Machining fault early warning method and system based on digital twinning

Abstract

The invention relates to a machining fault early warning method and system based on digital twinning. The method comprises the following steps: acquiring multi-modal data of interaction between a cutter and a workpiece in a machining process; the multi-modal data comprises real-time process parameters, temperature data, vibration data and current data; preprocessing the multi-modal data, and obtaining time sequence data based on timestamp synchronization; inputting the time sequence data into a digital twinning model to obtain a twinning state variable; inputting the time sequence data and the twinborn state variable into the depth migration inference model to obtain a prediction result; the prediction result comprises a prediction fault label and a probability value; and triggering a corresponding fault early warning instruction and a tool adjustment strategy according to the prediction result. By adopting the method, high-precision and high-reliability fault prediction and early warning with real-time response capability can be realized in the machining process through organic fusion of digital twinning, deep migration reasoning and intelligent control feedback, and the operation safety is improved.

Description

Technical Field

[0001] The present invention belongs to the field of data processing, and in particular relates to a mechanical processing fault early warning method and system based on digital twins. Background Art

[0002] With the development of intelligent manufacturing technology, equipment operation status perception and intelligent maintenance technology have emerged. By deploying a variety of sensors, the physical data of processing equipment during operation is collected in real time, and with the help of edge computing and data processing algorithms, the health status of equipment is analyzed and managed. Among them, fault identification and has been widely used in typical industrial scenarios such as CNC machine tools, robots, and conveying systems.

[0003] In traditional technology systems, monitoring the operating status of machining equipment typically utilizes threshold-based alarm mechanisms or rule-based knowledge-based diagnostic methods. The system sets a fixed temperature, vibration amplitude, or current threshold. When the real-time signal exceeds the threshold, an alarm is triggered or the equipment shuts down for protection. In addition, some expert systems based on empirical rules exist to perform pattern recognition and judgment of typical fault types. Some high-end systems also attempt to infer equipment status through the establishment of mathematical or physical models, but these typically have limited accuracy and are difficult to adapt to complex operating conditions and individual changes.

[0004] However, traditional methods use static and fixed threshold settings, which cannot adapt to processing tasks and environmental changes, resulting in frequent false alarms or missed alarms. They rely on responding only after an anomaly has occurred or is approaching a critical point, and lack the ability to provide early warning. Methods based on rules or mathematical models often rely on expert experience, have difficulty modeling complex physical behaviors, and have poor generalization capabilities. Even if a fault is detected, traditional systems are often unable to quickly coordinate process parameter adjustments to achieve a true closed loop of prediction, control, and optimization. Summary of the Invention

[0005] Based on this, it is necessary to provide a digital twin-based mechanical processing fault warning method and system that can integrate multi-source perception support prediction and decision feedback to address the above technical problems.

[0006] In a first aspect, the present application provides a machining fault early warning method based on digital twins, comprising:

[0007] Acquire multimodal data of the interaction between the tool and the workpiece during machining; the multimodal data includes real-time process parameters, temperature data, vibration data, and current data;

[0008] Preprocess multimodal data and obtain time series data based on timestamp synchronization;

[0009] Input time series data into the digital twin model to obtain twin state variables;

[0010] Input the time series data and twin state variables into the deep migration inference model to obtain the prediction results; the prediction results include the predicted fault label and probability value;

[0011] The corresponding fault warning instructions and tool adjustment strategies are triggered according to the prediction results.

[0012] In one embodiment, the digital twin model is constructed by:

[0013] Obtaining structural data, process path planning data and tool attribute parameters of mechanical machine tools;

[0014] Construct basic geometric models by analyzing structural data;

[0015] Importing process path planning data into the basic geometric model as initial boundary conditions to obtain a process path grid structure model; the process path grid structure model includes multiple process nodes;

[0016] Bind the tool attribute parameters to the process path grid structure model to obtain a digital twin model.

[0017] In one embodiment, the digital twin model obtains the twin state variables by the following method, including:

[0018] Projecting real-time process parameters onto corresponding process nodes in the digital twin model to obtain a multi-physics simulation boundary, which includes cutting heat sources, vibration excitation sources, and charge loads.

[0019] Based on historical risk factors, temperature data, vibration data and current data are subjected to finite source simulation according to the multi-physics field simulation boundary to obtain twin state variables; finite element simulation includes thermal field simulation, vibration field simulation and electric power field simulation.

[0020] In one embodiment, temperature data, vibration data, and current data are subjected to finite source simulation according to a multi-physics field simulation boundary to obtain twin state variables, including:

[0021] Conduct thermal field simulation of temperature conduction through temperature data to obtain temperature field distribution;

[0022] Extract the maximum temperature rise rate of the temperature field distribution and obtain the heat load factor;

[0023] Perform structural dynamics simulation of vibration based on vibration data to obtain stress field distribution; stress field distribution includes displacement and velocity;

[0024] Extract the resonance peak amplitude and energy dissipation ratio of the stress field distribution to obtain the vibration index;

[0025] Perform power field simulation based on current data to obtain power field distribution; power field distribution includes power and energy consumption density;

[0026] Extract the current mutation ratio of the power field distribution and obtain the current anomaly factor;

[0027] The thermal load factor, vibration index and current abnormality factor are integrated to obtain the twin state variables;

