A model training method and system based on dynamic ontology evolution and heterogeneous reasoning network

Abstract

The application belongs to the technical field of artificial intelligence and industrial application intelligent technology, and discloses a model training method and system based on dynamic ontology evolution and heterogeneous reasoning network. The model training method comprises the following steps: constructing an initial domain ontology; constructing a heterogeneous reasoning network comprising a symbolic reasoning module, a neural representation module and an attention gate module; in the training process of the network, performing model reasoning, dynamically evolving and task-aware compressing the domain ontology based on performance feedback, and performing a closed loop of updating knowledge to realize the collaborative evolution of the model and the knowledge; finally, obtaining a final reasoning model through domain-aware contrastive learning and task fine-tuning. The application solves the technical bottlenecks of the existing knowledge enhancement model, such as knowledge staticization, compression and task disconnection, and model uninterpretability when deployed on edge devices, and constructs an industrial-grade reasoning system with high precision, self-adaptation and interpretability in a resource-constrained environment.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and intelligent industrial applications. Specifically, it relates to a model training method and system, and more particularly to a model training method, an interpretable reasoning system, an equipment fault diagnosis method, and an electronic device based on dynamic ontology evolution and heterogeneous reasoning networks. Background Technology

[0002] In the field of artificial intelligence and intelligent industrial applications, injecting domain knowledge (such as industry rules, entity relationships, and causal logic) into pre-trained models to enhance their professional reasoning capabilities has become a key path to achieving intelligent applications. Especially in technical scenarios with extremely high requirements for real-time performance, accuracy, and interpretability, such as industrial equipment fault diagnosis, medical auxiliary decision-making, and real-time risk control in financial transactions, reasoning models based on domain knowledge graphs are often deployed on resource-constrained edge computing devices.

[0003] However, existing knowledge augmentation model deployment schemes, when addressing the aforementioned industrial-grade technical challenges, have revealed the following insurmountable technical bottlenecks: The mismatch between static knowledge bases and dynamic industrial scenarios leads to a rapid decline in the effectiveness of models: Equipment, processes, and rules in industrial environments are constantly evolving. Once the knowledge graph on which models deployed on edge devices rely is solidified, it cannot be updated online as the scenario changes. This causes a sharp drop in inference accuracy when the model faces new failure modes or new compliance requirements, necessitating frequent and time-consuming retraining and deployment to the cloud, severely disrupting the continuous and stable operation of industrial systems.

[0004] The disconnect between knowledge compression and specific reasoning tasks leads to low resource utilization and loss of critical information on edge devices: To adapt to the limited computing power and storage of edge devices, large-scale knowledge graphs need to be compressed. Existing compression methods (such as pruning) are general and irrelevant to the goals of downstream reasoning tasks. This indiscriminate compression easily discards logical paths that are crucial for specific fault diagnosis or compliance review tasks, resulting in a reduced-size model but compromised critical reasoning capabilities, falling into the dilemma of "compressing for the sake of compression, sacrificing core performance." This not only wastes computing resources but may also lead to serious underreporting or misjudgment.

[0005] Pure neural network "black box" models face application barriers in industrial trustworthy decision-making: existing solutions mostly rely on end-to-end deep neural networks, whose decision-making process is untraceable. In equipment failure root cause analysis or medical protocol compliance review, they only output a conclusive label without providing a chain of evidence consistent with domain logic. This makes it difficult for on-site engineers or doctors to understand and trust the model's judgment, hindering subsequent human-machine collaborative decision-making and operations. Consequently, the intelligent system is "outsourced" at critical stages and cannot be truly integrated into core business processes.

[0006] Therefore, those skilled in the art have long faced a pressing technical challenge: how to build and continuously maintain an industrial-grade inference system that can maintain high accuracy and provide clear and reliable logical explanations in critical tasks, without relying on the powerful computing power of the cloud and only utilizing the limited resources of edge devices. Existing technologies lack a systematic solution that can form a feedback control loop between the model inference process and the knowledge base evolution process, achieve task awareness, and integrate symbolic logic to ensure interpretability. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this paper provides a model training method, an interpretable reasoning system, a device fault diagnosis method, and an electronic device based on dynamic ontology evolution and heterogeneous reasoning networks. These solutions address the technical problems of existing knowledge-enhanced models when deployed on edge devices, such as the inability of static knowledge to adapt to dynamic scenarios, the disconnect between knowledge compression and specific reasoning tasks leading to the loss of key information, and the difficulty of integrating pure neural network models into high-reliability industrial processes due to the uninterpretable decision-making process.

[0008] This invention provides a model training method based on dynamic ontology evolution and heterogeneous inference networks, comprising the following steps: S1. Initial Domain Ontology Construction Steps: Obtain physical entity operation data and associated text records from industrial scenarios, and construct an initial domain ontology containing entities, attributes, relationships, and domain rules; S2. Steps for constructing a heterogeneous inference network: Construct a heterogeneous inference network; The heterogeneous inference network includes: The symbolic reasoning module is used to perform logical reasoning based on the initial domain ontology; A neural representation module is used to perform vectorized representation and reasoning on the initial domain ontology and input data; Attention gating module is used to generate dynamic fusion weights based on the input context; The heterogeneous inference network is configured to fuse the outputs of the symbolic inference module and the neural representation module according to the weights generated by the attention gating module to obtain the inference result. S3. Collaborative Training and Ontology Evolution Steps: During the training process of the heterogeneous inference network, using the initial domain ontology as the currently used domain ontology, the following iterative optimization operation based on performance feedback is executed repeatedly until a preset termination condition is met: a) Model Inference and Evaluation: The currently used domain ontology is input into the heterogeneous inference network, and a performance feedback signal is generated based on the inference performance evaluation results of the heterogeneous inference network on the domain validation set; the domain validation set is obtained from the historical data of the industrial scenario. b) Dynamic evolution and compression of ontology: Based on the performance feedback signal, the domain ontology currently in use is dynamically iteratively evolved and compressed in a task-aware manner, wherein the domain ontology is represented by a graph structure to form an optimized ontology that adapts to the current heterogeneous inference network state and task requirements. c) Knowledge update: Update the optimized ontology obtained in step b) to the domain ontology used by the heterogeneous inference network for the next round of training; S4. The heterogeneous inference network is pre-trained using a domain task-aware contrastive learning strategy and fine-tuned on a specific industrial inference task to obtain the final inference model.

