Machining process self-adaptive optimization and abnormal working condition recognition system and method

Abstract

The invention provides a processing process adaptive optimization and abnormal condition identification system and method, and the method employs a hierarchical decoupling architecture, a front-end display layer serves as an interaction interface, and the display content is obtained from a core algorithm layer through a data transfer and storage support layer; the core algorithm layer adopts a real-time and non-real-time service separation architecture, and the real-time and non-real-time service separation architecture is decoupled to ensure that real-time processing is not The hardware interaction layer realizes bidirectional data interaction with a machine tool numerical control system and a sensor through adaptive equipment and covers data acquisition and feedback control; the data transfer and storage support layer adopts a Redis publishing and subscribing mechanism, and a plurality of fixed channels are divided to realize data transmission decoupling; the system acquires processing full-dimension data through three-source data acquisition. Each layer is decoupled, data transmission is efficient, monitoring real-time performance, optimization accuracy and abnormity identification reliability in the machining process are effectively improved, and loss caused by abnormal working conditions is reduced.

Description

Technical Field

[0001] This invention relates to the fields of machining and industrial automation technology, specifically to a machining process adaptive optimization and abnormal condition identification system and method, and more particularly to a machining process adaptive optimization and abnormal condition identification system and method driven by multi-dimensional real-time monitoring, and a machining process data acquisition, multi-dimensional monitoring visualization, adaptive process parameter optimization and abnormal condition identification system, which is applicable to CNC machining scenarios of complex aerospace structural parts. Background Technology

[0002] In the CNC machining of complex aerospace structural components, the process faces numerous challenges due to their inherent complexity (often thin-walled, deep-cavity, and irregularly shaped curved surfaces, requiring micron-level dimensional accuracy), extremely high machining precision requirements (directly impacting the assembly accuracy and operational reliability of aerospace equipment), and the use of difficult-to-machine materials such as high-temperature alloys and composites (high hardness, high toughness, and high cutting resistance). Furthermore, the machining process is cumbersome, production batches are small, and customization is highly specialized. These challenges result in the following prominent technical difficulties: a. The monitoring of the machining process is limited to a single dimension. Traditional monitoring methods for CNC machining only focus on time-domain signals, such as spindle load, vibration, or spatial position information. However, the coupling and correlation between the two are poor, making it impossible to accurately locate the machining data corresponding to the workpiece cutting position. The data is difficult to analyze effectively and directly, resulting in an incomplete judgment of the machining status. b. Optimization of process parameters relies on manual experience or offline settings, which is difficult to adapt to changing cutting conditions. In difficult-to-machine areas, problems such as sudden increase in tool load and chipping may occur. Conservative parameter settings will significantly reduce cutting efficiency and fail to fully utilize the performance of machine tools and cutting tools. c. Abnormal working condition identification often uses a single signal static threshold method, which only judges single parameters such as vibration and load. The false alarm rate is high, and it lacks the ability to predict complex abnormalities such as tool wear, collision, and loose clamping, making it difficult to achieve early warning and traceability. d. The large volume and heterogeneous types of data being processed make real-time transmission and processing prone to lag. The low degree of coordination and coupling among multiple modules (monitoring, optimization, alarm) affects the overall response speed of the system.

[0003] In existing technologies, processing monitoring, parameter optimization, and anomaly identification are mostly independent systems, lacking an integrated solution, which cannot meet the processing requirements of high precision, high efficiency, and high reliability for complex aerospace structural components. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a system and method for adaptive optimization of the processing process and identification of abnormal working conditions.

[0005] The adaptive optimization and abnormal working condition identification system for processing provided by the present invention adopts a layered decoupling architecture, which includes: a front-end display layer, a core algorithm layer, a hardware interaction layer, and a data transfer and storage support layer. The front-end display layer is the display and interaction interface of the recognition system, and its display content is obtained from the core algorithm layer through the data transfer and storage support layer. The core algorithm layer adopts a separate architecture including real-time services and non-real-time services. The real-time services are used to complete real-time data processing, and the non-real-time services are used to execute time-consuming tasks. The real-time services and the non-real-time services are decoupled from each other. The hardware interaction layer communicates bidirectionally with the machine tool CNC system and sensors via a CNC system communication adapter and an external sensor acquisition card. The bidirectional data interaction includes data acquisition during the machining process and feedback control of the machining process. The data relay and storage support layer adopts the Redis publish-subscribe mechanism, and the data relay and storage support layer is divided into multiple fixed data transmission channels to achieve data transmission decoupling; The identification system uses a three-source data acquisition method to obtain full-dimensional data of the processing process.

[0006] Preferably, the real-time service has a fixed calculation period of 50ms; And / or, the plurality of fixed data transmission channels include: a high-frequency real-time channel, a low-frequency real-time channel, a command issuance channel, and a command feedback channel; the high-frequency real-time channel has a data transmission frequency greater than or equal to 20Hz and is used for continuous data update and transmission; the low-frequency real-time channel has a data transmission frequency less than 20Hz and is used for continuous data update and transmission; the command issuance channel is used to issue request commands to the front-end interactive interface; the command feedback channel is used to return the back-end processing results; And / or, the data acquired using the three-source data acquisition method includes: internal machine tool data acquired from the machine tool CNC system via the HNC adapter or NC-link protocol, external machining data acquired from external vibration sensors via the sensor acquisition card, and NC code execution data during the machining process.

[0007] Preferably, the core algorithm layer includes: a multi-dimensional monitoring and visualization module, an adaptive optimization control module, and an abnormal operating condition identification and prediction module; The multi-dimensional monitoring and visualization module constructs a spatiotemporally coupled digital twin monitoring environment, dynamically correlates the time-domain monitoring signals of the processing process with the three-dimensional spatial position, and realizes multi-dimensional visualization of the CNC machining process in the workshop through color mapping and state machine. The adaptive optimization control module is based on an optimization parameter database constructed from real-time load data and offline analysis. It uses a fuzzy control algorithm to dynamically calculate the feed rate optimization value and writes and executes the optimization parameters in real time through the CNC system interface. The abnormal working condition identification and prediction module establishes a learning and monitoring mechanism based on the spindle power signal. It generates an envelope curve model by performing sliding window statistics and boundary fusion on multiple historical qualified machining records of the same CNC machining task. During the machining process, it monitors in real time whether the vibration intensity exceeds the alarm boundary, wear boundary or load lower limit, thereby realizing the identification and early warning of abnormal working conditions including chipping, tool collision and clamping loosening.