[0028] The temperature field distribution is obtained through the following formula:

[0029]

[0030] Where ρ is the material density; c is the material specific heat capacity; is the temperature field distribution; k is the thermal conductivity; is the temperature gradient obtained from the temperature data; Q I It is the cutting heat source;

[0031] The stress field distribution is obtained by the following formula:

[0032]

[0033] Where M is the mass matrix; is the acceleration; C is the damping matrix; is velocity; K is stiffness matrix; u is displacement; F cut (t) is the vibration excitation source;

[0034] The power field distribution is obtained through the following formula:

[0035] P(t)=U(t)·I(t)=F cut (t)·v(t)+P loss

[0036]

[0037] Among them, P(t) is power; U is voltage; I is current data; F cut is the vibration excitation source; v(t) is the instantaneous linear velocity of the tool; P loss is the no-load power loss; q joule is the energy consumption density; σ is the conductivity; E is the electric field intensity; R is the equivalent resistance of the main axis; and V is the current volume.

[0038] In one embodiment, the deep migration inference model obtains prediction results by the following method, including:

[0039] The time series data is encoded at multiple scales by stacking coefficient autoencoding layers, and synchronously spliced and fused with the twin state variables to obtain a high-dimensional feature representation vector.

[0040] The high-dimensional feature representation vector is mapped to the discriminant space through the migration adaptation layer to obtain the predicted label and the corresponding probability value; the predicted labels include normal, slight wear, severe wear and chipping.

[0041] In one embodiment, triggering corresponding fault warning instructions and tool adjustment strategies based on the prediction results includes:

[0042] The warning response level is obtained based on the probability value; the warning response levels include green, yellow and red;

[0043] Warning instructions and tool adjustment strategies are generated based on the warning response level and prediction tags. Warning instructions are used to instruct the CNC machine tool to send reminder information to the user terminal according to the warning response level. Tool adjustment strategies include feed rate, spindle speed, cutting depth and coolant pressure.

[0044] In one embodiment, the method further comprises:

[0045] In response to the obtained prediction result confirmation instruction corresponding to the twin state variable, the historical risk factor is updated using the following formula:

[0046] R new =ω·R old +(1-ω)·R obs

[0047] Among them, R new is the new historical risk factor; R old Historical risk factor; R obs is the prediction result corresponding to the twin state variable; ω is the weight factor.

[0048] In a second aspect, the present application also provides a machining fault early warning system based on digital twins, comprising:

[0049] Physical perception module, used to obtain multimodal data of the interaction between tool and workpiece during machining;

[0050] The data synchronization processing module is used to pre-process the multimodal data and obtain time series data based on timestamp synchronization;

[0051] The digital twin module is used to input time series data into the digital twin model to obtain twin state variables;

[0052] The fault prediction module is used to input time series data and twin state variables into the deep migration inference model to obtain prediction results;

[0053] The fault warning module is used to trigger corresponding fault warning instructions and tool adjustment strategies based on the prediction results.

[0054] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of any of the above-mentioned digital twin-based machining fault warning methods.

[0055] In a fourth aspect, the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of any of the above-mentioned digital twin-based machining fault warning methods.

[0056] The digital twin-based machining fault warning method and system collects multimodal data, including temperature, vibration, current, and process parameters, covering the main energy transfer and dynamic change mechanisms during machining. It provides rich sensory information and strong state representation capabilities. Unified preprocessing and timestamp alignment effectively eliminate information bias caused by asynchronous data from different sensors, ensuring the consistency and accuracy of multimodal feature fusion. The digital twin model generates internal physical state variables, addressing the shortcomings of traditional methods that rely solely on surface data and fail to capture underlying evolutionary trends. This significantly improves prediction lead time and fault sensitivity. Deep transfer learning methods address the domain shift between twin simulation data and real-world machining data, enabling training to rely primarily on simulation data while enabling accurate adaptation to field data, significantly enhancing the model's generalizability and practicality. Prediction results not only provide alarms but also directly drive the CNC system to adjust feed rate, spindle speed, depth of cut, and coolant pressure, achieving prediction-feedback-optimization closed-loop control and improving overall machining safety and equipment protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0058] Figure 1 Schematic diagram of the process of the digital twin-based mechanical processing fault early warning method of the present invention;

[0059] Figure 2 A schematic diagram of the process of constructing the digital twin model of the present invention;

[0060] Figure 3 Schematic diagram of the step-by-step process of step S103;

[0061] Figure 4 This is a structural diagram of the mechanical processing fault warning system based on digital twins of the present invention. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0063] In one embodiment, Figure 1 As shown, a method for early warning of mechanical processing faults based on digital twins is provided. This embodiment uses the method applied to a terminal as an example. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0064] S101 , acquiring multimodal data of the interaction between the tool and the workpiece during machining; the multimodal data includes real-time process parameters, temperature data, vibration data, and current data.

