A bearing intelligent machining process recommendation system based on a knowledge graph

Abstract

The application relates to the technical field of intelligent manufacturing and precision machining, and discloses a bearing intelligent machining process recommendation system based on a knowledge graph, which comprises a graph construction module, a process recommendation module, a data sensing module, a quality prediction module, a feedback control module and a self-evolution module.The graph construction module generates a graph by fusing multi-source data by means of a large model and an alignment algorithm, the process recommendation module generates a recommended scheme and extracts a target value by combining hybrid retrieval and rule checking, the data sensing module realizes time sequence alignment of multi-dimensional sensing data based on a workpiece ID, the quality prediction module realizes real-time prediction of machining errors by means of an LSTM network, the feedback control module calculates a compensation amount according to the predicted errors and drives a machine tool to execute dynamic deviation correction, and the self-evolution module feeds back measured data to the graph based on SPC analysis.The application constructs a closed loop of pre-recommendation, in-process control and post-iteration, solves the problems that process design relies on experience and lacks real-time correction, and improves bearing machining precision and stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and precision machining technology, specifically to a bearing intelligent machining process recommendation system based on knowledge graphs. Background Technology

[0002] As a key basic component of mechanical equipment, the quality of bearing manufacturing directly determines the performance, lifespan, and reliability of the main machine. With the increasing demands for precision and consistency in bearings from high-end equipment, the design and control of manufacturing processes have become core aspects of the manufacturing process.

[0003] Current bearing manufacturing process design primarily relies on engineers consulting numerous scattered process manuals, industry standards, and historical production records. Due to the diverse formats of this data, and the lack of an effective automated fusion mechanism between unstructured technical documents and structured parameter databases, the process design process is highly dependent on the individual experience and knowledge of the designers. This dependence not only results in low efficiency in generating process solutions but also easily leads to poor consistency in solutions due to differences in personnel experience, making it difficult to meet the demands of large-scale customized production for standardized and intelligent process design.

[0004] In actual cutting or grinding processes, the machining environment is constantly changing. Time-varying factors such as tool wear, machine tool thermal deformation, and material stress release affect the workpiece's geometric accuracy in real time. However, existing machining control methods typically employ post-processing offline measurements or simple monitoring based on single signal thresholds. They lack the ability to accurately align and process multi-source sensor data over time, and they haven't incorporated deep learning models to effectively capture and predict the temporal characteristics of the machining process. This lagging control mode prevents the system from anticipating dimensional deviations before they occur, making it difficult to achieve millisecond-level real-time dynamic compensation, resulting in insufficient stability in machining quality.

[0005] Furthermore, traditional manufacturing systems typically operate in an open-loop manner, lacking an automatic feedback and iteration mechanism from production results to process design. Although a large amount of finished product measurement data is generated at the end of production, this data is often only used for quality judgment of the current batch, failing to establish effective correlation analysis and knowledge backtracking with upstream process parameters. The system cannot automatically learn from successful or failed processing cases, resulting in the process knowledge base remaining static and stagnant for a long time, unable to evolve and optimize itself with the accumulation of production data. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a knowledge graph-based intelligent bearing machining process recommendation system, which solves the problems of existing technologies, such as process design relying on human experience and difficulty in multi-source data fusion, lack of real-time dynamic control capabilities based on time-series prediction, and lack of a knowledge closed-loop iterative mechanism based on measurement data feedback.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] This invention provides a knowledge graph-based intelligent bearing machining process recommendation system, including a knowledge graph construction module, a process recommendation module, a data perception module, a quality prediction module, a feedback control module, and a self-evolution module.

[0009] The graph construction module receives technical documents and structural data, performs extraction and fusion processing based on a large language model and alignment algorithm, and generates process graphs.

[0010] Specifically, the graph construction module receives externally imported technical documents through the extraction unit, uses a large language model to perform semantic parsing and entity relation extraction on the technical documents, and outputs the original triples.

[0011] The alignment and fusion unit receives externally imported structural data and original triples, extracts the entities contained within them, maps the entities into high-dimensional vectors using an embedding model, and calculates cosine similarity using a vector algorithm and a similarity formula.

[0012] When the cosine similarity exceeds a pre-set alignment threshold based on a large-scale labeled dataset, the alignment algorithm determines that entities with different names are the same object and performs attribute merging, outputting a unique entity. The graph storage unit, based on a predefined graph pattern, utilizes a graph database to construct nodes and edges and establish indexes, generating a process graph.

[0013] The process recommendation module is used to respond to user needs and retrieve process maps, and generate process recommendation schemes by combining hybrid retrieval and rule verification.

[0014] Specifically, the requirement parsing unit receives user requirements from the user interface, uses semantic algorithms to identify entities and intentions, and extracts constraint parameters.

[0015] The hybrid retrieval unit combines constraint parameters and process diagrams, and uses retrieval algorithms to perform vector-based semantic similarity retrieval and graph-based adjacency traversal retrieval to construct candidate subgraphs.

[0016] The rule verification unit introduces hard rules pre-set based on physical limits and safety specifications, and uses the rule engine to perform logical verification and pruning on the process paths in the candidate subgraphs, and outputs compliant paths.

[0017] The scheme generation unit integrates structured information using the generative model and fills in specific parameter suggestions to generate recommended process schemes and extract target values ​​from them.

[0018] The data sensing module is used to collect environmental status data, and combines it with the workpiece ID output by the machine tool control system to complete data timing alignment and feature extraction, and output feature vectors.

[0019] Specifically, the data acquisition unit uses displacement sensors, strain sensors, and acoustic emission sensors to convert the physical quantities of the environmental conditions at the processing site into electrical signals.