[0008] Preferably, the multi-dimensional monitoring and visualization module includes: a spatiotemporal coupling monitoring unit, a spatial location monitoring unit, a state machine and UI interaction unit, and a UI interface partition design unit; The spatiotemporal coupling monitoring unit uses the ECharts dynamic line chart component to display the spindle load, triaxial vibration intensity, synthetic vibration intensity and time domain signal in real time at a preset refresh rate, and supports switching between multiple data sources and filtering of data items. Real-time communication between the front-end and back-end is achieved through a WebSocket push mechanism. Historical average curves and alarm threshold lines are overlaid on the chart, and the deviation between the real-time status and the historical baseline is calculated using the following formula: in, The signal value at the current moment. This represents the average signal value at the corresponding position in the same historical code segment. For standard deviation, when An exception flag is triggered when the value is greater than 3; The spatial position monitoring unit constructs a 3D rendering scene based on the Python VTK engine, and sets up a CuttingTool class and a Part class; the CuttingTool class includes tool type, tool diameter, tool cutting edge length, and tool overhang attributes; the Part class includes workpiece name, workpiece material, workpiece STL model vertex array, and workpiece normal vector attributes; The spatial position monitoring unit reads the tool and workpiece models in STL format, constructs an octree spatial index to accelerate collision detection, and realizes real-time visualization of the tool motion trajectory through a 4×4 homogeneous coordinate transformation matrix. The homogeneous transformation matrix is ​​as follows:

[0009] in, It is a 3×3 rotation matrix. The vector is a 3×1 translation vector, and T is a 4×4 homogeneous transformation matrix; simultaneously, the data size is processed by mapping the color depth to the corresponding position, and the color mapping function is:

[0010] in, For position Vibration intensity or load value at the location, , These are the minimum and maximum values ​​of the currently displayed data across the entire domain. For rainbow-colored lookup table functions; The state machine and UI interaction unit construct a finite state machine model, defining stable states and state transition events; the stable states include S0 standby, S1 start, S2 processing, and S3 end; the UI top bar status indicator color is updated in real time through the state transition events: blue for standby, yellow flashing for start, green for processing, and gray for end, and the corresponding business logic is triggered when the state changes, and the optimization parameter library is automatically loaded when entering the processing state; The UI interface partition design unit adopts a Flex flexible layout to divide the interface into multiple areas, including: top toolbar, left status bar, central main view area, bottom NC code bar, and bottom status bar. The top toolbar is used to switch between spatiotemporal monitoring views and abnormal monitoring views, as well as to enable adaptive function switches and system settings. The left status bar displays the machine tool status description, current program name, operator information, and current alarm list. The central main view area displays a 3D rendering canvas and real-time curve graphs. The bottom NC code bar highlights the currently executing code line in real time and supports clicking to navigate to the corresponding position in the 3D view. The bottom status bar displays the database connection status, the activation status of each functional module, system time, and network latency.

[0011] Preferably, the adaptive optimization control module includes: an optimization parameter database construction unit, an intelligent control algorithm unit, and a parameter execution interface unit; The optimization parameter database construction unit analyzes historical machining data offline and extracts the no-load load P1 and target load P2 for each combination of program name and tool number; it then uses a sliding window algorithm with a preset window width and preset step size to scan the no-load segment and extract a stable no-load segment that meets the conditions. The no-load load P1 is defined as satisfying P1≥P min +0.1×(P max P min The minimum power value for which the continuous data duration is not less than 5 seconds and the load value is less than P1 is selected, and transient fluctuation data points at the moment the cutting fluid is turned on or off are excluded; P max This refers to the maximum load power collected during the machining process corresponding to the same program name and tool number combination; P minThis refers to the minimum load power collected during the machining process corresponding to the same program name and tool number combination; The target load P2 is defined as the minimum power value in the stable processing section where the proportion of data points with load values ​​less than P2 is not less than 95%, used to characterize the typical load level of this process; at the same time, the vibration intensity under the corresponding load is recorded as a reference benchmark V. ref ; The intelligent control algorithm unit adopts a two-dimensional fuzzy PID control algorithm, with the error E(t) between the real-time load P(t) and the target load P2 being equal to P2. P(t) and the error derivative Ec(t) = dE / dt are the input quantities, quantized by factor K. e K ec Mapping to the fuzzy domain: E * =round(K e E), K e =6 / P2×0.5 Ec * =round(K ec Ec), K ec =6 / P2×0.1 Among them, E * and Ec * The input fuzzy subsets are all {NB, NM, NS, ZO, PS, PM, PL}; the universe of discourse for error E(t) is {-6, -4, -2, 0, 2, 4, 6}; the universe of discourse for error differential Ec is {-6, -4, -2, 0, 2, 4, 6}; the output is the feed rate increment ΔU, with the same fuzzy subset as the input, and the universe of discourse is {-3, -2, -1, 0, 1, 2, 3}. The membership function adopts a triangular distribution. A fuzzy rule table is constructed, fuzzy inference is performed using the Mamdani synthesis inference method, and fuzziness is resolved using a weighted average method.

[0012] Where, μ i Let C be the activation degree of the i-th rule. i The center value of the fuzzy set is output for the corresponding rule; the final output control quantity ΔU=Ku ΔU*, Ku is the scaling factor; the actual feed rate optimization value F opt (t+1)=F opt (t)+ΔU, and limited to the range of 30% to 150%, F opt (t) refers to the actual optimized feed rate at the current time t; a vibration feedback correction term is also introduced:

[0013] Where V(t) is the real-time vibration intensity, The target vibration reference is η, which is the vibration suppression coefficient; when the vibration intensity exceeds 1.5 times... At that time, active speed reduction control is executed; The parameter execution interface unit is designed for CNC systems. It writes R parameters through the OPC UA method node and calls synchronous action instructions to activate the magnification coverage. It adopts a dual-channel redundant writing mechanism, using dual writing through PLC data block channel 1 magnification DB21 and axis magnification DB31, to ensure that the magnification activation success rate is not less than 99.9%. Final feed rate F final =F opt ×F panel / 100%, of which F panel Set the control panel knob value to achieve a fusion control of optimized magnification and manual settings.