[0009] As described above, in the model training method, step S1, "constructing an initial domain ontology containing entities, attributes, relationships, and domain rules," specifically includes: The associated text records are subjected to text analysis processing to extract domain entities, attributes, and relationships; The initial domain ontology is constructed by combining the state parameters and events parsed from the physical entity's operational data, as well as the rules obtained from the domain specification.

[0010] In the model training method described above, the "ontology dynamic evolution and compression" in step S3 is specifically implemented through an adversarial ontology compression network, which includes: A generator is used to compress the currently used domain ontology, which is represented in a graph structure, to generate a compressed ontology subgraph with fewer nodes and edges than the input ontology. A discriminator is trained to distinguish whether an input ontology subgraph comes from the compressed output of the generator or from a pre-defined reference ontology that supports the current inference task. The generator and the discriminator are optimized through adversarial training so that the generator can compress the domain ontology while retaining information that is key to the current reasoning task.

[0011] In the model training method described above, the "performance evaluation" in step S3 specifically involves: calculating the difference between the inference result output by the heterogeneous inference network and the standard answer, and generating the performance feedback signal using domain-specific performance evaluation metrics.

[0012] As described above, in the model training method, the "dynamic iterative evolution" in step S3 includes at least one of the following operations: Based on the performance feedback signal, the currently used domain ontology is refined, enhanced, or modified. The refinement involves using a task-aware compressed ontology subgraph; the enhancement involves incorporating relevant knowledge extracted from new data into the ontology; and the correction involves adjusting the trainable confidence weights of the logical rules in the domain.

[0013] In the model training method described above, the "domain task-aware contrastive learning strategy" in step S4 includes: Construct positive sample pairs from the same industrial context and negative samples from different contexts or with contradictory conclusions; The heterogeneous inference network is trained using a contrastive loss function so that its output feature representation satisfies the following condition: the similarity between positive sample pairs is higher than the similarity between positive and negative samples.

[0014] The present invention also provides an interpretable reasoning system, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the model training method described above to obtain a trained heterogeneous reasoning network model and an optimized domain ontology associated with it. The system is configured to receive input data from an industrial scenario and perform collaborative reasoning using the trained heterogeneous reasoning network model and the optimized domain ontology to output a reasoning result containing a logical chain of evidence.

[0015] The interpretable reasoning system described above is deployed in a resource-constrained edge computing device, industrial gateway, or local server.

[0016] The present invention also provides a method for diagnosing equipment faults, comprising: Obtain sensor operating data and relevant historical maintenance logs of the industrial equipment to be diagnosed; The sensor operating data and related historical maintenance logs are input into any of the above-mentioned interpretable reasoning systems to perform fault diagnosis using the system. Obtain the fault diagnosis results output by the system, wherein the diagnosis results include the inferred faulty component, the cause of the fault, and the logical evidence path supporting the inference.

[0017] The present invention also provides an electronic device, comprising: One or more processors; Memory, used to store one or more computer programs; When the one or more computer programs are executed by the one or more processors, the one or more processors implement the model training method described in any of the preceding claims. Attached Figure Description

[0018] Figure 1A schematic diagram of the model training method based on dynamic ontology evolution and heterogeneous reasoning network provided in an embodiment of the present invention; Figure 2 A schematic diagram of the architecture of an industrial-grade interpretable reasoning system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation methods, structures, features, and effects of the model training method based on dynamic ontology evolution and heterogeneous reasoning networks, the industrial-grade interpretable reasoning system, the equipment fault diagnosis method, and the electronic device proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The following description, in conjunction with the accompanying drawings, details the specific solutions provided by the present invention, including the model training method based on dynamic ontology evolution and heterogeneous reasoning networks, the industrial-grade interpretable reasoning system, the equipment fault diagnosis method, and the specific solutions for an electronic device.

[0022] Example 1: Please refer to Figure 1 This is a flowchart illustrating the model training method based on dynamic ontology evolution and heterogeneous inference networks provided in the embodiments of this application. Figure 1 As shown, the process of model training based on dynamic ontology evolution and heterogeneous inference networks is specifically illustrated.

[0023] The model training method based on dynamic ontology evolution and heterogeneous reasoning networks of the present invention is implemented as follows: S1. Initial Domain Ontology Construction Steps: Obtain physical entity operation data and associated text records from a specific industrial scenario, and construct an initial domain ontology containing entities, attributes, relationships, and domain rules; It should be noted that the specific industrial scenarios mentioned above include: intelligent manufacturing, smart healthcare, or financial risk control scenarios. This implementation details one specific method for implementing the above steps, which includes the following sub-steps: S101. Multi-source data acquisition and preprocessing: For the target industrial scenario (e.g., predictive maintenance of rotating machinery), raw data is obtained from the following sources: 1. Physical Entity Operation Data: Time-series signal data is collected during equipment operation using vibration sensors, temperature sensors, current sensors, etc., deployed on equipment (such as fans and pumps). This physical entity operation data includes, but is not limited to, time-series signal data. In some application scenarios, it may also include image data captured by equipment monitoring cameras or audio data used to analyze abnormal equipment noise. The time-series signal data undergoes signal preprocessing such as standardization, filtering, and noise reduction to extract key state parameters (such as RMS vibration values, spectral characteristics, and average temperature) and identify abnormal events (such as over-temperature alarms and excessive vibration).

[0024] 2. Associated Text Records: Unstructured text records related to the equipment are obtained from the enterprise's Manufacturing Execution System (MES) and Equipment Maintenance Management System, including equipment maintenance work orders, fault description reports, operation manual sections, parts lists, etc. These records undergo text preprocessing, including word segmentation, stop word removal, and stemming. The associated text records include, but are not limited to, equipment maintenance logs and fault reports, and may also cover operation manuals, electronic medical records (in medical scenarios), or financial transaction records (in risk control scenarios).

[0025] S102. Structured Information Extraction: 1. Extracting Knowledge Triples from Text Records: Pre-trained named entity recognition models (e.g., BERT-based models) are used to process pre-processed text, automatically identifying and extracting domain entities (such as "bearing", "gearbox", "lubricating oil") and their attributes (such as "model: 6205", "temperature threshold: 80℃"). Furthermore, relation extraction models or rule-based pattern matching are used to extract relationships between entities from sentences (such as "<main shaft, connection, gearbox>", "<lubricating oil, used for lubrication, bearing>").