[0020] The timing alignment unit combines the original signal and the workpiece ID output by the machine tool control system, and uses an alignment algorithm with a sliding window mechanism to extract data within the corresponding processing cycle using the workpiece ID as an index, thereby achieving timing alignment of the data.

[0021] The feature extraction unit performs wavelet denoising and transform processing on the aligned data, performs time-domain statistics and frequency-domain analysis, and extracts time-frequency features to construct feature vectors.

[0022] The quality prediction module is used to input feature vectors into the LSTM network for time series modeling and error mapping, and output prediction error.

[0023] Specifically, the temporal modeling unit inputs the feature vector into the LSTM network, combines it with the internal state of the previous time step to learn the temporal features, and uses the state update formula to calculate the hidden state.

[0024] The error mapping unit uses a fully connected layer network structure to perform nonlinear mapping and regression analysis on the hidden state, and outputs a quantified prediction error.

[0025] The feedback control module is used to calculate the control deviation based on the prediction error and the target value, calculate the compensation amount through the PID algorithm, and generate control commands to drive the machine tool to execute.

[0026] Specifically, the deviation calculation unit uses a subtractor to calculate the difference between the prediction error and the target value to obtain the control deviation.

[0027] The compensation calculation unit uses the compensation formula and the PID algorithm to perform proportional, integral, and derivative operations on the control deviation to calculate the compensation amount.

[0028] The strategy distribution unit uses a communication program to convert the compensation amount into a machine tool-recognizable protocol instruction based on the current machining process type. For turning processes, it outputs the tool offset compensation value; for grinding processes, it outputs the feed rate instruction.

[0029] The self-evolution module is used to perform quality analysis and knowledge management based on measurement data, recommended process schemes, and prediction errors, outputting updated maps and feeding them back to the map construction module.

[0030] Specifically, the quality analysis unit receives measurement data from the measuring instrument, uses the SPC algorithm to perform statistical calculations on the distribution of the measurement data, and outputs the evaluation results.

[0031] Based on the evaluation results, recommended process schemes, and prediction errors, the knowledge management unit uses an update algorithm to generate case nodes and inserts them into the process graph to establish new relationships. It then outputs the updated graph and feeds it back to the graph construction module, completing the closed-loop update of the knowledge base.

[0032] This invention provides a knowledge graph-based intelligent bearing machining process recommendation system. It has the following beneficial effects:

[0033] 1. This invention achieves intelligent and standardized process design by constructing a graph construction module that includes extraction, alignment, and fusion mechanisms, combined with a process recommendation module based on hybrid retrieval and rule verification. The system can automatically process multi-source heterogeneous technical documents and structured data, use large language models and embedding models to solve entity alignment problems, and generate reliable process recommendation schemes by combining hard rule constraints. This achieves effective fusion of unstructured technical documents and structured data, reduces the reliance on human experience in process design, and improves the efficiency and accuracy of process scheme generation.

[0034] 2. This invention achieves millisecond-level real-time closed-loop control of the bearing machining process through the collaborative work of data perception, quality prediction, and feedback control modules. It solves the problem of asynchronous data from multiple sensors by using a sliding window alignment mechanism based on workpiece ID. Combined with the state update logic of the LSTM network, it accurately captures the time-varying characteristics of the machining process and can output the prediction error before the workpiece size deviation is formed. Based on the deviation between the prediction error and the target value, the system automatically generates specific compensation instructions for turning or grinding processes using a PID algorithm, effectively suppressing machining errors caused by tool wear and thermal deformation, and improving the geometric accuracy and consistency of the finished product.

[0035] 3. This invention introduces a self-evolutionary module to construct a complete data closed loop of pre-recommendation, in-process control, and post-iteration. By statistically analyzing the actual measurement data through the SPC algorithm, the system can automatically identify high-quality process cases or failure cases and feed these cases back into the process map, enabling the system to have the ability to learn continuously. With the accumulation of processing data, the knowledge coverage of the process map and the reliability of the process recommendation scheme will continue to improve, thereby adapting to the ever-changing production conditions and quality requirements. Attached Figure Description

[0036] Figure 1 This is an architecture diagram of a knowledge graph-based intelligent bearing machining process recommendation system according to the present invention.

[0037] Figure 2This is a process recommendation module architecture diagram for the present invention;

[0038] Figure 3 This is a diagram of the quality prediction module architecture of the present invention.

[0039] Among them, 100 is the map construction module; 200 is the process recommendation module; 300 is the data perception module; 400 is the quality prediction module; 500 is the feedback control module; and 600 is the self-evolution module. Detailed Implementation

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please see the appendix Figure 1 This invention provides a knowledge graph-based intelligent bearing machining process recommendation system, which adopts a collaborative deployment architecture of cloud, edge, and terminal. The terminal is equipped with various sensors and machine tool control interfaces for physical data acquisition and command execution; the edge is equipped with an edge computing gateway and a real-time processing server for real-time data alignment and low-latency control calculations; and the cloud is equipped with a graph database and large model services for knowledge storage, complex reasoning, and model training.

[0042] The system mainly includes a map construction module 100, a process recommendation module 200, a data sensing module 300, a quality prediction module 400, a feedback control module 500, and a self-evolution module 600.