[0014] Preferably, the abnormal operating condition identification and prediction module includes: a learning model construction unit, a collision detection unit, and an abnormal alarm and backtracking unit; The learning model construction unit is built based on the recording, learning and monitoring mechanism of the spindle power signal; the learning model construction unit is used to analyze the temporal characteristic changes of the spindle power signal in order to realize real-time judgment of abnormal working conditions including chipping, tool collision, and loose clamping. The learning model building unit supports training corresponding independent learning models for different CNC programs. Each model is associated with the program name and tool number and is used for anomaly monitoring and identification in the repeated machining process of the same program. The process of generating the learning model includes the following steps: Sample selection steps: Select one or more qualified machining records with the same program name and the same tool number as the training sample set; Parameter initialization steps: Set the time window length W to 0.5s and the alarm boundary bias value Δ alarm 0.2g, wear boundary deviation value Δ wear 0.1g, lower limit of load bias Δ low It is 0.05 times the load fluctuation range; Time axis alignment step: The dynamic time warping algorithm is used to uniformize the timestamps of each processing record in the training sample set to align the time axes of multiple processing records. Early warning boundary calculation steps: For each processing record, traverse the entire processing process with a time window length W, and calculate the boundary value at each time window position t. i At this point, the statistical characteristics of the monitoring signal within the calculation window are used to generate the upper and lower warning boundary sequences for this record; The upper warning boundary point is defined as: The lower warning boundary point is defined as: in, This is the spindle power signal sequence within this window. This represents the maximum power within the window. This represents the minimum power value within the window. The average power within the window. These are boundary coefficients; Envelope curve fusion steps: For a sample set containing K learning records, at position t in the i-th time window... i At this point, the upper and lower warning boundaries generated from each record are merged to obtain the final envelope curve model: Upper envelope curve: , Lower envelope curve: , in, , These represent the upper and lower warning boundaries of the k-th record in the i-th window. , The paranoia value set for the user. The value of Δ corresponds to alarm , The value of Δ corresponds to wear or Δ low ; Dynamic boundary generation steps: based on the fused envelope curve , Generate three boundary curves for real-time monitoring: Alarm boundary: B alarm (t)=U env (t)+Δ alarm ×σ U (t) Wear boundary: B wear (t)=U env (t) Δ wear ×σ U (t) Lower load limit: B low (t)=L env (t) Δ low ×σ L (t) Where, σ U (t), σ L (t) represents the standard deviation of each learning record at time t corresponding to the upper and lower boundaries, which is used to characterize the data fluctuation of the sample set; During the machining process, the collision detection unit is used to monitor the vibration intensity signal V(t) in real time and compare it with the three boundary curves to determine the tool condition. When V(t) > B alarm When (t) occurs, it is determined to be a serious abnormality, including chipping or collision, triggering an emergency stop command; When B wear (t) <V(t)≤B alarm If the error is detected as tool wear or a minor abnormality, a level two alarm will be triggered. When V(t) low When (t) occurs, it is determined that the clamping is loose or the air cutting is abnormal, triggering a level one warning. The learning model building unit supports model version management and iterative optimization. The specific steps are as follows: Model evaluation steps: Select the optimal model by comparing the prediction errors of multiple learning models on the same validation set; the prediction error metrics include root mean square error (RMSE) and mean absolute percentage error (MAPE). The formula for calculating the root mean square error (RMSE) is as follows:

[0015] The formula for calculating the Mean Absolute Percentage Error (MAPE) is as follows:

[0016] Incremental update steps: Periodically use newly accumulated qualified processing data to incrementally learn and update the selected optimal model to adapt to changes in production conditions.

[0017] Preferably, the collision detection unit is used to monitor the vibration intensity of the turning / milling channel in real time, and sets a collision judgment threshold through a preset formula. When the real-time monitored vibration intensity exceeds the collision judgment threshold, a machine tool stop signal is triggered. The formula for calculating the collision determination threshold is as follows: T collision =V ref ×(1+α)+β σ ref Among them, V ref For reference vibration intensity, α is the base percentage coefficient, β is the dynamic margin coefficient, and σ ref For reference vibration standard deviation; The reference vibration intensity is obtained by the learning model construction unit through learning qualified processing records; the reference vibration standard deviation corresponds to the reference vibration intensity and is obtained synchronously by the learning model construction unit through learning qualified processing records.

[0018] ​Preferably, the abnormal alarm and backtracking unit triggers graded alarms and shutdown commands based on the comparison result between the vibration intensity and the preset boundary: When the vibration intensity exceeds the wear boundary or the lower limit of the load, the abnormal alarm and backtracking unit triggers a graded alarm. When the vibration intensity exceeds the alarm threshold, the abnormal alarm and backtracking unit triggers a machine tool stop command. The abnormal alarm and backtracking unit supports the abnormal event backtracking function, and can retrieve and analyze the parameters, waveforms and status records related to the alarm during the current and historical processing.

