (specifically used to monitor database connections and performance) can be deployed to collect the utilization rate of the database connection pool of each business node.
[0081] Specifically, the aforementioned route weights can be determined based on the priority score of the route. This priority score corresponds to the product of the semantic matching score (base_score) of the description information of each functional node and the corresponding historical multimodal work order request, the health_penalty of the business node corresponding to each functional node, and the dependency_criticality coefficient of the business system corresponding to the functional node.
[0082] For example, the priority score for each route can be calculated using the following code: def calculate_priority(system): # Basic semantic matching score (range 0 - 1), calculated based on NLP similarity between work order content and system function. # A weight of 0.6 indicates that semantic matching plays a dominant but not absolute role in routing decisions. base_score = nlp_similarity * 0.6 # Calculation of system health penalty factor (range 0 - 1, the lower the value, the worse the system health). # system.cpu_usage / 70: Penalties are incurred when CPU utilization exceeds 70% (normalized to 0 - 1). # system.error_rate*10: A significant penalty is incurred for every 10% increase in the error rate (amplifying the impact of the error rate). # min(1, ...): Ensures the penalty factor does not exceed 1, avoiding over-penalty. health_penalty = 1 - min(1, system.cpu_usage / 70 + system.error_rate* 10) # Dependency criticality coefficient (≥1), weights set based on the importance of system services in the topology graph. # For core systems (such as ERP), this value > 1; for peripheral systems, = 1, ensuring that critical business processes are prioritized. dependency_criticality = get_dependency_criticality(system.id) #Final Priority Score = Semantic Matching Score × Health Factor × Keyness Coefficient # Reflecting multi-dimensional decision-making: business relevance, system availability, and business importance return base_score * health_penalty * dependency_criticality In another embodiment of this application, a specific method for updating the weight values of the aforementioned health assessment degree, semantic matching degree, and attribute information is also provided.
[0083] In one possible implementation, for each historical multimodal work order request, the actual functional node that will ultimately process the historical multimodal work order request is first determined; then, based on the number of all historical multimodal work order requests and the number of work order requests in all historical multimodal work order requests whose actual functional nodes differ from the target functional node determined based on the work order knowledge graph, the weight values of the health assessment degree, semantic matching degree, and attribute information are updated.
[0084] In this embodiment, based on the number of work orders that are secondarily dispatched (i.e., the actual functional node corresponding to the final processing of the historical multimodal work order request is different from the target functional node determined based on the work order knowledge graph), the weight values of each business system to which the work order request belongs are updated in a timely manner, so as to achieve dynamic routing balance while avoiding rigidity.
[0085] For example, the adjustment range for each weight value can be calculated using the following formula:
[0086] Where α represents each weight value (e.g., the initial weight value for semantic matching is 0.6); η represents the learning rate used to control the adjustment step size (e.g., 0.01); reroute_count represents the number of work order requests that require secondary reassignment; total_tickets represents the total number of work orders; reroute_count / total_tickets represents the optimization objective: minimizing the secondary reassignment rate, i.e., (1-reroute_count / total_tickets) corresponds to maximizing the accuracy of the first dispatch; gradient term / α is used to characterize the direction of the effect of changes in α on the objective function. That is, if increasing α can reduce the transfer rate, then the gradient is positive and α increases; otherwise, α decreases.
[0087] In another embodiment of this application, the distribution of processing modules of the work order processing platform 11 is also provided.
[0088] In one possible implementation, Figure 6 This is a schematic diagram of the module distribution of the work order processing platform 11 provided in this application embodiment, as shown below. Figure 6 As shown, the work order processing platform 11 also includes a dynamic system topology map module 112, a multi-strategy routing decision module 113, and an adaptive learning module 114. The multimodal input module 111 specifically includes an input submodule 1111 and a context-aware parsing engine submodule 1112.