[0043] The knowledge base construction module 100 is responsible for building the system's domain-specific technology knowledge base. This module receives imported technical documents through an extraction unit, parses the documents using a large language model, and outputs raw triples. The alignment and fusion unit receives imported structural data and extracted entities from the raw triples, maps the entities to high-dimensional vectors using an embedding model, and calculates the cosine similarity between vectors using a vector algorithm. During this process, the system uses a similarity formula to calculate the degree of association between entities. The similarity formula is as follows:

[0044] ;

[0045] In the formula: For entities With entity Cosine similarity between them; This is the high-dimensional feature vector generated by mapping entity A; This is the high-dimensional feature vector generated after mapping entity B; For vectors The modulus length; For vectors The length of the module.

[0046] When the calculated cosine similarity exceeds a preset alignment threshold, the alignment algorithm determines that entities with different names are the same object and performs attribute merging, outputting a unique entity. The graph storage unit receives the unique entity and its relationships, uses the graph database to construct nodes and edges, and finally forms the process graph.

[0047] The process recommendation module 200 responds to specific processing requests based on the established process map. The requirement parsing unit receives user requirements, uses semantic algorithms to identify entities and intentions, and extracts constraint parameters. The hybrid retrieval unit receives constraint parameters and the process map, and uses retrieval algorithms to perform vector retrieval and graph traversal retrieval to construct candidate subgraphs.

[0048] The rule verification unit receives candidate subgraphs and, in conjunction with preset hard rules, uses a rule engine to logically verify the process path, outputting a compliant path. The solution generation unit receives the compliant path and constraint parameters, integrates the information using a generation model, and generates a recommended process solution containing process sequences and equipment parameters. The target value in this recommended process solution will be passed to subsequent modules for control reference.

[0049] The data sensing module 300 is responsible for real-time data processing at the machining site. The data acquisition unit collects environmental data through displacement sensors, strain sensors, and acoustic emission sensors, and outputs raw signals. The timing alignment unit receives the raw signals and the workpiece ID output by the machine tool, and uses an alignment algorithm with a sliding window mechanism to extract the periodic data corresponding to the workpiece ID, outputting the aligned data. The feature extraction unit receives the aligned data, uses a feature algorithm for denoising and time-frequency analysis, and constructs feature vectors.

[0050] The quality prediction module 400 predicts processing quality based on real-time data. The temporal modeling unit receives feature vectors and uses an LSTM network for temporal feature learning. During this process, the system calculates the hidden state using a state update formula. The state update formula is as follows:

[0051] ;

[0052] In the formula: For the current discrete time step The hidden state; The symbol for the Sigmoid activation function; This is the weight matrix of the output gate; The hidden state of the previous moment With current input features Vector concatenation; This is the bias vector for the output gate; This is the Hadamard product operator; The symbol for the hyperbolic tangent activation function; For the current discrete time step The cellular state.

[0053] The error mapping unit receives the hidden state, performs regression analysis using a fully connected layer, and outputs the prediction error.

[0054] The feedback control module 500 performs closed-loop regulation based on the prediction results. The deviation calculation unit receives the prediction error and the target value from the recommended process scheme, calculates the difference between the two using a subtractor, and outputs the control deviation. The compensation calculation unit receives the control deviation and performs calculations using a PID algorithm. During this process, the system uses a compensation formula to calculate the compensation amount. The compensation formula is as follows:

[0055] ;

[0056] In the formula: For the current continuous time moment The output compensation amount; This is the proportional control coefficient; For the current continuous time moment Control deviation; These are integral control coefficients; From time 0 to the current time The definite integral operator; For integration variables; These are the differential control coefficients; For time Differential operators for finding derivatives.

[0057] The strategy distribution unit receives the compensation amount and converts it into control commands that the machine tool can recognize using a communication program. For turning processes, the control commands are expressed as tool offset compensation values; for grinding processes, the control commands are expressed as feed rate commands.

[0058] The self-evolution module 600 is used for continuous system optimization. The quality analysis unit receives measurement data from measuring instruments, calculates the process capability index using the SPC algorithm, and outputs the evaluation results. The knowledge management unit receives the evaluation results, recommended process schemes, and prediction errors, generates case nodes using the update algorithm, inserts them into the knowledge graph, and outputs the updated knowledge graph. The updated knowledge graph is fed back to the knowledge graph construction module 100, completing the closed-loop update of the knowledge base.

[0059] The knowledge graph construction module 100 serves as the domain-specific knowledge base for the entire system. Its core function is to transform multi-source heterogeneous data into a structured knowledge network that is understandable and reasonable by computers. The processing flow of the knowledge graph construction module 100 covers the entire process from acquiring, cleaning, extracting, and fusing raw data to final storage.

[0060] For processing unstructured data, the extraction unit receives unstructured technical documents imported from external sources. These unstructured technical documents include, but are not limited to, process manuals in the bearing manufacturing field, national and industry standards (such as GB / T 18254 high carbon chromium bearing steel standard), and historical failure case logs accumulated on the production line.

[0061] Before performing substantive information extraction, the system first performs data cleaning and preprocessing on these documents. The preprocessing uses regular expressions to remove irrelevant characters, garbled text, and formatting control characters to improve the accuracy of subsequent processing. The cleaned text data is then input into a large language model. This large language model uses a pre-trained model based on the Transformer architecture (such as DeepSeek-V3 or GPT-4 architecture) and has been fine-tuned using a specialized corpus in the bearing manufacturing field (containing millions of process-related terms) to enable it to understand specific process terminology.

[0062] The large language model performs semantic parsing and entity relation extraction on the text, identifies key entities in the text (such as GCr15, vacuum quenching, martensite) and their interrelationships (such as applicable process, causing defects), and outputs the original triplet containing entity, relation, and entity structure.