[0065] Multimodal data are sensor signals generated by different physical mechanisms during machining and reflecting different operating states. They include real-time process parameters, temperature data, vibration data, and current data. Real-time process parameters primarily include spindle speed, feed rate, and depth of cut, which can be directly output by the numerical control system (CNC) and collected via a PLC bus or industrial Ethernet. Temperature data can be collected using K-type thermocouples, which are attached to the tool edge or the workpiece cutting surface. The sampling period can be controlled within 50 to 100 milliseconds. Vibration data is acquired in real time by a three-axis accelerometer mounted on the spindle or toolholder, reflecting the structural response and impact characteristics during machining. Current data is derived from the spindle motor's current sensor module and is primarily used to reflect fluctuations in cutting load and abnormal power consumption.

[0066] S102: Preprocess the multimodal data and obtain time series data based on timestamp synchronization.

[0067] Schematically, the data sampling frequency, starting time and time accuracy of various sensors often differ, and preprocessing is required. Specifically, each sensor data is denoised and filtered using sliding average, wavelet transform or Kalman filtering to eliminate environmental interference and transient spikes; data of different dimensions are standardized through Z-score normalization or Min-Max linear mapping to unify the scale for convenient model input; further, based on millisecond timestamps, the timing alignment of multiple signals is achieved through interpolation or sliding window matching to construct a set of continuous three-channel time series tensors, where each time step corresponds to a temperature value, a vibration amplitude and a current value.

[0068] S103. Input the time series data into the digital twin model to obtain the twin state variables.

[0069] Schematically, based on the CAD drawings and processing path files provided by CNC, the spatial geometric structure and motion trajectory of the tool, workpiece and fixture are reconstructed; then, according to the thermal conductivity, elastic modulus, damping ratio and other properties of the processed material, the thermal-mechanical-electrical coupling characteristics of the structure are defined, and the real-time sensor data is bound as the driving source to perform physical field mapping of the data twin model. Specifically, the digital twin model is a multi-physics field simulation system constructed on the basis of a three-dimensional structural model, integrating process path planning and material physical properties. By simulating and solving the heat conduction equation, structural vibration equation and electric power distribution model, the actual collected boundary data is mapped to the physical field response in the virtual space, thereby obtaining deep-level state indicators that traditional sensors cannot directly provide. Specifically, a finite element solver is used to numerically solve the thermal field, vibration field and electric field to obtain state variables including heat load factor, vibration intensity factor and current fluctuation factor to reflect phenomena such as heat accumulation, modal instability or load mutation.

[0070] S104. Input the time series data and twin state variables into the deep migration inference model to obtain a prediction result; the prediction result includes a predicted fault label and a probability value.

[0071] The inference model is based on the stacked sparse autoencoder (SSAE) structure, which realizes deep feature extraction of time series data through multi-layer nonlinear coding, and realizes feature alignment of source and target domains through the maximum mean difference (MMD) method in the middle layer, thereby enhancing the adaptability of the model under actual processing tasks. Specifically, the currently acquired time series and the twin state variables are spliced into a unified feature vector, and the fault labels are output after model encoding. For example, they include mild wear, chipping risk, strong earthquake instability, etc. and their corresponding probability values. For example, the tool chipping risk label is predicted in a milling process with a confidence level of 87%, which means that the model judges that the current processing state is very likely to have entered the chipping precursor stage.

[0072] S105: triggering corresponding fault warning instructions and tool adjustment strategies according to the prediction results.

[0073] Based on the predicted results, the corresponding fault warning instructions and tool adjustment strategies are automatically triggered, thus achieving a closed-loop operation of prediction, control, and correction. For example, if the predicted probability is below the set threshold, normal operation is maintained. If it is in the intermediate range, a minor adjustment strategy is triggered, for example, reducing the feed rate by 10% or the cutting depth by 15% to alleviate the machining load. If the predicted probability exceeds the high-risk threshold, an alarm signal is immediately issued, and manual maintenance or tool replacement is recommended, and the machine tool operation can be suspended simultaneously.

[0074] The aforementioned digital twin-based machining fault early warning method captures multimodal data from tool-workpiece interaction during machining, including real-time process parameters, temperature, vibration, and current data. This method comprehensively covers the three primary physical mechanisms of change during machining: thermal, force, and electrical energy. This ensures multidimensional and high-integrity state perception, strengthening the fundamental data support for fault detection and trend prediction. Preprocessing and timestamp-based synchronization of multimodal data effectively addresses heterogeneity in sampling frequency, start and end times, and scale dimensions among different sensor data, ensuring temporal consistency and scale uniformity of the input data, significantly improving the accuracy and robustness of subsequent modeling and inference. By inputting this preprocessed time series data into the digital twin model, the thermal, force, and electrical fields within the machining system can be inferred through multiphysics simulation based on the actual structure and material properties. This generates twin state variables that are difficult to directly observe using traditional sensors. These twin state variables provide a deep understanding of the physical evolution trends during machining, enhancing the ability to detect latent faults early on. By synchronously inputting time series data and twin state variables into the deep transfer inference model, high-dimensional feature fusion and transfer learning mechanisms can effectively adapt to distribution shift issues in actual machining environments, leveraging the advantages of digital twin simulation data training. This improves the model's inference accuracy and generalization performance in real-world scenarios. Based on the predicted results obtained through inference, corresponding fault warning instructions and tool adjustment strategies can be automatically triggered, implementing a closed-loop control mechanism from fault trend identification to adaptive optimization of process parameters. This coordinated control not only delays fault development and reduces equipment damage, but also improves the overall stability and production efficiency of the machining process, reducing tool and equipment maintenance costs. By organically integrating digital twins, deep transfer inference, and intelligent control feedback, high-precision, high-reliability, and real-time responsive fault prediction and warning can be achieved in the machining field, significantly improving the level of intelligent machining and operational safety.