[0026] 2. Entity-Data Association and Instantiation: The device component entities extracted in S102.1 are associated with the status parameters and abnormal events parsed from the physical entity operation data in S101.1. Specifically, by matching the device identifier (ID) with the data timestamp, sensor data (status / events) from the same device within the same time period are bound to the corresponding entity, thereby creating entity instances with real-time operating status (e.g., "Drive end bearing of device A - current vibration value: 7.5 mm / s"). Thus, it is evident that the physical entity operation data is not constructed independently of the ontology, but is deeply integrated with the structured information extracted from text records, jointly contributing to the construction of a complete initial domain ontology that includes both static domain concepts and dynamic entity operation instances.

[0027] S103. Domain Rule Integration: Formalized domain rules are collected from domain expert knowledge bases, industry standard documents, or equipment technical specifications. These rules are typically represented in the form of "IF-THEN" or first-order logic. For example, a fault diagnosis rule can be expressed as: "IF The vibration characteristic X of entity A (bearing) contains high-frequency components AND The temperature Y of entity A is consistently greater than the threshold Z THEN Fault F (bearing wear) exists (confidence level: 0.9)". Such rules are encoded into machine-readable logical expressions.

[0028] S104. Initial Ontology Construction and Representation: The entities (and their attributes), relations (and their attributes), and domain rules extracted in the above steps are organized according to an ontology modeling language (such as OWL) or a graph structure to construct the initial domain ontology. Specifically, it can be stored in a graph database in the form of an attribute graph, where: Nodes: Represent entities or literals (such as parameter values).

[0029] Edge: Represents the relationship between entities, or the "having attribute" relationship between an entity and an attribute value.

[0030] Additional attributes of nodes / edges: logical representations used to attach contextual information or rules.

[0031] Thus, the initial domain ontology was constructed from multi-source heterogeneous data within a specific industrial scenario, becoming a structured, machine-understandable ontology. This ontology provides the initial knowledge foundation for subsequent collaborative training and evolution.

[0032] S2. Heterogeneous reasoning network construction steps: Construct a heterogeneous reasoning network, which includes a symbolic reasoning module, a neural representation module, and an attention gating module; It should be noted that the heterogeneous inference network is a core model specifically designed for end-to-end training in order to achieve "cooperative training" and "interpretable reasoning" in the subsequent S3 and S4 steps.

[0033] S201. Overall Network Architecture Definition: The constructed heterogeneous inference network is built as a unified directed acyclic computation graph at the implementation level. In this computation graph, nodes represent specific executable operations, and edges represent flowing data (tensors). The network takes as input the initial domain ontology (represented as a graph structure in the computer) constructed in step S1 and the specific problem or data instance to be inferred. Its final output is a reasoning conclusion and its corresponding logical evidence path. The entire computation graph is designed to consist of three functionally complementary and data flow-coupled core trainable modules to achieve end-to-end joint optimization.

[0034] S202. Specific construction of the symbolic reasoning module: This module is a key reasoning node in the computation graph, designed to perform domain-rule-based, differentiable logical derivations to provide interpretable reasoning paths.

[0035] 1. Module Structure and Interface: This module is implemented in the computation graph as a callable, differentiable function node. Its inputs are the domain ontology (containing entities, relations, and formal rules) and the current query. Its core integrates a parameterized neural-symbolic reasoning engine.

[0036] 2. Trainability Implementation Details: To achieve trainability and compatibility with deep learning frameworks, this module employs the following specific technical methods: Parametric rule representation: Each logical rule parsed from the domain ontology (e.g., "IF vibration characteristics of entity A (bearing) X THEN fault F (bearing wear)") is associated with a trainable confidence weight w. i This weight can be initialized to a default value (e.g., 1.0) during the initial training phase.

[0037] Differentiable reasoning process: When performing logical derivation, the module executes rule-based forward chain reasoning. However, unlike traditional Boolean logic, the deterministic measure of the conclusion (which can be viewed as a continuous value) is determined by the confidence weight w of the rules involved. i The degree of matching of the preconditions (calculated through vector similarity) jointly determines the gradient. The entire derivation process is designed as a series of differentiable operations in a continuous space to ensure that the gradient can propagate backward.

[0038] Vectorization of symbols: To facilitate the calculation of matching degree, entities and relations can also be represented as trainable low-dimensional embedding vectors. These vectors are optimized during training to better capture semantic similarity.

[0039] Output: The module outputs one or more possible "logical proof trees" and their corresponding comprehensive confidence scores. These proof trees form the structured basis for the final "logical evidence chain".

[0040] 3. Trainable Parameters: In summary, the trainable parameters of this module mainly include the rule confidence weight set W. i And the embedding vectors of entities / relationships. These parameters enable the symbolic reasoning process to learn from the data, thereby supporting the dynamic evolution of domain ontology rules in step S3.

[0041] S203. Specific construction of the neural representation module: This module is a semantic learning and matching node in the computation graph, designed to process unstructured information and complex patterns through deep neural networks.

[0042] 1. Module Structure and Interface: This module consists of a series of standard, trainable neural network layers in the computation graph. Its inputs are also the domain ontology graph and the current query.

[0043] 2. Specific implementation: Ontology Encoder: Encodes the input domain ontology graph using a graph neural network (e.g., a graph attention network, GAT). Each layer of the GNN performs trainable message passing and aggregation operations, ultimately generating a high-dimensional vector representation of each entity and relation in the graph that contains its topological and semantic information.

[0044] Query encoder: Uses a text encoder (e.g., Transformer Encoder) to encode the vectorized representation of a natural language query or structured query to obtain a semantic vector of the query.

[0045] Inference computation: By calculating the similarity between the query vector and each entity vector in the vector space (such as cosine similarity), or by performing multi-hop, attention-based information propagation on the graph representation, a probabilistic inference result distribution based on distributed representation is output (e.g., a probability distribution vector over all candidate entities).

[0046] 3. Trainable Parameters: The trainable parameters in this module are all the weight matrices and bias vectors that constitute the GNN and the text encoder. These parameters are optimized in steps S3 and S4 using the standard backpropagation algorithm.

[0047] S204. Specific construction of the attention gating module: This module is a dynamic routing and control node in the computation graph. Its core function is to evaluate the context of the current input and generate a dynamic scalar gating weight g to indicate the degree of dependence on symbolic reasoning paths and neural representation paths in subsequent fusion steps.