[0063] For processing structured data and fusing multi-source data, the alignment and fusion unit is responsible for performing entity mapping and similarity calculation. This unit receives an externally imported structured database and the extracted entities from the aforementioned generated original triples.

[0064] The structured database contains standardized parameter information, specifically including: a material thermophysical property parameter library (covering density, Young's modulus, and thermal conductivity data for high-carbon chromium steel and carburized steel, sourced from JMatPro software calculations or standard material libraries), an equipment parameter library (covering inherent equipment properties such as lathe spindle accuracy and grinding wheel speed range), and a mechanical performance index library. Because different data sources may describe the same object differently (e.g., GCr15 versus high-carbon chromium bearing steel), entity alignment is required.

[0065] The alignment and fusion unit utilizes an embedding model to map standard entities in the structured database and extracted entities from the original triples into high-dimensional vectors. This embedding model, based on BERT (Bidirectional Encoder Representations from Transformers) or a similar deep learning architecture, can transform textual semantics into dense vector representations. Subsequently, vector algorithms are used to operate on the generated vectors to calculate the cosine similarity between entities.

[0066] This invention uses a similarity formula to calculate the cosine similarity between entities, as follows:

[0067] ;

[0068] In the formula: For entities With entity Cosine similarity between them; This is the high-dimensional feature vector generated by mapping entity A; This is the high-dimensional feature vector generated after mapping entity B; For vectors The modulus length; For vectors The length of the module.

[0069] After calculating the cosine similarity, the system compares it with a preset alignment threshold. This threshold is an empirical value set based on the principle of maximizing F1 scores on large-scale labeled datasets, for example, 0.85. When the calculated cosine similarity exceeds this threshold, the alignment algorithm determines that entities with different names are the same physical object and performs an attribute merging operation, adding attributes extracted from unstructured documents to the attribute list of the structured entity, outputting a unique entity with a globally unique identifier (UUID).

[0070] The graph storage unit receives the unique entities and their relationships generated by the above processing. This unit utilizes a graph database to construct and index nodes and edges according to a predefined graph schema. The graph schema defines node types (e.g., material, process, defect) and edge types (e.g., contain, precede, cause). The graph database employs a native graph storage engine (e.g., Neo4j) that supports property graph models, enabling efficient depth-first traversal and relational queries. The resulting process graph contains end-to-end knowledge from material properties and process parameters to quality feedback, providing data support for subsequent intelligent recommendations and decision-making.

[0071] See attached document Figure 2The process recommendation module 200 utilizes the constructed process graph to transform unstructured user intents into standardized, engineering-executable process solutions. This module integrates natural language processing, graph retrieval algorithms, and rule-based reasoning techniques, achieving full automation from semantic understanding to solution generation.

[0072] The requirements parsing unit receives user requirements from the user interface. These requirements are typically input in the form of natural language text, covering product type (e.g., tapered roller bearings), geometric parameters (e.g., inner diameter, outer diameter, width), and specific constraints (e.g., cost priority or high fatigue life). To accurately understand this unstructured text, the requirements parsing unit uses semantic algorithms to process the text.

[0073] This semantic algorithm is built upon the BERT (Bidirectional Encoder Representations from Transformers) model. Specifically, it employs Named Entity Recognition (NER) technology to locate key entities in the text, and combines this with slot filling technology to map the extracted information onto predefined parameter templates. Through this processing, the system extracts structured constraint parameters from fuzzy natural language, such as explicit material grades, tolerance levels (IT grades), heat treatment hardness requirements, and cost limits.

[0074] The hybrid retrieval unit receives the extracted constraint parameters and the process graph generated by the process knowledge graph construction module. To address the limitations of a single retrieval mode, this unit employs a dual retrieval strategy using a retrieval algorithm.

[0075] First, a vector-based semantic similarity retrieval is performed, converting non-numerical descriptions in the constraint parameters (such as high wear resistance) into vectors to recall historical case nodes with similar semantic features in the graph. Second, a graph-based adjacency traversal retrieval is performed, starting from the recalled case node or a specific entity node (such as GCr15), and performing a multi-hop traversal along the applicable process and equilateral relationships of the included processes in the graph to search for associated equipment nodes and standard parameter nodes. The results of these two retrievals are then weighted and fused to form a candidate subgraph containing multiple potential process paths.

[0076] The rule verification unit receives candidate subgraphs and uses preset hard rules for logical filtering. These hard rules are a set of logic based on physical limit formulas, safety production regulations, and national standards (such as GB / T 18254) in the bearing manufacturing field, and are stored in the rule engine. The rule engine uses the Drools engine and supports forward chaining. The rule verification unit matches and verifies each process path in the candidate subgraph against the hard rules.

[0077] For example, the system has a speed threshold rule: when the workpiece's designed speed exceeds the threshold, a dynamic balancing correction process must be added before the fine grinding process. This speed threshold is set based on the bearing rotor dynamics stability requirements, for example, 15,000 rpm. If a process path for high-speed bearings in the candidate subgraph lacks a dynamic balancing process, the rule engine will automatically determine that the path is in violation and prune it. After traversing and verifying all rules, the system outputs a compliant path that fully conforms to physical and engineering constraints.

[0078] The solution generation unit receives the verified compliant path and the original constraint parameters. This unit uses a generative model to transform the structured path information into a final process recommendation solution. The generative model employs a Retrieval Enhanced Generation (RAG) framework, using Prompt engineering techniques to input the compliant path as contextual prompts into the large language model, guiding the model to fill in specific process parameter suggestions (such as specific cutting speed values ​​and feed rate ranges). The generated process recommendation solution includes two formats: one is JSON format data for inter-system data exchange, which explicitly defines target values ​​(e.g., roundness error ≤ 0.8μm, diameter fluctuation ≤ 0.03mm), which will serve as the benchmark for subsequent feedback control modules; the other is a natural language report easy for engineers to read, detailing the process sequence, recommended equipment models, and key considerations.