[0019] This invention also provides a method for adaptive optimization of the processing process and identification of abnormal working conditions, based on the above-mentioned adaptive optimization of the processing process and identification of abnormal working conditions, including the following steps: System initialization steps: Start the system's layered decoupled architecture, load the CNC machine tool configuration parameters, sensor acquisition parameters and the system's stored learning model library, complete the initial deployment of each functional module, and ensure normal linkage between the front-end display layer, core algorithm layer, hardware interaction layer and data transfer and storage support layer; Learning model construction steps: For a preset CNC machining program, select the qualified machining records corresponding to the history of the CNC machining program as training samples. Through the learning model construction unit of the abnormal working condition identification and prediction module of the core algorithm layer, the envelope curve model is generated by the sliding window statistics and boundary fusion algorithm. The envelope curve model is associated with and stored with the CNC machining program and the corresponding tool number. Real-time machining monitoring steps: The machining process data is acquired in all dimensions through the system’s three-source data acquisition method, forming three-source data. The three-source data includes internal data acquired from the machine tool CNC system through the HNC adapter or NC-link protocol, external data acquired from the external vibration sensor through the sensor acquisition card, and NC code execution data. The acquired data is distributed to the corresponding functional modules through multiple fixed channels of the data transfer and storage support layer. Multi-dimensional monitoring and visualization steps: Through the multi-dimensional monitoring and visualization module of the core algorithm layer, based on the real-time data distributed in the real-time processing monitoring steps, the three-dimensional digital twin monitoring scene is updated and the real-time dynamic curve is drawn, so as to realize the multi-dimensional visualization of the processing process. Adaptive optimization control steps: The adaptive optimization control module of the core algorithm layer, based on the fuzzy control algorithm and combined with real-time load data and optimization parameter database, calculates the feed rate optimization value in real time, and writes it into the CNC system for execution through the parameter execution interface unit; Abnormal working condition identification steps: Through the collision detection unit and learning model construction unit of the abnormal working condition identification and prediction module of the core algorithm layer, the spindle power signal and vibration signal are collected. The real-time vibration signal is compared with the alarm boundary, wear boundary and load lower limit corresponding to the envelope curve model generated in the learning model construction step to determine the abnormal processing state. The abnormal alarm and backtracking unit of the abnormal working condition identification and prediction module of the core algorithm layer triggers the corresponding graded alarm, shutdown or early warning processing. Data storage and model iteration steps: Processing data collected in the real-time processing monitoring step, optimization parameter data in the adaptive optimization control step, and abnormal condition identification data in the abnormal condition identification step are stored in the database of the data transfer and storage support layer. Newly accumulated qualified processing data are used periodically to incrementally train the learning model through the learning model building unit to achieve model iterative optimization.

[0020] Preferably, in the real-time processing monitoring step, the three-source data includes: First source data: Acquire internal data such as spindle load, axis coordinates, and feed rate from the machine tool CNC system via HNC adapter or NC-link protocol; Second source data: Three-dimensional vibration signals of the turning / milling channel are acquired from an external vibration sensor via a sensor acquisition card; Third source data: Obtain the current execution line number and code content through NC code parsing; The data relay and storage support layer uses a Redis publish-subscribe mechanism to divide the data into four fixed data transmission channels, which are as follows: The high-frequency real-time channel transmits monitoring data during the processing and data required for updating the digital twin scene at a data transmission frequency of greater than or equal to 20Hz. The low-frequency real-time channel transmits model-generated and interface-requested data at a data transmission frequency of less than 20Hz. The command issuing channel is used to receive request instructions issued by the front-end presentation layer and forward the request instructions to the corresponding back-end functional modules; The command feedback channel is used to receive the processing results of various backend functional modules and forward the processing results to the frontend display layer to realize data interaction between the frontend and backend.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention proposes a layered and decoupled system architecture and a four-level data channel. The core algorithm layer separates real-time services from non-real-time services to ensure that time-consuming tasks do not affect the stable operation of high-frequency monitoring.

[0022] 2. This invention constructs an adaptive envelope curve anomaly recognition model based on spindle power signal learning. It records and learns the dynamic machining process based on CNC program, and generates three envelope curves: alarm boundary, wear boundary, and load lower limit. It can recognize abnormal working conditions such as chipping, tool collision, and loose clamping.

[0023] 3. This invention constructs an adaptive optimization method that combines fuzzy PID control with vibration feedback correction. Based on the error and error derivative between the real-time load and the target load, the optimal value of the feed rate is dynamically calculated, thereby improving processing efficiency while ensuring processing quality in the early stage. Attached Figure Description

[0024] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a system block diagram of the adaptive optimization and abnormal operating condition identification system according to an embodiment of the present invention; Figure 2 The principle of the adaptive optimization and abnormal operating condition identification method in this embodiment of the invention Figure 1 ; Figure 3 The principle of the adaptive optimization and abnormal operating condition identification method in this embodiment of the invention Figure 2 . Detailed Implementation

[0025] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0026] Example 1 like Figures 1 to 3 As shown, this embodiment provides a multi-dimensional real-time monitoring-driven adaptive optimization and abnormal working condition identification system and method for the processing process. It adopts a layered decoupled architecture, including a front-end display layer, a core algorithm layer, a hardware interaction layer, and a data transfer and storage support layer, wherein: The front-end display layer is a pure display and interactive interface, without data calculation and statistical functions. All display content is obtained from the core algorithm layer through the data transfer layer. The core algorithm layer adopts a real-time service and non-real-time service separation architecture. It completes high-frequency data processing with a fixed operation cycle of 50ms, while the non-real-time service is responsible for time-consuming tasks such as algorithm learning and data statistics. The two are decoupled to ensure that the real-time service is not blocked. The hardware interaction layer achieves bidirectional data interaction with the machine tool CNC system and sensors through the CNC system communication adapter and external sensor acquisition card, including two major functions: data acquisition and feedback control. The data relay and storage support layer uses the Redis publish-subscribe mechanism to divide the data into four fixed channels to achieve data transmission decoupling, including: a high-frequency real-time channel for continuous data updates at 20Hz, a low-frequency real-time channel for low-frequency data updates with strong randomness, a command issuance channel for front-end request issuance, and a command feedback channel for back-end processing result return. The system acquires machining process data through a three-source data acquisition method, including: internal data acquired from the machine tool CNC system via an HNC adapter or NC-link protocol, external data acquired from an external vibration sensor via a sensor acquisition card, and NC code execution data, thereby achieving multi-dimensional perception of the machining process.