[0075] In one embodiment, Figure 2 As shown in the figure, the digital twin model is constructed by the following methods:

[0076] S201: Acquire structural data, process path planning data, and tool attribute parameters of a mechanical machine tool.

[0077] Schematically, the structural data of a mechanical machine tool can be a CAD three-dimensional drawing of the machine tool body in a format such as STEP, IGES or STL, which includes the dimensional parameters and geometric topology information of parts such as the machine tool foundation, spindle system, fixture device and worktable; the process path planning data is the program sequence that represents the tool trajectory and operation logic during the processing process, which usually comes from the G code generated by the CNC controller or the path form exported by the CAM system, which includes dynamic process information such as feed rate, cutting depth, spindle speed and processing sequence; the tool attribute parameters, including tool type, diameter, material, chip groove characteristics, cooling type, etc., determine the physical response characteristics of the cutting area.

[0078] Optionally, data can be acquired through the machine tool manufacturer's open interface, a digital master model parser, or manually imported by an operator through a parametric modeling platform.

[0079] S202: Construct a basic geometric model by analyzing the structural data.

[0080] By analyzing the structural data, a basic geometric model is constructed, which serves as the static skeleton of the twin and is the boundary basis for the operation simulation of the digital twin. Schematically, the CAD file is read through the three-dimensional geometry engine, the topological relationship between the components is identified, and the solid voxel grid is reconstructed in the form of finite element modeling to form a three-dimensional space domain with simulation characteristics. Furthermore, in order to improve the stability and accuracy of subsequent simulations, mesh refinement can be performed on key areas such as the tool-workpiece contact area to retain the motion constraints between key structural levels and linkage components. For example, for a five-axis linkage machine tool, the matching mechanisms of the rotating A-axis, B-axis and spindle Z-axis are modeled respectively, and a relative coordinate reference system is set for the tool clamping position to ensure that the path simulation is aligned with the physical one.

[0081] S203 , importing the process path planning data into the basic geometric model as initial boundary conditions to obtain a process path grid structure model; the process path grid structure model includes multiple process nodes.

[0082] The process path grid structure refers to the distribution of path segments and control points in accordance with the processing sequence on the static geometric skeleton to form a spatiotemporal driving framework for the processing process. The model consists of multiple process nodes, each of which corresponds to a processing position, a set of process parameters and a timestamp. Node information includes spindle speed, tool feed rate, current tool position coordinates and processing methods, including rough milling, fine milling, drilling, etc., which can be regarded as a scheduling unit for twin operation. During simulation execution, physical solutions will be performed between these nodes, and the thermal field, force field and electromagnetic field states will be dynamically updated accordingly. For example, in a contour milling task, the node path will reflect the changing process of the tool in and out of the cut, drive the time distribution of the twin heat source term, and thus reflect the true response of the temperature rise trend.

[0083] S204. Bind the tool attribute parameters to the process path grid structure model to obtain a digital twin model.

[0084] Tool attribute parameters include tool geometry, such as the area of influence of the cutting edge width, and tool material properties, namely thermal conductivity, elastic modulus, resistivity, etc. Tool attribute parameters are introduced into the physical simulation equations as important parameters affecting the evolution of thermal-mechanical-electrical behavior. Optionally, characteristics such as cooling method and coating type can also be used as dynamic boundary conditions in the model to affect the simulation boundary. For example, the difference in heat flow boundaries between dry cutting and wet cutting can be simulated. By completing the attribute binding, a high-fidelity digital twin model with real geometric structure, machining path scheduling and material physical characteristics is obtained, which can support subsequent multi-physics field state variable solution and risk factor prediction tasks.

[0085] In one embodiment, the digital twin model obtains the twin state variables by the following method, including:

[0086] S31. Project the real-time process parameters onto the corresponding process nodes in the digital twin model to obtain a multi-physics simulation boundary; the multi-physics simulation boundary includes a cutting heat source, a vibration excitation source, and a charge load.

[0087] Schematically, real-time process parameters are projected onto corresponding process nodes in the digital twin model, thereby constructing a multi-physics simulation boundary for the machining process. Specifically, real-time process parameters include spindle speed, feed rate, depth of cut, tool radius compensation, and cooling method, reflecting the current operating configuration of the machining tool. The digital twin model contains a pre-set process path grid structure, in which each process node records the spatial position and machining action at the corresponding time point. Real-time process parameters are aligned to the current machining node through timestamp matching, and the physical excitation conditions associated with each node are set accordingly, forming three types of simulation boundaries. These include, schematically, a cutting heat source boundary, which estimates the cutting power per unit time and allocates the equivalent heat source intensity to the contact area between the tool and the workpiece; a vibration excitation source boundary, which defines the vibration excitation function based on the discrete derivative of the tool feed path and the structural frequency spectrum; and a charge load boundary, which infers the spindle load fluctuations from the current measurement data and injects them into the simulation area as electrical power density. For example, a hole machining path is simulated, where the feed rate at process node P1 is 800 mm / min and the spindle speed is 6000 rpm. After calculation by the power model, a 95 W heat source term is applied to the corresponding tool contact surface, and a transient force function with a frequency of 280 Hz is applied in the structural excitation module.