[0048] 1. Module Structure and Interface: This module itself is a trainable, lightweight neural network, implemented as a function node in the computation graph. For example, it can be specifically implemented as a multilayer perceptron. Its inputs are the output features from the symbolic reasoning module and the neural representation module, as well as optional raw query context features. These input features typically include: the combined confidence of the logical proof tree output by the symbolic reasoning module, the entropy of the probability distribution output by the neural representation module (used to measure uncertainty), and the semantic vector of the query encoding, etc. These features are concatenated into a combined context feature vector. Its output is a scalar gating weight g, where 0 ≤ g ≤ 1. This weight g will be used as a control signal and output to the downstream independent weighted fusion computation nodes.

[0049] 2. Internal Implementation and Trainability: This MLP network performs a series of non-linear transformations on the input context feature vector (through fully connected layers and activation functions such as ReLU).

[0050] The last layer of the network uses a sigmoid activation function to map the output to the interval [0, 1], thus obtaining the final gating weights g.

[0051] The trainable parameters of this module are all the weight matrices and bias vectors of its internal MLP. Through training, this module learns to interpret different contextual features: for inputs with clear rules, strict compliance requirements, or deterministic reasoning, it tends to output a larger g value (approaching 1) to guide subsequent fusion to rely more on symbolic paths; for inputs with ambiguous semantics, relying on similarity matching, or lacking clear rules, it tends to output a smaller g value (approaching 0) to guide subsequent fusion to rely more on neural paths.

[0052] S205. Module Integration and Unified Optimization: This step integrates the three core trainable modules (symbolic reasoning module, neural representation module, and attention gating module) constructed above according to their technical functions and data dependencies to form a complete, end-to-end trainable computation graph, and defines a unified forward reasoning and backward optimization process for this computation graph.

[0053] 1. Computational graph integration and data flow definition: The modules constructed from S202 to S204 are used as nodes and connected according to their functions to form a directed acyclic computation graph with a clear data flow. The critical path is as follows: Input layer: The computation graph receives two inputs: a) the domain ontology represented in graph structure; b) specific query / input data.

[0054] Parallel processing layer: Input is fed into two parallel branches simultaneously: Symbolic reasoning path: Input data flows through the symbolic reasoning module (node). This module performs differentiable logical deduction based on the rules in the ontology, and outputs a vectorized symbolic reasoning result R. symbolic And its related logical proof information.

[0055] Neural Representation Path: Input data flows through the neural representation module (node). This module learns semantic representations through a GNN and encoder, and outputs a probabilistic neural inference result R. neural .

[0056] Gated weight generation layer: Attention gating module (node) receives intermediate features (such as R) from two paths. symbolic confidence level, R neuralThe distribution entropy, etc., and query context features are used to calculate and output a dynamic scalar gating weight g.

[0057] Fusion and Output Layer: A weighted fusion computation node is set up in the computation graph. This node receives three data points from three upstream paths: R symbolic R neural And the gating weight g. This node performs a deterministic, differentiable weighted summation operation: R final =g·R symbolic +(1 g)·R neural The result of the operation is R final This is the final output of the computation graph, namely the reasoning result containing probabilistic information. Simultaneously, it is related to R... symbolic The associated logical proof information is preserved, together forming an interpretable reasoning result and a chain of evidence.

[0058] 2. Unified optimization mechanism: Parameter set: The computation graph described above integrates the trainable parameters of all modules, forming a unified parameter set. This set include: Rule confidence weights and entity / relation embedding vectors in the symbolic reasoning module; The neural network weights and biases of the GNN and encoder in the neural representation module; MLP weights and biases in the attention gating module.

[0059] End-to-end training: During the training phase, pairs of (input, desired output) data are fed into the computation graph for forward propagation to obtain the predicted output R. final The error (loss value L) between the prediction and the expectation is calculated using a loss function (such as cross-entropy loss).

[0060] Joint backpropagation: Utilizing the automatic differentiation capability of deep learning frameworks, the gradient of L propagates back along the computational graph. Since all operations in the computational graph (including differentiable inference in the symbolic inference module) are designed to be differentiable, the gradient can flow through all nodes in the graph and ultimately compute the loss L against the parameter set. gradient of each parameter .

[0061] Parameter update: The optimizer (such as Adam) updates the parameters based on the gradients described above. Update the entire parameter set This means that a single training iteration can simultaneously and jointly optimize the rule weights of symbolic reasoning, the representational power of the neural network, and the fusion strategy of attention gating.

[0062] At this point, a complete computational graph of a "heterogeneous inference network" with a clear data flow and end-to-end joint optimization capabilities has been constructed. This network is the core object and basic architecture for all subsequent training processes, including S3 (cooperative training and ontology evolution) and S4 (task adaptation optimization).

[0063] S3. Collaborative Training and Ontology Evolution Steps: During the training process of the heterogeneous inference network, starting from the initial domain ontology, the following iterative optimization operations based on performance feedback are executed cyclically: a) Model reasoning and evaluation: Input the currently used domain ontology into the heterogeneous reasoning network, and generate a performance feedback signal based on the reasoning performance evaluation results of the heterogeneous reasoning network on the domain validation set; b) Dynamic evolution and compression of ontology: Based on the performance feedback signal, the domain ontology currently in use is dynamically iteratively evolved and compressed in a task-aware manner, wherein the domain ontology is represented by a graph structure to form an optimized ontology that adapts to the current heterogeneous inference network state and task requirements. c) Knowledge update: Update the optimized ontology obtained in step b) to the domain ontology used by the heterogeneous inference network for the next round of training; It should be noted that this step is the core of the invention to achieve "model-knowledge co-evolution", which places the training of heterogeneous reasoning networks and the optimization of domain ontology in a dynamic closed loop.

[0064] S301. Co-training Initialization: After completing the initial domain ontology (generated in step S1, denoted as G), init After constructing the initial domain ontology G and the heterogeneous inference network (generated in step S2), the collaborative training process is initiated. init Let G be the domain ontology currently in use. cur and initialize G cur =G init Initialize the training round counter t=0.

[0065] S302. Specific Implementation of Model Reasoning and Evaluation: This sub-step aims to obtain the current heterogeneous reasoning network based on the current domain ontology graph G. cur It improves inference performance and generates quantized feedback signals.