[0089] For example, the dynamic system topology graph module 112 is used to construct the above-mentioned work order processing knowledge graph; the multi-strategy routing decision module 113 is used to determine the target business system and target functional node of the multimodal work order request; and the adaptive learning module 114 is used to dynamically optimize the work order routing strategy.
[0090] Specifically, the adaptive learning module 114 enables the work order processing system 10 to continuously improve in four dimensions: accuracy, timeliness, adaptability, and user experience through continuous feedback and incremental learning, while avoiding the situation where existing experience is forgotten due to new knowledge.
[0091] In one example, Table 1 is a four-dimensional evaluation matrix of the adaptive learning module 114 provided in an embodiment of this application, for feedback-driven optimization based on the four-dimensional evaluation matrix: Table 1. Four-dimensional evaluation matrix of adaptive learning module 114
[0092] Correspondingly, the initialization operation of the above four-dimensional evaluation can be implemented using the following code: def __init__(self): # Initialize a four-dimensional evaluation indicator system, corresponding to the four dimensions of the solution's objectives. self.metrics = { 'accuracy': {# Accuracy dimension: measures the correctness of routing} 'reroute_count': 0, # Number of times the work order is rerouted (count of incorrect routes) 'total': 0# Total number of work orders processed}, 'timeliness': {# Timeliness dimension: measures SLA compliance} 'sla_violations': [], # Stores a list of ticket IDs that violate the SLA. 'target': '95%' # Target SLA compliance rate (95% of work orders completed on time)}, 'adaptability': {# Adaptability dimension: measures the ability to learn new processes} 'new_flow_learning_time': timedelta() # Time required to learn the new business process}, 'ux': {# User Experience Dimension: Measuring the Degree of Automation 'auto_fill_completeness': 0 # Form autofill completeness (0 - 100%) The update operation for the above four-dimensional evaluation can be implemented using the following code: #Real-time updated metrics: Processing feedback data after each work order is completed. def update(self, ticket): #Accuracy Statistics: If a work order has a rerouting record, it means the initial route was incorrect. if ticket reroute times>0: self.metrics['accuracy']['reroute_count']+= 1 self.metrics['accuracy']['total'] += 1# Total number of work orders # Timeliness Statistics: Check whether the work order processing time exceeds the SLA-agreed time. if ticket.duration>ticket.sla: self.metrics['timeliness']['sla_violations'].append(ticket.id) # Note: In actual implementations, update logic for adaptability and UX dimensions should also be included. # For example: recording the learning time for new processes, the success rate of form autofill, etc. The optimization operations for the above four-dimensional evaluation can be implemented using the following code: # Daily Optimization Suggestions: Generating specific system improvement suggestions based on indicator data. def generate_advice (self): advice = [] # Accuracy optimization suggestion: When the rerouting rate exceeds 10%, sample learning needs to be enhanced. if self.reroute_rate()>0.1: advice.append("Add OA-EHR process comparison sample") # Target common error routing scenarios # Timeliness Optimization Suggestion: The reward mechanism needs to be adjusted if the SLA violation rate exceeds 5%. if self.sla_violation_rate()>0.05: advice.append("Adjusting the RL reward function for production work orders") # Optimizing the reinforcement learning reward function # Optimization suggestion generation logic can be expanded to include other dimensions. return advice In another example, regarding the incremental training mechanism, the adaptive learning module 114 can employ Elastic Weight Fixation (EWC) to prevent catastrophic forgetting; while in terms of data acquisition, key samples are replayed, high-value historical work orders (e.g., high-frequency or complex cases) are stored, and replays are mixed in new training for active learning, while low-confidence work orders are manually labeled, and samples with more information are prioritized for learning.
[0093] Correspondingly, Figure 7 This is a schematic diagram of the update based on the adaptive learning module 114 provided in an embodiment of this application, as shown below. Figure 7 As shown, the work order closed loop is analyzed based on the evaluation matrix. When the accuracy of the analysis results decreases, a sample retrieval operation is triggered, and EWC constraint training is performed. When the analysis results fail to meet the SLA, the RL reward is adjusted and the policy gradient is updated. Then, the model is used for verification. If the verification is successful, it is deployed to the production environment; otherwise, it is rolled back and manually intervened.