[0088] S32. Based on historical risk factors, temperature data, vibration data and current data are subjected to finite source simulation according to the multi-physics field simulation boundary to obtain twin state variables; finite element simulation includes thermal field simulation, vibration field simulation and electric power field simulation.

[0089] Schematically, in the thermal field simulation, the non-steady-state heat conduction equation is used as the solution model. After the finite element discrete solution, the temperature distribution map of the tool-workpiece contact area is obtained, and the maximum temperature rise gradient, heat accumulation rate and other indicators are further extracted to form a thermal load factor that expresses the thermal risk level. This factor can be used to measure the thermal wear trend and local overheating risk.

[0090] In vibration field simulation, the second-order vibration equations of structural dynamics are used, and the process path excitation function and the structural stiffness matrix are input into the simulation solver. Through time-domain deconstruction and spectral analysis of the modal response, the presence of near-resonant modes and nonlinear vibration during machining can be identified. Characteristics such as peak response amplitude and frequency shift are calculated to generate a vibration risk factor reflecting dynamic stability. A vibration risk factor value close to 1 indicates that the machining tool is in the critical modal region, making it highly susceptible to failures such as tool chipping.

[0091] In the electric power field simulation, real-time current data is mapped as a spindle load function and input into the electric power model along with material resistivity and motor parameters to perform power flow distribution and Joule heating calculations. By measuring indicators such as current slope per unit time, peak power fluctuation, and transient energy density, the current fluctuation factor, which characterizes energy input stability, is extracted. This factor reflects load anomalies or tool gnawing during the cutting process and is crucial for detecting load fluctuation-related faults.

[0092] In one embodiment, temperature data, vibration data, and current data are subjected to finite source simulation according to a multi-physics field simulation boundary to obtain twin state variables, including:

[0093] S41. Perform thermal field simulation of temperature conduction using temperature data to obtain temperature field distribution.

[0094] The temperature field distribution is obtained through the following formula:

[0095]

[0096] Where ρ is the material density; c is the material specific heat capacity; is the temperature field distribution; k is the thermal conductivity; is the temperature gradient obtained from the temperature data; Q I It is the cutting heat source;

[0097] In schematic form, ρ represents the material density. When the machining fault is a warning of tool body heating or vibration, the material density is the tool material density. When the machining fault is a warning of a fault in the machining area or workpiece heating area, the material density is the workpiece material density. c represents the specific heat capacity of the material. represents the temperature field distribution function to be solved, which characterizes the temperature rise rate over time; k represents the thermal conductivity, that is, the heat transfer ability of the material, and is the temperature gradient, which comes from the spatial distribution measured by the thermocouple sensor; Q I is the heat source intensity per unit volume, primarily derived from the heat input to the tool cutting zone, and can be derived by inverse calculation from the actual power or cutting parameters. By numerically solving the partial differential equation using finite element methods, a temporal and spatial distribution of the temperature near the tool-workpiece interface can be obtained.

[0098] S42. Extract the maximum temperature rise rate of the temperature field distribution and obtain the heat load factor.

[0099] The thermal field simulation results were analyzed, and the maximum temperature rise rate of the temperature field in the cutting area was extracted. After normalization, it was defined as the thermal load factor. This factor is used to measure the thermal shock capacity of the tool per unit time. The higher the value, the more likely the current thermal environment is to cause tool wear, thermal fatigue, or surface quality deterioration.

[0100] S43. Perform structural dynamics simulation of vibration based on vibration data to obtain stress field distribution; the stress field distribution includes displacement and velocity.

[0101] The stress field distribution is obtained by the following formula:

[0102]

[0103] Where M is the mass matrix; is the acceleration; C is the damping matrix; is velocity; K is stiffness matrix; u is displacement; F cut (t) is the vibration excitation source;

[0104] Among them, M is the mass matrix, i.e. the inertia distribution of the tool structure in all directions; C is the damping matrix, i.e. the energy dissipation such as internal friction of the material and structural energy consumption; K is the stiffness matrix, which reflects the structural deformation resistance; F cut (t) is the vibration excitation source function. The cutting force generated by the contact between the tool and the workpiece changes with time. Its value is calculated by the machining path derivative and the vibration sensor data. By solving the mode of the motion system, the vibration response characteristics and resonant energy distribution at different frequencies can be obtained.

[0105] S44. Extract the resonance peak amplitude and energy dissipation ratio of the stress field distribution to obtain the vibration index.

[0106] The maximum response amplitude corresponding to the resonant frequency is extracted in the spectral domain, and combined with the vibration energy attenuation rate, the vibration index is calculated. Specifically, when the vibration index approaches or exceeds 1, it usually indicates that the structure is in a metastable state or at the resonance boundary, and there are processing risks such as chipping and tool jumping.

[0107] S45. Perform power field simulation based on the current data to obtain power field distribution; the power field distribution includes power and energy consumption density.