[0066] 1. Model Inference: In the collaborative training process, a domain validation set needs to be prepared in advance. This validation set is partitioned from the historical data of the industrial scenario and includes the input query and its corresponding standard answer (or label). It is used to periodically evaluate the generalization performance of the model during training and does not participate in the direct optimization of model parameters. The domain validation set can be an externally labeled independent dataset or a portion statically partitioned from the total training data according to a preset proportion (e.g., 20%). The query of each sample is compared with the current domain ontology graph G. cur The data are input together into the heterogeneous inference network, and forward propagation is performed to obtain the network's inference prediction result R. final And its associated logical proof information.

[0067] 2. Performance Evaluation: The network's prediction results R... final Calculate performance evaluation metrics by comparing the results with the standard answer. These metrics should be specific and quantifiable, for example: For classification tasks: use accuracy, F1 score, precision, and recall.

[0068] For more complex structured predictions, customized evaluation metrics based on task design can be used (e.g., mean average precision (mAP) in object detection and BLEU score in machine translation), and calculated using the corresponding evaluation function.

[0069] 3. Generate performance feedback signal: Use the calculated performance evaluation index (such as the average F1 score of this batch) or its transformed form (such as 1 - F1, as a loss signal) as the performance feedback signal P. t The signal P t It is a scalar value, and the lower the value, the better the performance of the current "network + ontology" combination.

[0070] S303. Specific Implementation of Ontology Dynamic Evolution and Compression: This sub-step is the key to the invention, aiming to optimize the domain ontology in a targeted manner based on model performance feedback. Specifically, it is implemented through an adversarial ontology compression network, which is responsible for "task-aware compression," while the evolution logic is driven by feedback signals.

[0071] 1. Drive signal: Performance feedback signal P t This is the core driver of ontology evolution. One strategy is to set a performance threshold T. If P t If the value is >T, then the current performance is considered insufficient, triggering an enhancement evolution and compression operation on the entity.

[0072] 2. Workflow of the adversarial ontology compression network: Compression (Generator): The generator G in the adversarial ontology compression network receives the current domain ontology graph G.cur And the inference task that the current model focuses on. The goal of G is to output a compressed domain ontology subgraph, denoted as G. comp Its number of nodes and edges is significantly less than that of G. cur However, it retains as much of the entities, relationships, and rule paths that are essential to completing the Task as possible.

[0073] Evaluation (Discriminator): The discriminator D is a neural network classifier. During training, D receives two types of input: one is a compressed neighborhood ontology subgraph G generated by the generator G. comp (Marked as "false"); secondly, a reference ontology subgraph that can reliably support the task (marked as "true"). The reference ontology subgraph can be an uncompressed raw G... cur It can also be a high-quality ontology fragment selected based on prior knowledge. The training goal of D is to accurately classify input samples as "true" or "false". In practice, D outputs a scalar value between 0 and 1 through functions such as Sigmoid, which represents its confidence in determining whether the input subgraph is "informative".

[0074] Adversarial Training and Task Awareness: The generator G and discriminator D are jointly optimized through adversarial training. G attempts to generate a highly compressed G... comp G "deceives" D into giving a high confidence level; D, in turn, tries to identify the subgraphs generated by G that are missing information due to over-compression. Through this game process, G eventually learns to perform selective and intelligent compression based on specific tasks, rather than blindly pruning.

[0075] 3. Dynamic evolution operation: based on feedback signal P t In response to the result of compression, the system can automatically or semi-automatically perform one or more of the following evolutionary operations to form the optimized ontology graph G for this iteration. iter_opt : Refined: G generated directly using an adversarial ontology compression network comp As G iter_opt It is a core component.

[0076] Enhancement: If P t If the model performs poorly on a certain type of entity or relationship, relevant knowledge can be extracted from the latest data and added to the dataset. cur In the graphical representation, G is formed enhanced Then input it into the adversarial ontology compression network to obtain G. iter_opt .

[0077] Correction: If a rule is ignored by the network in multiple inferences (through activation tracking of the symbolic reasoning module) and is associated with an error, then the trainable confidence weight w associated with it is reduced through backpropagation.i (See S202) This implements a soft modification to the rule. This modification directly affects G. cur The result is reflected in G based on the rule weight parameters it carries. iter_opt middle.

[0078] S304. Specific implementation of knowledge update: After the sub-step is completed, the evolution results are fed back to the model for subsequent training.

[0079] The optimized ontology graph G generated in this iteration of step S303 iter_opt Update the knowledge base for the next round of training, that is, let G cur =G iter_opt At the same time, the training round counter is updated to t=t+1.

[0080] S305. Iteration and Termination: Return to step S302 and use the updated G. cur A new round of model reasoning, evaluation, ontology evolution, and knowledge updating is then performed on the current heterogeneous reasoning network. This cycle continues until a preset termination condition is met. As a specific implementation, the termination condition can be at least one of the following: 1. The model's performance evaluation metrics (such as F1 score) on the domain validation set reach the preset target value (such as 0.95) and the fluctuation is less than the threshold δ in N consecutive iterations; 2. The compression ratio (node ​​reduction ratio) of the domain ontology meets the preset requirements; 3. Complete the specified maximum number of training epochs (T).

[0081] Thus, through iterative operations S301 to S305, the "co-training and dynamic evolution" of the heterogeneous inference network and the domain ontology was achieved. The model continuously adapts and optimizes the knowledge base during training, while the knowledge base continuously evolves and compresses under the performance feedback of the model. Both co-evolve, ultimately converging to a high-performance, high-efficiency state. When the termination condition is met, the current domain ontology graph G... cur This is the final output of the co-training and evolution step (S3), denoted as the ontology graph G after co-evolution. co This provides an optimal starting point for the subsequent task-specific fine-tuning in the S4 step. S4. The heterogeneous inference network is pre-trained using a domain task-aware contrastive learning strategy and fine-tuned on a specific industrial inference task to obtain the final inference model.

[0082] It should be noted that this step builds upon the results of step S3, "Co-training and Ontology Evolution," by further performing domain-general representation learning and task-specific fine-tuning on the co-trained heterogeneous inference network and its associated dynamically optimized domain ontology to obtain a deployable final inference model. The "pre-training" in this invention is not the general representation learning performed at the beginning of model training in the traditional sense. Instead, it involves further strengthening and generalizing the representation capabilities of the system using broader domain-specific unlabeled data after the model and knowledge base have formed a mutually compatible preliminary dedicated system through iterative co-optimization in step S3. Subsequent task-specific fine-tuning enables a smooth and efficient transition from "strong domain-general capability" to "high accuracy for specific tasks."