[0094] The intelligent work order processing system 10 based on multimodal perception and dynamic system topology provided in this application embodiment, by integrating natural language processing (i.e., NLP), knowledge graph reasoning, and reinforcement learning techniques, can significantly improve the efficiency and intelligence level of work order processing in complex information environments of enterprises; its beneficial effects specifically include: First, through multimodal semantic understanding and dynamic routing decisions, the system improves the accuracy of initial order dispatch while shortening the processing cycle. Furthermore, the high auto-fill rate of form fields reduces manual input time, significantly reducing operational burden and improving work order processing efficiency. Second, by constructing a work order processing knowledge graph, the system can analyze API call chains and data flows in real time, automatically identifying business relationships between business systems. When anomalies in the call interface are detected (e.g., ERP-MES synchronization delays), the system can dynamically adjust routing paths to avoid process interruptions due to single points of failure. Third, the system achieves adaptive and continuous evolution through a four-dimensional evaluation matrix (i.e., accuracy, timeliness, adaptability, and user experience) and incremental training techniques (e.g., EWC anti-forgetting). Fourth, through multimodal interaction (voice, image, text) and intelligent guidance, the system reduces knowledge dependence and training costs. It can also automatically associate historical work order requests with solutions to form a knowledge base, thereby ensuring the processing efficiency of specific processes.
[0095] The following is for reference. Figure 8 , Figure 8 A schematic diagram of a computer device suitable for implementing embodiments of this application is shown, such as... Figure 8 As shown, the computer system 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 802 or programs loaded from storage section 809 into random access memory (RAM) 803. RAM 803 also stores various programs and data required for the system's operating instructions. CPU 801, ROM 802, and RAM 803 are interconnected via bus 804. Input / output (I / O) interface 805 is also connected to bus 804.
[0096] The following components are connected to I / O interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to I / O interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 810 as needed so that computer programs read from it can be installed into storage section 808 as needed.
[0097] Specifically, according to embodiments of this application, the flowchart above refers to... Figure 3 The described process can be implemented as a computer software program. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. In such an embodiment, the computer program contains program code for performing the methods shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by central processing unit (CPU) 801, it performs the functions defined in the system of this application.
[0098] It should be noted that the computer-readable medium shown in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.
[0099] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operational instructions of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two connected blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified functions or operational instructions, or using a combination of dedicated hardware and computer instructions.
[0100] The units or modules described in the embodiments of this application can be implemented in software or hardware. The described units or modules can also be housed in a processor; for example, a processor may be described as including a semantic extraction unit, a weight allocation unit, and a determination unit. The names of these units or modules do not necessarily constitute a limitation on the unit or module itself.
[0101] On the other hand, this application also provides a computer-readable storage medium, which may be included in the computer device described in the above embodiments, or may exist independently and not assembled into the computer device. The aforementioned computer-readable storage medium stores one or more programs that, when used by one or more processors, execute the methods described in this application. For example, it may execute... Figure 3 The steps of the method shown.
[0102] This application provides a computer program product including instructions that, when executed, cause the method described in this application to be performed. For example, it can execute... Figure 3 The steps of the method shown.
[0103] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.
Claims
1. A work order processing method, characterized in that, The method includes: Receive a multimodal work order request, perform semantic recognition on the multimodal work order request, and determine the target business system for processing the multimodal work order request and the intent of the multimodal work order request; Based on the work order processing knowledge graph, multiple candidate functional nodes corresponding to the target business system are determined, and a target functional node matching the intent of the multimodal work order request is determined from the multiple candidate functional nodes; wherein, the multiple candidate functional nodes are slave nodes of the system node corresponding to the target business system in the work order processing knowledge graph; Based on the flow order of each business node under the target function node in the work order processing knowledge graph, the business nodes are called sequentially to process the multimodal work order request; the business nodes include: business processing nodes and / or data nodes.