[0108] P(t)=U(t)·I(t)=F cut (t)·v(t)+P loss

[0109]

[0110] Among them, P(t) is power; U is voltage; I is current data; F cut is the vibration excitation source; v(t) is the instantaneous linear velocity of the tool; P loss is the no-load power loss; q joule is the energy consumption density; σ is the conductivity; E is the electric field intensity; R is the equivalent resistance of the main axis; and V is the current volume.

[0111] Schematically, calculate the power field distribution F cut It can be obtained from the force exerted by the tool on the workpiece, P loss For the energy that does not participate in cutting but is consumed, the voltage and current data can be the working output of the spindle motor.

[0112] S46. Extract the current mutation ratio of the power field distribution to obtain the current anomaly factor.

[0113] High-frequency sudden changes or peak jumps in current can significantly increase the value of the current abnormality factor, which is often closely related to tool edge breakage, tool gnawing, overload instability and other faults.

[0114] S47. The heat load factor, the vibration index, and the current abnormality factor are integrated to obtain the twin state variables.

[0115] In one embodiment, Figure 3 As shown in the figure, the deep transfer inference model obtains prediction results through the following methods, including:

[0116] S301. Perform multi-scale time series encoding on the time series data through the stacked coefficient self-encoding layer, and synchronize it with the twin state variables to obtain a high-dimensional feature representation vector.

[0117] The time series data obtained through multimodal perception and twin modeling are input into the stacked sparse autoencoder layer for multi-scale time series encoding. Specifically, a tensor representation is constructed for the three-channel signal sequences of temperature, vibration, and current in each fixed time window during the processing process, and input into a multi-layer sparse autoencoder network. The sparse autoencoder is an unsupervised feature compression model that learns the potential structure of the input data by reconstructing the objective function and sparsity constraints to obtain high-order feature expressions. It can effectively extract the frequency characteristics, trend changes, mutation point information, etc. contained in the time series, and eliminate the high-dimensional redundant noise in the original data. Furthermore, the twin state variables of this section of the processing process are obtained from the digital twin simulation module, and they are synchronously spliced with the time series features output by the autoencoder layer to form a unified high-dimensional feature representation vector, thereby realizing physically enhanced feature expression.

[0118] S302 , mapping the high-dimensional feature representation vector to the discriminant space through the migration adaptation layer to obtain prediction labels and corresponding probability values; the prediction labels include normal, slight wear, severe wear and chipping.

[0119] The high-dimensional feature representation vector is sent to the transfer adaptation layer, and the projection mapping from the feature space to the discriminant space is completed in this layer. The transfer adaptation layer adopts the domain alignment mechanism based on Maximum Mean Discrepancy (MMD). During the training process, Minimize the distribution difference between the source domain and the target domain in the latent space, that is, the gap between the large amount of virtual processing data simulated by the digital twin model and the actual data, and output the distribution offset of the data in the high-dimensional space, so as to achieve consistent mapping of the feature space and ensure the accuracy of fault prediction. Specifically, this layer not only performs the classification transformation operation in the traditional neural network, but also dynamically adjusts the network parameters during the training phase, so that the output distribution remains consistent when different data sources are input, preventing the degradation of prediction performance due to task drift or equipment differences. Furthermore, the migration adaptation layer outputs a multi-classification probability distribution vector, with the category corresponding to the maximum probability as the prediction label for the time window, and simultaneously outputs the confidence probability value of the classification result.

[0120] For example, in a high-intensity milling task, the state of a tool was inferred and judged. The input was a three-channel time series of 300 sampling points in the past 10 seconds and the corresponding time period twin variables heat load factor, vibration index, and current anomaly factor (θ = 0.74, ζ = 0.92, γ = 0.88). After encoding by the self-encoding layer, a 128-dimensional feature vector was extracted, which was then concatenated with the twin variables to form a 131-dimensional fusion vector. After processing by the migration adaptation layer, the output classification result was severe wear, with a corresponding probability value of 0.89. This result indicates that the current machining state is very likely to have entered the critical wear stage, and process deload or tool change measures must be taken as soon as possible.

[0121] In one embodiment, triggering corresponding fault warning instructions and tool adjustment strategies based on the prediction results includes:

[0122] S51. Obtain warning response levels based on probability values; warning response levels include green, yellow, and red.

[0123] Schematically, the predicted probability value output by the deep transfer inference model determines the warning response level corresponding to the current processing state. The predicted probability value refers to the model's confidence level in classifying a fault label, and the value range is typically 0 to 1. To enhance the interpretability and operability of the response, the probability value is divided into three levels and mapped to three color-coded response levels. For example, when the probability value is less than 0.6, the green level is determined, indicating normal status or a slight fluctuation range, and no intervention is required. When the probability value is between 0.6 and 0.85, the yellow level is determined, indicating that the processing system has certain abnormal signs and lightweight parameter adjustments are recommended. When the probability value equals or exceeds 0.85, the red level is determined, indicating that the processing system is in a high-risk state, requiring immediate alarm triggering, forced adjustment of process parameters, or even suspension of operations. This level determination not only takes into account the absolute value of the confidence level but also supports a dynamic threshold fine-tuning mechanism, which can be configured differently based on user experience or different task scenarios.