[0079] The data sensing module 300 is responsible for converting the multi-source heterogeneous physical quantities at the processing site into standardized digital features. This data sensing module 300 mainly includes a data acquisition unit, a timing alignment unit, and a feature extraction unit.

[0080] The data acquisition unit is responsible for acquiring the real-time status of the physical machining environment. This unit achieves multi-dimensional information perception by deploying various types of sensors in key parts of the machine tool.

[0081] Specifically, displacement sensors are used for acquiring position information. In grinding scenarios, non-contact inductive displacement sensors (such as the probes in the Marbus measurement and control system) or capacitive displacement sensors are selected and installed on the measuring arm or spindle box of the grinding machine. They are used to monitor changes in workpiece diameter and the feed position of the grinding wheel in real time. The sampling frequency is set to a value that satisfies the Nyquist sampling theorem (e.g., 10kHz) to ensure that micron-level dimensional fluctuations can be captured.

[0082] For acquiring cutting force information, strain sensors are used. Strain sensors (such as resistance strain gauges) are attached to the force-sensitive area of ​​the lathe tool post or grinding headstock to sense changes in cutting resistance during machining, thereby indirectly reflecting the tool wear condition or the risk of grinding burn.

[0083] Acoustic emission sensors are used to collect data on contact status and abnormal vibrations. These sensors are mounted on machine tool components (such as the tailstock or grinding wheel head) near the machining area to capture high-frequency elastic wave signals generated at the moment of contact between the grinding wheel and the workpiece (end of idle stroke), as well as abnormal collision signals during the machining process.

[0084] The aforementioned sensor converts the collected physical quantities into analog electrical signals, which are then processed by an analog-to-digital converter (ADC) to output a multi-channel raw signal containing position, force, and acoustic emission information.

[0085] The timing alignment unit receives raw signals from multiple channels and the workpiece ID output by the machine tool control system. Because different sensors may have different data acquisition frequencies (e.g., acoustic emission signals are high-frequency, while machine tool load signals are low-frequency), and data transmission may have varying network delays, data arriving at the system at the same time may not correspond to the same physical moment. Furthermore, workpieces on a continuous production line not only flow continuously, but the system needs to clearly distinguish which specific workpiece each segment of data corresponds to.

[0086] To address this issue, the timing alignment unit utilizes an alignment algorithm to perform data synchronization. This algorithm employs a sliding window mechanism, using the workpiece ID as an index for data extraction. The workpiece ID is a unique code obtained by reading a specific address from the machine tool's PLC register (or a macro variable in the CNC system). When the machine tool begins machining a workpiece, the PLC writes the workpiece's ID into the designated register; when machining ends, the ID is updated.

[0087] The alignment algorithm monitors changes in the register in real time, using the moment of ID change as the start and end point of a time window, and captures the data stream from all sensor channels within that time window. Simultaneously, it utilizes the unified time synchronization service (based on NTP or PTP protocol) provided by the edge computing gateway to perform timestamp correction on the data from all channels, ensuring that the timestamp synchronization error of the aligned data is controlled within milliseconds (e.g., ≤10ms). After the above processing, the system outputs aligned data that strictly corresponds to a specific workpiece ID.

[0088] The feature extraction unit receives the aligned data and performs noise reduction and feature engineering on it using a feature algorithm to extract the information most representative of the processing quality. The feature algorithm first performs wavelet denoising on the aligned data, using wavelet transform to decompose the signal into different frequency bands, filtering out high-frequency noise interference, and retaining the main trend terms of the signal.

[0089] Subsequently, the algorithm extracts features from both the time and frequency domains. In terms of time-domain statistics, it calculates the mean (reflecting the average cutting load), variance (reflecting the fluctuation of the cutting process), and peak value (reflecting the maximum impact force) of the cutting force signal. In terms of frequency-domain analysis, it performs a Fast Fourier Transform (FFT) on the acoustic emission signal or vibration signal to calculate the energy spectral density of a specific frequency band in order to identify anomalous frequency components caused by flutter or grinding wheel imbalance.

[0090] To eliminate the order-of-magnitude differences between different physical units (such as force units N and position units mm), the feature algorithm performs Z-score standardization on all extracted feature values, converting them into standard distribution data with a mean of 0 and a standard deviation of 1. Finally, this unit constructs a standardized feature vector (i.e., a real-time feature vector) as input to the subsequent quality prediction model.

[0091] See attached document Figure 3 The quality prediction module 400 is responsible for using deep learning algorithms to mine the temporal correlation features of the processing data, thereby predicting the geometric quality indicators of the workpiece in advance before the actual processing is completed. The quality prediction module 400 mainly consists of a temporal modeling unit and an error mapping unit. Through an end-to-end mapping mechanism, it converts high-dimensional sensor signals into quantifiable quality deviation values.

[0092] The time-series modeling unit receives real-time feature vectors (denoted as ) from the data-aware module. Because the bearing machining process exhibits significant time-varying characteristics, the current machining quality depends not only on the current cutting parameters but also on historical conditions such as tool wear history and accumulated machine tool thermal deformation. Therefore, the temporal modeling unit employs a Long Short-Time Memory (LSTM) network as its core algorithm.