[0083] This step employs a two-stage training strategy. The first stage is pre-training for domain task awareness, aiming to enable the network to learn the general data distribution and inference patterns of the target industrial domain, thereby acquiring powerful domain feature representation capabilities. The second stage is fine-tuning for specific industrial inference tasks, aiming to quickly adapt the general capabilities gained from pre-training to specific downstream tasks, achieving task-optimal performance.

[0084] S401. Phase 1: Contrastive Learning Pre-training for Domain Task Awareness: The goal of this phase is to optimize the parameters of the heterogeneous inference network so that its output feature representation can better distinguish and associate the semantic relationships between different samples in the same domain.

[0085] 1. Training Data Preparation: Collect large-scale unlabeled or weakly labeled data from the target industrial sector. For example, in the field of smart manufacturing, massive amounts of equipment operation logs, historical work order texts, and sensor time-series snippets can be collected. This data does not require finely labeled inference results, but should cover typical entities, events, and relationships within the domain.

[0086] 2. Construction of positive and negative sample pairs: Based on domain knowledge, construct sample pairs for comparative learning.

[0087] Positive sample pairs: Samples from the same domain context that describe the same or highly related events. For example, two different maintenance reports describing the same type of fault in the same equipment; the same sensor data and its corresponding data-enhanced version (such as noise addition or time scaling).

[0088] Negative samples: Samples from different fields that describe unrelated events or have contradictory conclusions. For example, fan vibration reports vs. water pump temperature reports; normal operating data vs. faulty operating data.

[0089] Contrastive learning objective function: The contrastive loss function L based on InfoNCE Loss is adopted. contrastive Training is performed. For a positive sample pair (x) i,x j ), calculate its feature vector h extracted by the network. i with h j The comparison loss is used between the two. Specifically, the loss function encourages positive samples to pair with feature h. i with h j The similarity is higher than h i Similarity to features of other samples within the batch.

[0090] 3. Pre-training process: Using the heterogeneous inference network M obtained in step S3, whose parameters have been preliminarily co-optimized. co-trained and its co-evolutionary ontology graph G co This serves as the starting point for pre-training. During pre-training, both network parameters and ontology embedding vectors are compared using the contrastive loss L. contrastive Updates are performed under the guidance of [the relevant authority / system]. This process enables the network to extract domain-invariant features that are beneficial for generalization to downstream tasks and further optimizes the vector representations of entities / relationships in the ontology. After pre-training, the domain pre-trained model M is obtained. pre trained And the further optimized ontology diagram.

[0091] S402. Phase Two: Fine-tuning for Specific Industrial Reasoning Tasks: In this phase, the general capabilities acquired through pre-training will be quickly and efficiently transferred to specific, labeled industrial reasoning tasks.

[0092] 1. Fine-tuning the Task and Data: For specific industrial reasoning tasks (such as "cause and effect diagnosis of equipment failures," which also includes specialized domain tasks requiring high accuracy and interpretability, such as compliance review of medical solutions and judgment of abnormal logic in financial transactions), prepare high-quality, labeled task-specific datasets. The samples in this dataset include inputs (such as sensor data + maintenance logs) and expected outputs (such as specific faulty components, causes, and chains of evidence).

[0093] 2. Network structure adjustment and initialization: The core architecture of the heterogeneous reasoning network (symbolic reasoning module, neural representation module, attention gating module) remains unchanged.

[0094] Depending on the specific task's output format, a task-specific output layer is added at the end of the heterogeneous inference network (i.e., after the fusion layer). For example, for a K-classification task, a fully connected layer with K neurons and using the Softmax activation function is added.

[0095] Using the pre-trained model M pre trained The parameters are initialized for the entire network (including the added output layer), and the co-evolved ontology graph G is used. coThis serves as the knowledge base input used during the fine-tuning phase.

[0096] 3. Supervised Fine-Tuning Training: The network is trained using a task-specific dataset in a supervised learning manner. A loss function suitable for the task is employed, for example: For classification tasks: cross-entropy loss.

[0097] For tasks that require generating a chain of evidence: the answer prediction loss and the logical path generation loss (such as sequence generation loss) can be combined.

[0098] During fine-tuning, all network parameters (including attention gating modules, output layers, etc.) can be trained, while the entity and relation embedding vectors in the knowledge base are typically kept fixed or only allowed to be slightly adjusted to prevent catastrophic forgetting. The optimizer (such as AdamW) updates the network parameters under the guidance of the task loss function.

[0099] 4. Fine-tuning strategies: Strategies such as gradual unfreezing and tiered learning rates can be used to stabilize training and protect useful features obtained during pre-training.

[0100] S403. Final Inference Model Output and Deployment: After the above two-stage training, the final inference model M is obtained. final The model and the final optimized ontology graph G final Together, they constitute a complete industrial-grade interpretable reasoning system. This system can be deployed on resource-constrained edge devices, receiving real-time industrial data and utilizing model M. final With G as a knowledge base final It performs collaborative reasoning and outputs high-precision reasoning results with clear logical evidence chains, which can be directly applied to specific business scenarios such as equipment fault diagnosis and compliance review.

[0101] This completes the entire model training process (S1 to S4) from raw data to a deployable industrial-grade inference system. This process encompasses knowledge building, network architecture design, co-evolutionary training, domain pre-training, and task fine-tuning, forming a systematic, end-to-end solution that effectively addresses the technical bottlenecks identified in the background section.

[0102] Example 2: Please see Figure 2 This is a schematic diagram of the architecture of an industrial-grade interpretable reasoning system provided in an embodiment of this application. Figure 2As shown, a system architecture for an industrial-grade interpretable reasoning system is illustrated. This system is a direct application of the method of this invention. Its core lies in integrating an interpretable reasoning model obtained through a complete training process (S1-S4) with a dynamically optimized knowledge base, and deploying it in an industrial setting to solve the problems of static nature, black box nature, and low resource utilization mentioned in the background art.

[0103] 1. Specific composition of the system: The system 2000 of this embodiment includes a processor 2100, a memory 2200, and a computer program stored in the memory and executable on the processor, thereby obtaining a trained heterogeneous inference network model (denoted as M). final ) and its associated optimized domain ontology (denoted as G) final Specifically: Core model: M final It is a heterogeneous reasoning network trained through a complete process of steps S1 (ontology construction), S2 (network construction), S3 (co-training and evolution), and S4 (pre-training and fine-tuning). The network integrates symbolic reasoning modules, neural representation modules, and attention gating modules, and has the ability to combine rule logic and data-driven approaches, and to make dynamic decisions based on context.