[0027] Furthermore, the core algorithm layer further includes a multi-dimensional monitoring and visualization module, an adaptive optimization control module, and an abnormal operating condition identification and prediction module, wherein: The multi-dimensional monitoring and visualization module is used to construct a spatiotemporally coupled digital twin monitoring environment, dynamically associate time domain signals with three-dimensional spatial positions, and realize multi-dimensional visualization of the CNC machining process in the workshop through color mapping and state machine. The adaptive optimization control module is used to dynamically calculate the feed rate optimization value based on the optimization parameter database constructed by real-time load data and offline analysis, and realizes the real-time writing and execution of parameters through the CNC system interface. The abnormal working condition identification and prediction module is used for the learning and monitoring mechanism based on the spindle power signal. By performing sliding window statistics and boundary fusion on multiple qualified machining records of the same CNC machining, an envelope curve model is generated. In actual machining, the vibration intensity is monitored in real time to see if it exceeds the alarm boundary, wear boundary or load lower limit, so as to identify and warn of abnormal working conditions such as chipping, tool collision, and loose clamping.

[0028] Furthermore, the multi-dimensional monitoring and visualization module further includes a spatiotemporal coupling monitoring unit, a spatial location monitoring unit, a state machine and UI interaction unit, and a UI interface partition design unit, wherein: The spatiotemporal coupling monitoring unit uses an ECharts dynamic line chart component to display spindle load, triaxial vibration intensity, synthetic vibration intensity, and time-domain signal in real time at a 20Hz refresh rate. It supports switching between multiple data sources and filtering data items, and achieves real-time communication between the front-end and back-end through a WebSocket push mechanism. Historical average curves and alarm threshold lines are overlaid on the chart to intuitively compare the deviation between the real-time status and the historical benchmark. The deviation calculation formula is as follows: in, The signal value at the current moment. This represents the average signal value at the corresponding position in the same historical code segment. For standard deviation, when An exception flag is triggered when the value is >3; The spatial position monitoring unit constructs a 3D rendering scene based on the Python VTK engine. It establishes a CuttingTool class containing attributes such as tool type, diameter, cutting edge length, and overhang, and a Part class containing workpiece name, material, STL triangular facet vertex array, and normal vector. It reads the STL format tool / workpiece model and constructs an octree spatial index to accelerate collision detection. Real-time visualization of the tool motion trajectory is achieved through a homogeneous coordinate transformation matrix.

[0029] Where R is a 3×3 rotation matrix, t is a 3×1 translation vector, and T is a 4×4 homogeneous transformation matrix; the color mapping function is defined as follows: (The color mapping function maps the processing data size to the corresponding position based on color depth.)

[0030] Where A(p) is the vibration intensity or load value at position p. , The minimum / maximum value of the currently displayed data is given by f, which is a rainbow-colored lookup table function. The state machine and UI interaction unit construct a finite state machine model, defining four stable states: S0 standby, S1 start, S2 processing, and S3 end, as well as six state transition events. The state indicator lights on the top bar of the UI are updated in real time through the state transition diagram: blue for standby, flashing yellow for start, green for processing, and gray for end. The corresponding business logic is triggered when the state changes, such as automatically loading the optimization parameter library when entering the processing state. The UI interface partition design unit adopts a Flex layout and is divided into five zones: Top toolbar: Spatiotemporal monitoring view switch, anomaly monitoring view switch, adaptive function switch, system settings; Left status bar: Machine tool status description, current program name, operator, current alarm list; Central main view area: 3D rendering canvas, real-time curve graph; Bottom NC code bar: Real-time highlighting of the currently executing line of code, supports clicking to locate the corresponding position in the 3D view; Bottom status bar: Database connection status, activation status of each functional module, system time, network latency.

[0031] Furthermore, the adaptive optimization control module further includes an optimization parameter database construction unit, an intelligent control algorithm unit, and a parameter execution interface unit, wherein: The optimization parameter database construction unit analyzes historical machining data offline. For each program name and tool number combination, it extracts the no-load load P1 and the target load P2. Using a sliding window algorithm with a window width of 5s and a step size of 1s, it scans the no-load stroke segment and extracts the stable no-load segment that meets the conditions. The no-load load P1 is defined as satisfying P1≥P min +0.1×(P max P min The minimum power value with a load value less than P1 and a continuous data duration of ≥5s is selected, while transient fluctuation data points at the moment the cutting fluid is turned on / off are discarded. The target load P2 is defined as the minimum power value in the stable processing section where the percentage of data points with load values ​​less than P2 is ≥95%, reflecting the typical load level of this process. Simultaneously, the vibration intensity at the corresponding load value is recorded as a reference benchmark V. ref ; The intelligent control algorithm unit adopts a two-dimensional fuzzy PID control algorithm, and the input is the error E(t) between the real-time load P(t) and the target load P2 = P2. P(t), the differential error Ec(t) = dE / dt, and the quantization factor K e K ec Mapping to the fuzzy domain: E =round(K e E), Ke = 6 / P² × 0.5 Ec =round(K ec Ec), Kec = 6 / P² × 0.1 The fuzzy subsets are all {NB, NM, NS, ZO, PS, PM, PL}, with universes of discourse E={-6,-4,-2,0,2,4,6} and Ec={-6,-4,-2,0,2,4,6}, respectively. The output is the feed rate increment ΔU, with the same fuzzy subset as the input, and a universe of discourse {-3,-2,-1,0,1,2,3}. The membership function adopts a triangular distribution. The fuzzy rule table is constructed based on expert experience (49 rules in total). Fuzzy inference is performed using the Mamdani synthesis inference method, and the weighted average method is used to resolve fuzziness.

[0032] in, Let i be the activation value of the i-th rule. The center value of the fuzzy set is output for the corresponding rule, and the final output is the control quantity. , F is the scaling factor, representing the optimized value of the actual feed rate. opt (t+1)=F opt (t)+ΔU, and limited to the interval [30%, 150%], while f also introduces a vibration feedback correction term:

[0033] in For real-time vibration intensity, As the target vibration reference, This is the vibration suppression coefficient, which is applied when the vibration intensity exceeds 1.5 times the normal value. Actively reduce speed; The parameter execution interface unit, designed for the Siemens 840Dsl CNC system, writes the R parameter via the OPC UA method node and calls synchronous action instructions to activate the rate overlay. Simultaneously, a dual-channel redundant writing mechanism is employed, using both PLC data block channel 1 rate overlay DB21 and axis-specific rate overlay DB31 for dual writing, ensuring a rate overlay success rate of ≥99.9%, ultimately achieving the feed rate F. final =Fopt×F panel / 100%, where F panel Set the control panel knob value to achieve a fusion control between optimized values ​​and manual settings.