[0093] LSTM networks, by introducing gating mechanisms (forget gate, input gate, output gate), can effectively capture long-term dependencies in time-series data, solving the gradient vanishing problem inherent in traditional recurrent neural networks (RNNs). In terms of specific network structure design, a multi-layered stacked LSTM structure (e.g., 2 to 3 layers) is employed to extract deeper, more abstract features. The model operates at each discrete time step... Based on the internal state of the previous time step and the input features of the current time step, it updates its own cell state and hidden state.

[0094] This invention uses a state update formula to calculate the hidden state at the current moment, as follows:

[0095] ;

[0096] In the formula: For the current discrete time step The hidden state; This is the symbol for the Sigmoid activation function, used to limit the output between 0 and 1 as a gating signal; This is the weight matrix of the output gate, which is automatically learned and updated during model training using the backpropagation algorithm; The hidden state of the previous moment With current input features Vector concatenation; This is the bias vector for the output gate; This is the Hadamard product operator; This is the symbol for the hyperbolic tangent activation function, used for nonlinear transformations of cell states; For the current discrete time step The cell state, determined by the forget gate and the input gate, carries the memory of historical information.

[0097] The error mapping unit receives the hidden state output by the time series modeling unit. Since the hidden state is a high-dimensional vector, it cannot be directly used for control decisions and needs to be mapped to a specific physical quantity. The error mapping unit uses a fully connected layer to perform nonlinear mapping and regression analysis. The fully connected layer projects the high-dimensional hidden state space onto a low-dimensional label space.

[0098] In the model training phase, to prevent overfitting and improve the model's generalization ability under different conditions, the fully connected layer introduces the Dropout technique, which randomly discards some neuron connections with a certain probability during training.

[0099] The unit ultimately outputs a quantified prediction error. The specific physical meaning of this prediction error depends on the current machining process: in turning, the prediction error refers to the roundness error of the workpiece's cross-section (the radial offset relative to the least squares circle); in grinding, the prediction error refers to the deviation of the workpiece's outer diameter from the target diameter. This prediction error is trained through supervised learning by minimizing the mean square error (MSE) loss function, ensuring that the predicted value approximates the actual offline measurement value as closely as possible.

[0100] The feedback control module 500 connects the preceding quality prediction results with the underlying machine tool actuators, and is responsible for converting the prediction deviation in the digital space into corrective actions in the physical world, thereby achieving closed-loop control of machining accuracy. Specifically, the feedback control module 500 includes a deviation calculation unit, a compensation calculation unit, and a strategy distribution unit.

[0101] The deviation calculation unit receives the predicted error value from the quality prediction module and the target quality setting value from the process recommendation module (process recommendation scheme). The target quality setting value is an ideal geometric parameter determined based on product design drawings and process requirements, such as the nominal value of the roller outer diameter or the allowable limit of roundness error (such as 0 value). The predicted error value reflects the expected degree of dimensional deviation under the current processing conditions.

[0102] The deviation calculation unit uses a subtractor to perform a difference calculation operation, subtracting the prediction error value from the target quality setpoint to obtain the control deviation (denoted as ). This control deviation reflects the gap between the current processing state and the ideal state, and is a direct input to subsequent control algorithms.

[0103] The compensation calculation unit receives the control deviation and uses the PID control algorithm to determine the adjustment amount required to eliminate the deviation. The PID control algorithm, through a linear combination of proportional, integral, and derivative components, takes into account the response to the current error, the elimination of historical errors, and the prediction of error change trends.

[0104] This invention uses a compensation formula to calculate the control compensation amount, as follows:

[0105] ;

[0106] In the formula: For the current continuous time moment The output compensation amount; This is the proportional control coefficient, used to reflect control deviation. The size of the [something] determines the strength of the control effect; For the current continuous time moment Control deviation; These are integral control coefficients used to eliminate steady-state errors; From time 0 to the current time The definite integral operator represents the cumulative effect of the bias; For integration variables; These are differential control coefficients, used to reflect the changing trend of the deviation signal and provide proactive control; For time Differential operators for finding derivatives.

[0107] The coefficients in the above PID algorithm ( The parameters were determined through a combination of offline simulation and online self-tuning. In the offline simulation phase, historical machining data was used to establish a machine tool response model and initially tune a set of baseline parameters. In the online machining phase, the parameters were fine-tuned based on real-time feedback of the control effect.

[0108] The strategy distribution unit receives the calculated control compensation amount and converts it into specific protocol instructions recognizable by the machine tool's numerical control system (CNC) based on the current machining process type (turning or grinding). This unit utilizes a communication interface program to perform instruction conversion and distribution. The communication interface program supports industrial Ethernet buses (such as EtherCAT) or CNC-specific interface protocols (such as FANUC FOCAS) to ensure the real-time transmission of control instructions, with system control latency controlled within 100ms.

[0109] For grinding processes (such as centerless entry grinding of outer diameter), the strategy distribution unit executes real-time process control strategies. Because the grinding process involves continuous feed, control compensation is converted into feed rate override commands or specific grinding stage switching commands. For example, when the prediction error indicates that the workpiece size is close to the target value, the system sends a command to reduce the feed rate or triggers a signal to switch from the rough grinding stage to the finish grinding stage, directly intervening in the current machining cycle.

[0110] For turning processes (such as roller turning), the strategy distribution unit executes an inter-batch compensation strategy. Since the turning process is usually completed in one pass, and tool wear is the main cause of dimensional drift, the control compensation amount is converted into a tool offset compensation value. This compensation value includes fine-tuning amounts for the X-axis (radial) and Z-axis (axial).