2. The work order processing method according to claim 1, characterized in that, When the intent of the multimodal work order request matches multiple target function nodes, the process of sequentially calling each business node based on the flow order of the business nodes under the target function nodes in the work order processing knowledge graph includes: For each target functional node, the confidence level of the target functional node is determined based on the matching degree between the description information of the target functional node and the intent; Based on the flow order of the business nodes under the target function node with the highest confidence in the work order knowledge graph, the business nodes under the target function node with the highest confidence are called sequentially.
3. The work order processing method according to claim 2, characterized in that, The method further includes: When the confidence level of the target functional node with the highest confidence level is less than a preset threshold, supplementary parameters for the multimodal work order request are generated based on the processing requirements of the business nodes under the target functional node with the highest confidence level. The user is prompted with the follow-up confirmation information; the follow-up confirmation information is used to prompt the user to input the supplementary parameters.
4. The work order processing method according to claim 1, characterized in that, The creation process of the work order processing knowledge graph includes: Based on the historical logs generated by multiple business systems when processing multiple historical multimodal work order requests, multiple functional nodes associated with each business system are identified. For each functional node, at least one business node is determined based on the historical logs when the functional node processes the corresponding historical multimodal work order request, and the flow order of the business nodes under the functional node is formed based on the calling order of the at least one business node; wherein, the business node includes a business processing node and / or a data node; The work order processing knowledge graph is created based on the flow order of business nodes under multiple functional nodes.
5. The work order processing method according to claim 1, characterized in that, The step of performing semantic recognition on the multimodal work order request to determine the target business system for processing the multimodal work order request includes: Obtain the health assessment score of each business system, the semantic matching score between the multimodal work order request and each business system, and the attribute information of the multimodal work order request and each business system; Based on the weighted results of the health assessment degree, the semantic matching degree, and the attribute information, the target business system for processing the multimodal work order request is determined.
6. The work order processing method according to claim 5, characterized in that, The method further includes: For each historical multimodal work order request, determine the actual functional node that will ultimately process the historical multimodal work order request; Based on the number of all historical multimodal work order requests and the number of work order requests in all historical multimodal work order requests whose actual functional nodes differ from the target functional nodes determined based on the work order knowledge graph, the weight values of the health assessment degree, the semantic matching degree, and the attribute information are updated.
7. The work order processing method according to claim 1, characterized in that, The method further includes: When a business node under the target function node experiences an abnormal operation, the flow order of the business nodes under the target function node is updated.
8. A work order processing system, characterized in that, The work order processing system includes a work order processing platform and multiple business systems that interface with the work order processing platform. The work order processing platform is used to receive multimodal work order requests, perform semantic recognition on the multimodal work order requests, and determine the target business system for processing the multimodal work order requests and the intent of the multimodal work order requests from the multiple business systems. Based on the work order processing knowledge graph, multiple candidate functional nodes corresponding to the target business system are determined, and a target functional node matching the intent of the multimodal work order request is determined from the multiple candidate functional nodes; wherein, the multiple candidate functional nodes are slave nodes of the system node corresponding to the target business system in the work order processing knowledge graph; The work order processing platform is also used to, based on the work order processing knowledge graph, call the business nodes under the target function nodes associated with the target business system to process the multimodal work order requests.
9. The work order processing system according to claim 8, characterized in that, The work order processing platform includes a multimodal input module. The multimodal input module is used to receive multimodal work order requests in at least one of the following forms: voice, text, and image. The multimodal input module is further configured to obtain the input features of the multimodal work order request based on the input format of the multimodal work order request, and fuse the input features according to the text association characteristics.
10. A computer program product, characterized in that, The computer program product includes instructions that, when executed, cause the method as described in any one of claims 1-7 to be implemented.