[0034] Furthermore, the abnormal operating condition identification and prediction module further includes a learning model construction unit, a collision detection unit, and an abnormal alarm and backtracking unit, wherein: The learning model building unit is based on the recording, learning, and monitoring mechanism of spindle power signals and is designed for repetitive machining scenarios of batch production models. By analyzing the temporal characteristic changes of spindle power signals, it can realize real-time judgment of abnormal working conditions such as chipping, tool collision, and loose clamping. This unit supports training corresponding learning models for different CNC programs. Each model is stored independently and associated with the program name and tool number for subsequent abnormal monitoring and identification in the repetitive machining process of the same program. The process of generating the learning model includes: selecting one or more qualified machining records with the same program name and the same tool number as training samples, with a default time window length W of 0.5s and an alarm boundary bias value Δ. alarm 0.2g, wear boundary deviation value Δ wear 0.1g, lower limit of load bias Δ low To achieve a load fluctuation range of 0.05, a dynamic time warping algorithm is used to even out the timestamps of each processing record, ensuring time axis alignment between multiple records. For each learning record, a sliding time window is used to traverse the entire processing, and the statistical characteristics of the monitoring signal are calculated within each window to form the upper and lower warning boundary sequences for that record: Upper warning boundary point: Upper warning boundary point: in, The sequence of spindle power signals within the window. The maximum value within the window. The minimum value within the window. This is the average value within the window. These are boundary coefficients; For a sample set containing K learning records, at each time window position titi, the upper and lower warning boundaries generated by fusing the records are merged to obtain the final envelope curve model: in, , These represent the upper and lower warning boundaries of the k-th record in the i-th window. , The paranoia value set for the user can correspond to Δalarm, Δwear, or Δlow, respectively. Based on the fused envelope curve, three boundary curves are further generated for real-time monitoring: Alarm boundary: B alarm (t)=U env (t)+Δ alarm ×σ U (t) Wear boundary: B wear (t)=U env (t) Δ wear ×σ U (t) Lower load limit: B low (t)=L env (t) Δ low ×σ L (t) Where, σ U (t), σ L (t) represents the standard deviation of the upper and lower boundaries of each learning record at the corresponding time, reflecting the fluctuation between samples, so that the boundary curve can adaptively adjust its width; During actual machining, the vibration intensity V(t) of the corresponding channel is monitored in real time, and the tool condition is judged by comparing it with three envelope curves. When V(t) > B alarm When (t) is reached, it is judged as a serious abnormality (such as blade breakage or collision), triggering an emergency stop; When B wear (t) <V(t)≤B alarm When (t) occurs, it is determined to be tool wear or a minor abnormality, triggering a level two alarm; When V(t) low When (t) occurs, it is determined that the clamping is loose or there is a gap, triggering a level one warning. The learning model supports version management and iterative optimization. By comparing the prediction errors of models generated from multiple learning records on the same validation set, including root mean square error (RMSE) and mean absolute percentage error (MAPE), the optimal model is automatically selected as the current production version. The model can also be incrementally updated periodically using newly accumulated qualified processing data. The calculation formula is as follows:

[0035]

[0036] Furthermore, the collision detection unit monitors the vibration intensity of the turning / milling channel in real time, sets a threshold according to a formula, and triggers a machine tool stop signal when the threshold is exceeded. The formula is as follows: T collision =Vref×(1+α)+β σ ref Among them, V ref For reference vibration intensity (obtained through learning), α is the baseline percentage coefficient, β is the dynamic margin coefficient, and σ ref The standard deviation of vibration is used as a reference.

[0037] Furthermore, the abnormal alarm and backtracking unit triggers a graded alarm when the vibration intensity exceeds the wear boundary / load lower limit, and stops when it exceeds the alarm boundary; it supports abnormal backtracking and historical backtracking functions.

[0038] This embodiment also provides an identification method for the above-mentioned multi-dimensional real-time monitoring driven adaptive optimization and abnormal working condition identification system and method for processing, including the following steps: Step 1: System initialization, loading CNC machine tool configuration, sensor parameters and existing learning model library; Step 2: Learning model construction. For a specific CNC program, select the qualified machining records corresponding to the history of the CNC machining program, and generate an envelope curve model through sliding window statistics and boundary fusion algorithm. Step 3: Real-time processing monitoring. Processing data is acquired through a three-source data acquisition method and distributed to each functional module via four fixed channels in the data transfer layer. ​Step 4: Multi-dimensional monitoring and visualization, based on real-time data to drive the updating of 3D digital twin scenes and dynamic curve drawing; Step 5: Adaptive optimization control, based on fuzzy control algorithm, calculates the feed rate optimization value in real time and writes it into the execution; Step 6: Abnormal operating condition identification. Collect spindle power signal and vibration signal, compare them in real time with alarm boundary, wear boundary and load lower limit generated by the learning model, judge the abnormal state and trigger corresponding processing. Step 7: Data storage and model iteration. Store the processed data in the database and periodically use new data to incrementally train the learning model.

[0039] Furthermore, the three-source data acquisition method in step 3 includes: acquiring internal data such as spindle load, axis coordinates, and feed rate from the machine tool CNC system through an HNC adapter or NC-link protocol; acquiring three-dimensional vibration signals of the turning / milling channel from an external vibration sensor through a sensor acquisition card; and obtaining the current execution line number and code content through NC code parsing.

[0040] Furthermore, the data transfer layer uses four fixed channels defined by the Redis publish-subscribe mechanism: a high-frequency real-time channel that transmits monitoring and digital twin data at a frequency of 20Hz; a low-frequency real-time channel that transmits highly random data such as model generation and interface requests; a command publishing channel that receives front-end requests and sends them to the back-end; and a command feedback channel that receives back-end processing results and relays them to the front-end.