[0111] The strategy distribution unit writes the compensation value into the tool compensation register of the machine tool's CNC system. This compensation operation does not change the trajectory of the currently being machined workpiece, but takes effect when the machine tool is machining the next workpiece, thereby correcting the machining trajectory and offsetting dimensional deviations caused by tool wear.

[0112] In addition, the strategy distribution unit has a safety limiting logic. This logic sets a safety threshold based on the machine tool's physical travel limit and maximum allowable feed speed (e.g., maximum single compensation amount ≤ 0.05mm). If the calculated control compensation amount exceeds this threshold, the system will automatically cut off the output and trigger an alarm to prevent equipment collisions or workpiece scrap due to algorithm malfunctions.

[0113] The self-evolution module 600 is used to perform statistical analysis on the quality of the processed finished product and feed back the verified process data to the system's knowledge base to achieve iterative optimization of the knowledge graph. This self-evolution module 600 mainly includes a quality analysis unit and a knowledge management unit.

[0114] The quality analysis unit is responsible for receiving measurement data from measuring instruments. This data is acquired through offline inspection of the machined workpiece using high-precision measuring equipment in a temperature-controlled precision measuring chamber. For measuring the roundness and waviness of bearing rings or rollers, a high-precision roundness meter (e.g., the Taylor Hopson Talyrond 565 PRO) is used; for measuring contour dimensions, a profilometer or coordinate measuring machine is used. The measuring equipment transmits the measurement data to the quality analysis unit through the company's internal quality management system interface.

[0115] The quality analysis unit uses the SPC (Statistical Process Control) algorithm to perform statistical calculations on the distribution of measurement data. Based on the measurement data, the algorithm calculates the process capability index (Cpk) and standard deviation to assess the stability of the current process in actual production.

[0116] During this process, the system sets process capability judgment thresholds. These thresholds are pre-set based on the ISO 22514 statistical process control standard and the company's internal quality level requirements (for example, setting Cpk ≥ 1.33 as sufficient process capability and Cpk < 1.0 as insufficient process capability).

[0117] When the calculated process capability index exceeds the process capability judgment threshold, the system determines the current batch processing as high-quality processing; when the calculated process capability index is lower than the process capability judgment threshold, it is judged as having risk or failure. The quality analysis unit finally outputs the evaluation results, which include statistical indicators and quality judgment labels.

[0118] The knowledge management unit receives the evaluation results, the process recommendation scheme generated by the preceding process recommendation module, and the prediction error output by the quality prediction module. This unit uses an update algorithm to perform a graph update operation. When the evaluation result shows high-quality processing, the update algorithm extracts key process parameters (including grinding wheel linear speed, workpiece rotation speed, and feed rate) and the corresponding workpiece material and equipment model from the process recommendation scheme, constructing a successful case node. This successful case node is inserted into the process graph, and applicable association edges are established between this node and specific material nodes and equipment nodes. At the same time, the update algorithm searches for existing similar paths in the graph and increases the weight value of the process path in subsequent graph searches (e.g., increasing the confidence parameter of the edge weight by 5%).

[0119] When the assessment results indicate a risk or failure, the knowledge management unit performs correlation analysis based on the prediction error. If the deviation between the prediction error and the actual measurement data is within the preset allowable range, it indicates that the model prediction of the quality prediction module is accurate, but the process parameters themselves are improperly set. The system marks the combination of process parameters for this processing as a failure case node and inserts it into the process diagram to establish a correlation edge that leads to the defect. This correlation edge contains specific defect type attributes (such as out-of-tolerance roundness or grinding burns).

[0120] The failed node is marked as a negative sample in subsequent rule validation or hybrid retrieval to filter out similar erroneous paths. If the prediction error differs from the actual measurement data, the system marks the case as a model bias case and stores the relevant data in the log database for subsequent fine-tuning training of the LSTM network in the quality prediction module.

[0121] The knowledge management unit outputs an updated graph. This updated graph includes newly added case nodes and adjusted relationships, and is fed back to the graph construction module 100. The graph construction module 100 reconstructs the index based on the updated graph, enabling the system to prioritize and recommend high-quality parameters that have been verified in actual production based on the updated weights when handling similar user requests, or to automatically avoid known risky paths based on failed association edges.

[0122] This invention provides a knowledge graph-based intelligent bearing machining process recommendation system. Through a collaborative architecture of cloud, edge, and terminal, it achieves intelligent design and dynamic closed-loop control of bearing machining processes. The system utilizes a knowledge graph construction module 100 to transform scattered unstructured technical documents and structured databases into a unified process graph, establishing end-to-end knowledge associations. Based on this, a process recommendation module 200, combining hybrid retrieval and rule-based validation mechanisms, automatically generates process recommendation schemes containing process sequences and equipment parameters according to user needs, solving the problem of traditional process design relying on manual experience.

[0123] During actual machining, the data sensing module 300 collects the physical environment status in real time through multi-dimensional sensors and uses time-series alignment technology based on workpiece ID to provide accurate and synchronized feature vectors for subsequent analysis. The quality prediction module 400 uses a deep learning model to capture the time-varying characteristics of the machining process and predicts the dimensional deviations of the workpiece in advance. The feedback control module 500 calculates the compensation amount based on the predicted deviation and performs tool offset compensation or feed rate adjustment for turning or grinding processes respectively, achieving millisecond-level real-time correction.

[0124] After processing, the self-evolution module 600 uses the SPC algorithm to analyze the finished product measurement data, automatically identify high-quality or failure cases, and update the process diagram. This system constructs a complete closed loop of pre-recommendation, in-process control, and post-iteration, effectively improving the process design efficiency and processing quality stability of bearing manufacturing.