[0104] Core Knowledge Base: G final It is a domain ontology that evolves and optimizes along with the model. It is not a static graph, but a dynamic knowledge structure that has undergone iterative compression and refinement in S3 and optimization through pre-training in S4. It is compact, efficient for tasks, and deeply matched with the model's inference logic.

[0105] 2. System working principle: The system is configured to receive input data from industrial scenarios (such as real-time data from equipment sensors and related text queries), and utilize the trained heterogeneous inference network model M. final With the dynamically optimized domain ontology G final Perform collaborative reasoning. The workflow is detailed below: Input and Representation: The system receives raw industrial data and uses a pre-built data processing module to convert it into a format that the model can process (such as feature vectors, text encoding). Simultaneously, it reads data from G... final Load relevant domain knowledge (entities, relationships, rules) into the middle.

[0106] Collaborative Reasoning: Model M final It uses its internal heterogeneous architecture to process input and knowledge in parallel. The symbolic reasoning module is based on G... final The rules in the system are logically deduced; the neural representation module learns the semantic association between the input and the knowledge base; and the attention gating module dynamically weighs the reliability of the two results.

[0107] Output Generation: The system outputs a reasoning result containing a logical chain of evidence. This result not only includes the final conclusion (such as "bearing X is worn"), but more importantly, it includes a structured evidence path, such as: "Trigger rule R001 (IF vibration frequency within range A THEN suspected wear) → matching real-time feature value F → historical similar cases C support → neural network confidence 0.95". This makes the decision-making process transparent and traceable.

[0108] 3. Deployment Forms and Beneficial Effects: The system can be deployed on resource-constrained edge computing devices, industrial gateways, or local servers. This deployment approach provides the following advantages: data is processed locally on the device, avoiding cloud transmission delays, meeting industrial real-time requirements, and ensuring continuous system operation even with poor network conditions; thanks to the collaborative optimization of models and knowledge, the system maintains high accuracy in specialized tasks, while its output naturally includes understandable logical evidence, greatly improving the trust and efficiency of human-machine collaboration; through task-aware compression, the knowledge base size is significantly reduced, adapting to the limited computing power and storage of edge devices; and the system has the ability to adapt to slowly changing scenarios through continuous learning.

[0109] In this application, "industrial grade" specifically refers to the application scenarios and technical indicators that the methods, systems, or devices are designed for meet the stringent requirements of industrial production and critical business areas, including but not limited to scenarios with explicit requirements for the real-time performance of system response, the accuracy and reliability of inference results, and the interpretability and traceability of the decision-making process. The solution of this invention is specifically designed to address these "industrial grade" challenges.

[0110] Example 3: This embodiment aims to specifically illustrate a method for equipment fault diagnosis. The method fully describes how to utilize the industrial-grade interpretable reasoning system described in Embodiment 2 to achieve an end-to-end, high-precision, and interpretable equipment fault diagnosis process. This embodiment uses online condition monitoring and fault diagnosis of rotating machinery (such as centrifugal fans) as a specific scenario to demonstrate how this method operates at each step from data input to result output.

[0111] The equipment fault diagnosis method described in this embodiment includes the following three steps: Step F1: Obtain sensor operating data and relevant historical maintenance logs of the industrial equipment to be diagnosed; This step is the data input phase for diagnosis. Specifically, it includes: acquiring sensor operating data and relevant historical maintenance logs. Sensor operation data acquisition: Raw time-series signals are acquired in real time or at set intervals from multiple sensors deployed on the centrifugal fan to be diagnosed (such as vibration acceleration sensors at the drive end and non-drive end, temperature sensors, and motor current sensors). These signals undergo necessary preprocessing, such as filtering, noise reduction, and synchronization, and key characteristic parameters are calculated, including the effective value, peak value, and kurtosis of vibration at each measuring point, as well as characteristic frequency amplitudes extracted through Fast Fourier Transform (FFT) or envelope demodulation analysis (such as rotational frequency, blade passage frequency, and bearing fault characteristic frequencies). These characteristic parameters constitute the structured sensor operation data required for diagnosis.

[0112] Relevant historical maintenance logs: Query the maintenance records for this fan within a recent period (e.g., the past year) from the equipment management system. These logs are typically unstructured or semi-structured text, including but not limited to: past maintenance work orders (recording fault phenomena, handling measures, and replaced parts), daily inspection records, and abnormal event reports (such as "vibration alarm," "high temperature," etc.). These text records constitute the contextual background knowledge for diagnosis.

[0113] Step F2. Input the sensor operating data and relevant historical maintenance logs into the industrial-grade interpretable reasoning system as described in the above embodiments, so as to use the system for fault diagnosis; This step involves invoking the core inference system. The structured sensor operation data processed in step F1 (in the form of feature vectors) is combined with relevant historical maintenance logs (in text or vectorized representation) to form a complete query instance. This query instance is then input into an industrial-grade interpretable inference system deployed in the field (such as an edge server or industrial gateway). This system is based on the above embodiment. After receiving the input, the system's internal model will perform collaborative inference based on a knowledge base.

[0114] Step F3. Obtain the fault diagnosis results output by the system, wherein the diagnosis results include the inferred faulty component, the cause of the fault, and the logical evidence path supporting the inference: This step involves receiving and parsing the diagnostic results. After completing the reasoning, the system outputs a structured diagnostic report. This report specifically includes the following three parts: (1) Inferred faulty component: Clearly indicate the physical component most likely to fail. For example, “drive end bearing (model: 6313)”.

[0115] (2) Deduced causes of failure: Clarify the mechanism or mode of failure. For example, "fatigue wear of the outer ring of the bearing".