[0124] S52. Generate warning instructions and tool adjustment strategies based on the warning response level and the prediction tag; the warning instructions are used to instruct the CNC machine tool to send reminder information to the user terminal according to the warning response level; the tool adjustment strategy includes feed rate, spindle speed, cutting depth and coolant pressure.

[0125] The warning response level and the corresponding prediction label generate specific warning instructions and tool adjustment strategies. This process is completed by the control strategy scheduling module, whose output includes two types of information: warning instructions for human-machine interaction and parameter adjustment strategies for equipment control. The warning instructions are sent to the user terminal, such as the operation screen or remote monitoring platform, through the CNC machine tool's communication interface. The prompt content includes the current fault type, probability value, risk level, and recommended response measures. For example, the statement "Prompts that the risk of severe tool wear is 87%. It is recommended to suspend processing and replace the tool" is pushed to the operation interface to assist the operator in making quick decisions.

[0126] Furthermore, the tool adjustment strategy is generated in conjunction with the prediction label and the warning level, and is used to adjust the processing parameters in real time to reduce the load and delay the evolution trend of the fault. The core variables of the adjustment strategy include feed rate, spindle speed, cutting depth and coolant pressure. The adjustment range corresponding to different fault types and response levels is different. For example, under the combination of slight wear and yellow warning, the feed rate is reduced by 10% and the cutting depth is reduced by 15%; for severe wear and red warning, a more aggressive strategy is implemented, including a 20% reduction in spindle speed and a 30% increase in coolant pressure; if there is a risk of chipping and the confidence level is extremely high, a shutdown command will be forced to trigger and a tool replacement prompt will be prompted. At the same time, the event will be recorded in the risk log library for use in twin model updates.

[0127] For example, when a vibration anomaly prediction tag for a tool is detected, the corresponding probability value is 0.81. Based on the classification strategy, it is determined to be a yellow warning, which then generates the following control strategy: reduce the feed rate to 90% of the original value, maintain the spindle speed unchanged, and increase the coolant pressure by 1.3 times. This strategy reduces the risk of vibration in the machining structure without affecting the machining cycle, effectively preventing resonance-triggered chipping failures.

[0128] In one embodiment, the method further comprises:

[0129] In response to the obtained prediction result confirmation instruction corresponding to the twin state variable, the historical risk factor is updated using the following formula:

[0130] R new =ω·R old +(1-ω)·R obs

[0131] Among them, R new is the new historical risk factor; R old Historical risk factor; R obs is the prediction result corresponding to the twin state variable; ω is the weight factor.

[0132] In schematic form, the predicted faults are updated to the historical risk factors according to the performance ratio, so that the digital twin model can adaptively enhance or weaken the sensitivity of risks such as excessive thermal load, vibration instability or current mutation in subsequent state variables and simulations, so as to enable the model to continue to evolve and improve accuracy.

[0133] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0134] Based on the same inventive concept, the embodiments of the present application also provide a digital twin-based machining fault warning system for implementing the digital twin-based machining fault warning method mentioned above. The implementation solution provided by this device is similar to the implementation solution described in the above method. Therefore, the specific limitations of one or more digital twin-based machining fault warning system embodiments provided below can be found in the above-mentioned limitations of the digital twin-based machining fault warning method, and will not be repeated here.

[0135] In an exemplary embodiment, Figure 4 As shown, a machining fault early warning system based on digital twin is provided, including:

[0136] Physical perception module, used to obtain multimodal data of the interaction between tool and workpiece during machining;

[0137] The data synchronization processing module is used to pre-process the multimodal data and obtain time series data based on timestamp synchronization;

[0138] The digital twin module is used to input time series data into the digital twin model to obtain twin state variables;

[0139] The fault prediction module is used to input time series data and twin state variables into the deep migration inference model to obtain prediction results;

[0140] The fault warning module is used to trigger corresponding fault warning instructions and tool adjustment strategies based on the prediction results.

[0141] In one embodiment, the digital twin module is further configured to project real-time process parameters onto corresponding process nodes in the digital twin model to obtain a multi-physics simulation boundary; the multi-physics simulation boundary includes a cutting heat source, a vibration excitation source, and an electric charge load;

[0142] The digital twin module is also used to perform finite source simulation on temperature data, vibration data, and current data based on historical risk factors according to the multi-physics field simulation boundary to obtain twin state variables; finite element simulation includes thermal field simulation, vibration field simulation, and electric power field simulation.

[0143] In one embodiment, the fault prediction module is further configured to perform multi-scale time series encoding on the time series data through a stacked coefficient self-encoding layer, and synchronously concatenate and fuse the data with the twin state variables to obtain a high-dimensional feature representation vector.

[0144] The fault prediction module is also used to map the high-dimensional feature representation vector to the discriminant space through the migration adaptation layer to obtain the prediction label and the corresponding probability value; the prediction labels include normal, slight wear, severe wear and chipping.

[0145] In one embodiment, the fault prediction module is further configured to obtain an early warning response level according to the probability value; the early warning response levels include green, yellow, and red;

[0146] The fault prediction module is also used to generate warning instructions and tool adjustment strategies based on the warning response level and prediction label; the warning instructions are used to instruct the CNC machine tool to send reminder information to the user terminal according to the warning response level; the tool adjustment strategy includes feed rate, spindle speed, cutting depth and coolant pressure.