Claims

1. A knowledge graph-based intelligent bearing machining process recommendation system, characterized in that, include: The graph construction module (100) is used to integrate the received technical documents and structural data, extract and fuse them using a large language model and alignment algorithm, and construct a process graph. The process recommendation module (200) is used to search the process map according to user needs, generate a process recommendation scheme through hybrid search and rule verification, and parse the target value in the process recommendation scheme. The data sensing module (300) is used to collect environmental status and perform data timing alignment and feature extraction based on the workpiece ID output by the machine tool control system to generate a feature vector; A quality prediction module (400) is used to perform temporal modeling and error mapping on the feature vector through an LSTM network, thereby generating and outputting prediction error; The feedback control module (500) is used to calculate the control deviation of the prediction error relative to the target value, use the PID algorithm to solve the compensation amount and generate control commands to drive the machine tool to move. The self-evolution module (600) is used to combine measurement data, the recommended process scheme and the prediction error to perform quality analysis and knowledge management, and generate an updated map to be fed back to the map construction module (100). The process recommendation module (200) includes: The requirement parsing unit is used to receive the user requirements from the user interface, identify the entities and intentions in the user requirements using semantic algorithms, and extract constraint parameters. A hybrid retrieval unit is used to receive the constraint parameters and the process diagram, and to perform vector-based semantic similarity retrieval and graph-based adjacency traversal retrieval using retrieval algorithms to recall similar historical case nodes and associated standard parameter nodes, and to construct candidate subgraphs. The process recommendation module (200) also includes: The rule verification unit is used to receive the candidate subgraph and preset hard rules, use the rule engine to perform logical verification and pruning on the process paths in the candidate subgraph, and output compliant paths. The scheme generation unit is used to receive the compliance path and the constraint parameters, integrate structured information using the generation model and fill in specific parameter suggestions to generate the process recommendation scheme, and extract the target value from the process recommendation scheme; The aforementioned rigid rules are pre-set based on the physical limit formulas and safety production standards in the field of bearing processing.

2. The bearing intelligent machining process recommendation system based on knowledge graph as described in claim 1, characterized in that, The map construction module (100) includes: An extraction unit is used to receive the externally imported technical document, perform semantic parsing and entity relation extraction on the technical document using the large language model, and output the original triples; The alignment and fusion unit is used to receive the externally imported structural data and the original triplet, extract the entities contained in the structural data and the original triplet, map the entities into high-dimensional vectors using an embedding model, calculate the cosine similarity using a similarity formula using a vector algorithm, and when the cosine similarity exceeds the alignment threshold, use the alignment algorithm to determine that entities with different names are the same object and perform attribute merging to output a unique entity. The graph storage unit is used to receive the unique entity and its association, and to use the graph database to construct and index nodes and edges according to a predefined graph pattern to generate the process graph. The alignment threshold is pre-set based on the principle of maximizing F1 scores on a large-scale labeled dataset, and the predefined graph pattern defines the node type and edge type in the graph database.

3. The bearing intelligent machining process recommendation system based on knowledge graph as described in claim 1, characterized in that, The data sensing module (300) includes: The data acquisition unit is used to convert physical quantities in the environmental state of the processing site into electrical signals using displacement sensors, strain sensors and acoustic emission sensors, and output them as raw signals. The timing alignment unit is used to receive the original signal and the workpiece ID, and uses an alignment algorithm with a sliding window mechanism to extract data within the corresponding processing cycle using the workpiece ID as an index, and outputs the alignment data.

4. The bearing intelligent machining process recommendation system based on knowledge graph according to claim 3, characterized in that, The data sensing module (300) also includes: The feature extraction unit is used to receive the alignment data, perform wavelet denoising and transformation processing on the alignment data using a feature algorithm, perform time-domain statistics and frequency-domain analysis, and extract time-frequency features to construct the feature vector.

5. The bearing intelligent machining process recommendation system based on knowledge graph according to claim 1, characterized in that, The quality prediction module (400) includes: The temporal modeling unit is used to receive the feature vector, use the LSTM network to learn the temporal features in combination with the internal state of the previous time step, and use the state update formula to calculate and output the hidden state. An error mapping unit is used to receive the hidden state, perform nonlinear mapping and regression analysis on the hidden state using a fully connected layer network structure, and output the prediction error.

6. The bearing intelligent machining process recommendation system based on knowledge graph according to claim 1, characterized in that, The feedback control module (500) includes: A deviation calculation unit is used to receive the prediction error and the target value, perform difference calculation using a subtractor, and output the control deviation. The compensation calculation unit is used to receive the control deviation, calculate and output the compensation amount using the compensation formula and the PID algorithm.

7. The bearing intelligent machining process recommendation system based on knowledge graph according to claim 6, characterized in that, The feedback control module (500) further includes: A strategy distribution unit is used to receive the compensation amount, convert the compensation amount into a control command that the machine tool can recognize according to the current machining process type using a communication program, and output the control command to drive the machine tool. Wherein, when the current machining process type is turning, the control command is a tool offset compensation value; when the current machining process type is grinding, the control command is a feed rate command.

8. The bearing intelligent machining process recommendation system based on knowledge graph according to claim 1, characterized in that, The self-evolution module (600) includes: The quality analysis unit is used to receive the measurement data from the measuring instrument, perform statistical calculations on the distribution of the measurement data using the SPC algorithm, and output the evaluation results. The knowledge management unit is used to receive the evaluation results, the recommended process scheme, and the prediction error, generate case nodes based on the evaluation results using an update algorithm, insert them into the process map to establish new associations, and output the updated map.

Images