[0116] (3) The logical evidence path supporting this inference: This is the core of the "interpretability" value of this invention. The evidence path is presented in the form of a clear chain: Chain of Evidence 1 (based on real-time data and rules): "Sensor characteristics: The vertical vibration at the drive end significantly exceeds the standard at the 'Bearing Outer Ring Fault Characteristic Frequency (BPFO),' which is 5.2 times the baseline value." “Triggering Domain Rule: IF The vibration spectrum of the bearing exceeds the energy limit at BPFO THEN There is a bearing outer ring wear fault (Rule ID: R_B_001, Confidence Weight: 0.94).” Chain of Evidence 2 (based on historical case connections): "Historical log matching: A similar vibration characteristic record of the same model of wind turbine was found 3 months ago (work order #2023-1105), and the fault was finally confirmed to be 'bearing outer ring wear'." Chain of Evidence 3 (based on neural network confidence and context): "Neural network determination: Under the current operating conditions and load, the model's overall confidence level for 'bearing wear' type faults is 0.91 (high confidence). The attention gating weight g=0.7, indicating that this inference focuses more on symbolic rule paths." Based on the comprehensive assessment: "In summary, considering the real-time feature matching rules, historical case support, and the high confidence of the neural network, the fault is determined to be: fatigue wear of the outer ring of the drive end bearing (model: 6313)." Example 4: See Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application, such as... Figure 3 As shown, the electronic device 3000 includes a memory 3100, a processor 3200, and a computer program stored in the memory 3100 and executable by the processor. When the processor 3200 executes the program code for the method steps according to the present invention, it executes each step of the method in the present invention, thereby realizing a model training method based on dynamic ontology evolution and heterogeneous inference networks. The model training method based on dynamic ontology evolution and heterogeneous inference networks has been described in detail above and will not be repeated here.

[0117] The memory 3100 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. The memory 3100 has storage space for storing program code for performing any of the method steps described above, which is used to execute the method steps according to the invention. The program code for performing the method steps according to the invention may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The computer device may include multiple processors, each of which may be a single-core processor or a multi-core processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0118] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A model training method based on dynamic ontology evolution and heterogeneous reasoning networks, characterized in that, Includes the following steps: S1. Initial Domain Ontology Construction Steps: Obtain physical entity operation data and associated text records from industrial scenarios, and construct an initial domain ontology containing entities, attributes, relationships, and domain rules; S2. Steps for constructing a heterogeneous inference network: Construct a heterogeneous inference network; The heterogeneous inference network includes: The symbolic reasoning module is used to perform logical reasoning based on the initial domain ontology; A neural representation module is used to perform vectorized representation and reasoning on the initial domain ontology and input data; Attention gating module is used to generate dynamic fusion weights based on the input context; The heterogeneous inference network is configured to fuse the outputs of the symbolic inference module and the neural representation module according to the weights generated by the attention gating module to obtain the inference result. S3. Collaborative Training and Ontology Evolution Steps: During the training process of the heterogeneous inference network, using the initial domain ontology as the currently used domain ontology, the following iterative optimization operation based on performance feedback is executed repeatedly until a preset termination condition is met: a) Model Inference and Evaluation: The currently used domain ontology is input into the heterogeneous inference network, and a performance feedback signal is generated based on the inference performance evaluation results of the heterogeneous inference network on the domain validation set; the domain validation set is obtained from the historical data of the industrial scenario. b) Dynamic evolution and compression of ontology: Based on the performance feedback signal, the domain ontology currently in use is dynamically iteratively evolved and compressed in a task-aware manner, wherein the domain ontology is represented by a graph structure to form an optimized ontology that adapts to the current heterogeneous inference network state and task requirements. c) Knowledge update: Update the optimized ontology obtained in step b) to the domain ontology used by the heterogeneous inference network for the next round of training; S4. The heterogeneous inference network is pre-trained using a domain task-aware contrastive learning strategy and fine-tuned on a specific industrial inference task to obtain the final inference model.

2. The model training method according to claim 1, characterized in that, Step S1, "Constructing an initial domain ontology containing entities, attributes, relationships, and domain rules," specifically includes: The associated text records are subjected to text analysis processing to extract domain entities, attributes, and relationships; The initial domain ontology is constructed by combining the state parameters and events parsed from the physical entity's operational data, as well as the rules obtained from the domain specification.

3. The model training method according to claim 1, characterized in that, The "ontology dynamic evolution and compression" in step S3 is specifically implemented through an adversarial ontology compression network, which includes: A generator is used to compress the currently used domain ontology, which is represented in a graph structure, to generate a compressed ontology subgraph with fewer nodes and edges than the input ontology. A discriminator is trained to distinguish whether an input ontology subgraph comes from the compressed output of the generator or from a pre-defined reference ontology that supports the current inference task. The generator and the discriminator are optimized through adversarial training so that the generator can compress the domain ontology while retaining information that is key to the current reasoning task.

4. The model training method according to claim 1, characterized in that, The "performance evaluation" in step S3 specifically involves: calculating the difference between the reasoning result output by the heterogeneous reasoning network and the standard answer, and generating the performance feedback signal using domain-specific performance evaluation metrics.

5. The model training method according to claim 1, wherein the "dynamic iterative evolution" in step S3 includes at least one of the following operations: Based on the performance feedback signal, the currently used domain ontology is refined, enhanced, or modified. in, The refinement involves using a task-aware compressed ontology subgraph; the enhancement involves incorporating relevant knowledge extracted from new data into the ontology; and the correction involves adjusting the trainable confidence weights of the logical rules in the domain.

6. The model training method according to claim 1, characterized in that, The "domain task-aware contrastive learning strategy" in step S4 includes: Construct positive sample pairs from the same industrial context and negative samples from different contexts or with contradictory conclusions; The heterogeneous inference network is trained using a contrastive loss function so that its output feature representation satisfies the following condition: the similarity between positive sample pairs is higher than the similarity between positive and negative samples.

7. An interpretable reasoning system, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program implements the model training method as described in any one of claims 1 to 6 to obtain a trained heterogeneous inference network model and an optimized domain ontology associated therewith. The system is configured to receive input data from an industrial scenario and perform collaborative inference using the trained heterogeneous inference network model and the optimized domain ontology to output an inference result containing a logical chain of evidence.

8. The interpretable reasoning system according to claim 7, characterized in that, The system is deployed on resource-constrained edge computing devices, industrial gateways, or local servers.

9. A method for diagnosing equipment faults, characterized in that, include: Obtain sensor operating data and relevant historical maintenance logs of the industrial equipment to be diagnosed; The sensor operating data and related historical maintenance logs are input into the interpretable reasoning system as described in claim 7 or 8, so as to use the system for fault diagnosis; Obtain the fault diagnosis results output by the system, wherein the diagnosis results include the inferred faulty component, the cause of the fault, and the logical evidence path supporting the inference.

10. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more computer programs; When the one or more computer programs are executed by the one or more processors, the one or more processors implement the model training method as described in any one of claims 1 to 6.

Images