[0147] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0148] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0149] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0150] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A machining fault early warning method based on digital twin, characterized in that: The method comprises: Acquire multimodal data of the interaction between the tool and the workpiece during machining; the multimodal data shown includes real-time process parameters, temperature data, vibration data, and current data; Preprocessing the multimodal data and obtaining time series data based on timestamp synchronization; Inputting the time series data into the digital twin model to obtain twin state variables; Inputting the time series data and the twin state variables into a deep migration inference model to obtain a prediction result; the prediction result includes a predicted fault label and a probability value; The corresponding fault warning instructions and tool adjustment strategies are triggered according to the prediction results.

2. The method according to claim 1, characterized in that The digital twin model is constructed by the following method: Obtaining structural data, process path planning data and tool attribute parameters of mechanical machine tools; Constructing a basic geometric model by parsing the structural data; Importing the process path planning data into the basic geometric model as initial boundary conditions to obtain a process path grid structure model; The process path grid structure model includes a plurality of process nodes; The tool attribute parameters are bound to the process path grid structure model to obtain the digital twin model.

3. The method according to claim 2, characterized in that The digital twin model obtains the twin state variables by the following method, including: Projecting the real-time process parameters onto the corresponding process nodes in the digital twin model to obtain a multi-physics field simulation boundary; the multi-physics field simulation boundary includes a cutting heat source, a vibration excitation source, and an electric charge load; Based on historical risk factors, the temperature data, the vibration data and the current data are subjected to finite source simulation according to the multi-physics field simulation boundary to obtain the twin state variables; the finite element simulation includes thermal field simulation, vibration field simulation and electric power field simulation.

4. The method according to claim 3, characterized in that The step of performing finite source simulation on the temperature data, the vibration data, and the current data according to the multi-physics field simulation boundary to obtain the twin state variables includes: Conduct thermal field simulation of temperature conduction through temperature data to obtain temperature field distribution; Extracting the maximum temperature rise rate of the temperature field distribution to obtain a heat load factor; Performing a structural dynamics simulation of vibration according to the vibration data to obtain a stress field distribution; the stress field distribution includes displacement and velocity; Extracting the resonance peak amplitude and energy dissipation ratio of the stress field distribution to obtain a vibration index; Performing power field simulation based on the current data to obtain power field distribution; the power field distribution includes power and energy consumption density; extracting the current mutation ratio of the power field distribution to obtain a current anomaly factor; fusing the heat load factor, the vibration index, and the current abnormality factor to obtain the twin state variable; The temperature field distribution is obtained by the following formula: Where ρ is the material density; c is the material specific heat capacity; is the temperature field distribution; k is the thermal conductivity; is the temperature gradient obtained from the temperature data; Q I It is the cutting heat source; The stress field distribution is obtained by the following formula: Among them, M is the mass matrix; üu is the acceleration; C is the damping matrix; is velocity; K is stiffness matrix; u is displacement; F cut (t) is the vibration excitation source; The power field distribution is obtained through the following formula: P(t)=U(t)·I(t)=F cut (t)·v(t)+P loss Among them, P(t) is power; U is voltage; I is current data; F cut is the vibration excitation source; v(t) is the instantaneous linear velocity of the tool; P loss is the no-load power loss; q joule is the energy consumption density; σ is the conductivity; E is the electric field intensity; R is the equivalent resistance of the main axis; and V is the current volume.

5. The method according to claim 1, wherein The deep migration inference model obtains the prediction result by the following method, including: Performing multi-scale time series encoding on the time series data through a stacked coefficient self-encoding layer, and synchronously splicing and fusing it with the twin state variables to obtain a high-dimensional feature representation vector; The high-dimensional feature representation vector is mapped to the discriminant space through a migration adaptation layer to obtain the predicted label and the corresponding probability value; the predicted label includes normal, slight wear, severe wear and chipping.

6. The method according to claim 5, characterized in that The triggering of corresponding fault warning instructions and tool adjustment strategies according to the prediction results includes: According to the probability value, an early warning response level is obtained; the early warning response level includes green, yellow and red; A warning instruction and a tool adjustment strategy are generated according to the warning response level and the prediction label; the warning instruction is used to instruct the CNC machine tool to send a reminder message to the user terminal according to the warning response level; the tool adjustment strategy includes feed rate, spindle speed, cutting depth and coolant pressure.

7. The method according to claim 3, characterized in that The method further comprises: In response to the obtained prediction result confirmation instruction corresponding to the twin state variable, the historical risk factor is updated using the following formula: R new =ω·R old +(1-ω)·R obs Among them, R new is the new historical risk factor; R old Historical risk factor; R obs is the prediction result corresponding to the twin state variable mentioned this time; ω is the weight factor.

8. A machining fault warning system based on digital twins, characterized in that: The system comprises: Physical perception module, used to obtain multimodal data of the interaction between tool and workpiece during machining; A data synchronization processing module, configured to pre-process the multimodal data and obtain time series data based on timestamp synchronization; A digital twin module, configured to input the time series data into a digital twin model to obtain twin state variables; A fault prediction module, configured to input the time series data and the twin state variables into a deep migration inference model to obtain a prediction result; The fault warning module is used to trigger corresponding fault warning instructions and tool adjustment strategies according to the prediction results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images