[0041] This embodiment provides a multi-dimensional real-time monitoring-driven adaptive optimization and abnormal working condition identification system and method for machining processes, belonging to the field of mechanical processing and industrial automation technology. The system adopts a layered decoupled architecture, including a front-end display layer, a core algorithm layer, a hardware interaction layer, and a data transfer and storage support layer. The core algorithm adopts an architecture that separates real-time and non-real-time services, completing high-frequency data processing with a period of 50ms. The hardware interaction layer realizes the acquisition of three-source data through HNC adapters, NC-Link protocol, and sensor acquisition cards. The data transfer layer divides four fixed channels based on the Redis publish-subscribe mechanism to achieve data transmission decoupling. The system also includes multi-dimensional monitoring visualization, adaptive optimization control, and abnormal working condition identification modules.

[0042] This embodiment solves the problems of existing technologies, such as single monitoring, weak optimization capabilities, low accuracy in identifying abnormal operating conditions, and poor system coordination. It is particularly suitable for high-precision, high-efficiency, and high-reliability processing of complex aerospace structural components.

[0043] This embodiment aims to solve the problems of incomplete monitoring dimensions, weak adaptive parameter optimization, low accuracy of anomaly identification, and poor system coordination in existing processing control technologies, and provides an integrated system that integrates multi-dimensional monitoring, adaptive optimization, and anomaly prediction.

[0044] Example 2 Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1.

[0045] This embodiment provides a system and method for adaptive optimization of the processing process and identification of abnormal working conditions.

[0046] 1. System hardware deployment and connection relationship.

[0047] The hardware deployment underlying the multi-dimensional real-time monitoring-driven adaptive optimization and abnormal working condition identification system and method for the processing in this embodiment includes: CNC machine tools: Five-axis machining centers equipped with Siemens 840D sl CNC system or Huazhong CNC HNC-848 system are used for machining complex structural parts such as cabin, rudder, and wall panels. The machine tools support PLC data interaction and real-time writing of R parameters. Sensor System: A three-dimensional ICP accelerometer (PCB 356A25 or equivalent) is attached to the machine tool spindle near the tool using a magnetic base or adhesive. Three vibration sensors (corresponding to the X, Y, and Z directions respectively) are configured in each turning / milling channel. Acoustic emission sensors (PAC WD or equivalent) are also configured for high-frequency signal monitoring. The sensor signals are connected to an NI 9234 data acquisition card (4 channels, 51.2kS / s sampling rate, 24-bit resolution) and then connected to a data processing server via an Ethernet cable. Communication adapter: For the Siemens 840D sl system, configure an industrial Ethernet switch to communicate with the server via the OPC UA protocol; for the Huazhong CNC system, configure a dedicated HNC Ethernet adapter to achieve bidirectional data transmission via the NC-Link protocol. Server: High-performance workstations (configured with Intel Xeon processors, 64GB of memory, and NVIDIA RTX graphics cards) are deployed with a Redis real-time database (for high-speed data caching and distribution), a MySQL relational database (for historical data storage), and core algorithm services (including real-time and non-real-time service processes). Front-end terminal: It adopts an industrial panel PC or ordinary PC, connects to the server via the network, runs a web-based monitoring interface, and supports multiple users to access at the same time.

[0048] The hardware connections are as follows: the vibration sensor and acoustic emission sensor are connected to the NI data acquisition card via a low-noise cable, and the data acquisition card is connected to the server via Gigabit Ethernet; the server is connected to the machine tool CNC system (OPC UA / NC-Link protocol) via an industrial Ethernet switch to realize real-time reading of PLC data and writing of R parameters; the front-end terminal is connected to the server via the enterprise intranet wirelessly or via wired connection, and uses the WebSocket protocol to realize real-time data push and command interaction.

[0049] 2. System software architecture and data flow.

[0050] The system software architecture of this embodiment adopts a layered and decoupled design, including a front-end presentation layer, a core algorithm layer, a hardware interaction layer, and a data transfer and storage support layer. The specific implementation is as follows: 2.1 Front-end presentation layer.

[0051] The front-end presentation layer is developed based on the Vue.js framework, uses the ECharts component library for curve drawing, and Three.js / Mayavi for 3D rendering. This layer does not contain any data calculation or statistical functions; all displayed content is retrieved from the backend via WebSocket. The UI is divided into five areas: top toolbar (containing time-space monitoring view switching, anomaly monitoring view switching, adaptive function switch, and system settings), left status bar (displaying machine tool status, current program name, operator, and current alarm list), central main view area (3D rendering canvas occupies 60%, real-time curve chart occupies 40%), bottom NC code bar (real-time highlighting of the currently executing code line, supporting click to locate the corresponding position in the 3D view), and bottom status bar (displaying database connection status, activation status of each functional module, system time, and network latency).

[0052] 2.2 Core Algorithm Layer.

[0053] The core algorithm layer adopts a separate architecture for real-time and non-real-time services, implemented using a mix of C++ and Python programming. Real-time service process: Running with a fixed 50ms computation cycle and set to the highest priority, this process is responsible for high-frequency data processing tasks, including: vibration intensity calculation, sensor data parsing and feature extraction such as sliding window filtering, fuzzy PID calculation of adaptive optimization control algorithms, and real-time anomaly identification based on envelope curve comparison. This process obtains raw data by subscribing to a high-frequency real-time channel via Redis, processes it, and publishes the results through the same channel, ensuring an end-to-end latency of less than 50ms.

[0054] Non-real-time service process: Responsible for asynchronous processing of time-consuming tasks, including: training CNN-BiLSTM learning models, building optimization parameter databases, and historical data statistics and report generation. This process is decoupled from the real-time service process and communicates through Redis low-frequency real-time channels and command channels. Even if the non-real-time service is blocked, it will not affect the normal operation of the real-time service.

[0055] 2.3 Hardware Interaction Layer.

[0056] The hardware interaction layer consists of three independent programs, each corresponding to one of the three data acquisition methods: Sensor acquisition program: Developed based on NI-DAQmx driver, it synchronously acquires vibration sensor signals from three sensors each in the car / milling channels at a sampling rate of 51.2kHz. After anti-aliasing filtering, the signals are downsampled to 2560Hz. The unidirectional vibration intensity and triaxial composite vibration intensity are calculated in real time using the following formula:

[0057] in, The standard uncertainty (or root mean square error, standard deviation) in the x-direction represents the degree of dispersion of the measurement result in the x-direction; N represents the number of measurements (sample size), that is, the same quantity was measured N times repeatedly. This represents the error component (or deviation value) in the x-direction during the j-th measurement, which is the difference between the j-th measurement result and the true value / average value. It represents the combined standard uncertainty (or total uncertainty, combined error), which represents the total dispersion of the measurement results in three-dimensional space; , , These are the standard uncertainties in the three orthogonal directions: x, y, and z.

[0058] The calculation results are sent to the server via UDP multicast, with port numbers 5001 and 5002 respectively.

[0059] HNC adapter program: Developed based on Huazhong CNC SDK, it establishes a connection with the machine tool PLC via Ethernet. It uses a polling method to read data such as spindle load, axis coordinates, feed rate, and spindle speed at 100ms cycles, and also supports real-time writing of R parameters. The communication protocol follows the NC-Link standard, enabling unified format output of multi-source heterogeneous data.

[0060] NC-Link protocol stack program: used for Siemens 840D SL system, developed based on OPC UA client, automatically identifies server address space through discovery mechanism, subscribes to required variable nodes, and the data update frequency is set to 20Hz.

[0061] 2.4 Data transfer and storage support layer.

[0062] The data transfer layer is deployed using a Redis 6.0 cluster, and is divided into four fixed channels based on a publish-subscribe mechanism: High-frequency real-time channel: 20Hz frequency, used to transmit continuously updated data required for monitoring and visualization, including: spindle load, triaxial vibration intensity, synthetic vibration intensity, coordinates of each axis, and toolpath transformation matrix. The front end subscribes to this channel to achieve real-time curve updates and 3D scene driving.

[0063] Low-frequency real-time channel: 1Hz frequency, used to transmit data with strong randomness such as model generation and interface requests, including: envelope curve update requests, optimization parameter query results, and historical data statistics results.

[0064] Command publishing channel: The front end sends command requests through this channel, such as "Start learning model", "Save current record", "Query historical alarms". The request format is a JSON string, which includes command type, parameters and timestamp.

[0065] Command feedback channel: After the backend completes the processing, it returns the result through this channel. The frontend subscribes to it and updates the UI state, realizing the classic request-response pattern.

[0066] The data storage layer adopts a MySQL 8.0 sharding and partitioning architecture, with the main data tables including: process_record: The processing record table stores metadata for each processing operation, including program name, tool number, start time, end time, and operator name; sensor_data: Sensor data table, which stores time-series data such as vibration intensity and spindle load by time partitioning; alarm_record: An alarm record table that stores the time, type, level, and preceding and following waveform data of abnormal events; optimization_params: Optimization parameter table, storing the P1, P2, and Vref values ​​of the program name and its tool number; model_version: The model version table stores the version, parameters, and accuracy metrics of the envelope curve model.

[0067] 3. Taking the CNC machining of a certain aerospace cabin structural component as an example, the specific working steps of this system are explained in detail.

[0068] 3.1 Initialization phase.

[0069] After powering on, the operator opens the front-end monitoring interface, which is initially grayed out. Users can then select the desired monitoring functions using the top toolbar. Select "Spatiotemporal Monitoring": Enable 3D digital twin and real-time curve display; Check "Anomaly Monitoring": Enable the envelope curve anomaly detection function; Check "Adaptive": Enable the feed rate adaptive optimization function; Clicking the "Connect to Database" button sends a connection request from the front end through the command publishing channel. After the back end verifies the connection status between Redis and MySQL, it returns a connection success message through the command feedback channel. The database icon in the bottom status bar changes from gray to green, indicating that the system is ready.

[0070] 3.2 Data Acquisition Stage.

[0071] After the machine tool starts, the hardware interaction layer three-source acquisition program runs automatically: The sensor acquisition program continuously acquires vibration signals at a frequency of 2560Hz, calculates the triaxial vibration intensity, and then sends it to the server via UDP. The HNC adapter / NC-Link protocol stack polls PLC data every 100ms, including spindle load P(t), coordinates of each axis (X,Y,Z), feed rate F(t), and spindle speed S(t). The NC code parser reads the current line number L(t) and code content in real time.

[0072] The server-side real-time service process reads raw data from the hardware interface every 50ms to perform data cleaning and time alignment. The PTP precise time protocol is used to synchronize the system time of each acquisition device, and cubic spline interpolation is performed on asynchronously arriving data to ensure that the time error of each data stream is <1ms; The 3σ principle was used to eliminate gross errors, and a combination of forward padding and linear interpolation was used for missing data, with the window length set to 10ms.

[0073] The processed data is written to Redis in key-value pairs and simultaneously published via a high-frequency real-time channel (20Hz) for front-end monitoring and visualization.

[0074] 3.3 Monitoring and Visualization Stage.

[0075] The state machine and UI interaction unit determine the machine tool status based on real-time data: Initial state: Standby (S0), the indicator light on the top bar of the UI is blue, and the 3D scene displays a static workpiece model; After the program is loaded: Start (S1), the indicator light flashes yellow, and the NC code bar loads the current program; Spindle rotation and feed start: Normal machining (S2), indicator light is green, 3D scene begins dynamic update; Program End or Emergency Stop: End (S3), indicator light grays out, data cleared.

[0076] During machining, the spatiotemporal coupling monitoring unit displays four dynamic curves in real time at a 20Hz refresh rate: spindle load, milling channel X-axis vibration, milling channel Y-axis vibration, and milling channel Z-axis vibration. The gray dashed line overlaid on the curves represents the historical average, and the red solid line represents the alarm threshold. The deviation calculation formula is as follows: when When the value is greater than 3, the corresponding point on the curve is marked with a red dot to alert the operator.

[0077] The spatial position monitoring unit constructs a 3D rendering scene based on the Python VTK engine. It reads the workpiece STL model (a cabin blank with approximately 50,000 triangular faces) from MySQL, reads the tool position coordinates in real time from Redis, and updates the tool model position through a homogeneous transformation matrix.

[0078] The color depth maps to the vibration intensity at a corresponding location; the color mapping function is as follows:

[0079] Where f is a rainbow lookup table, red indicates the high vibration area of ​​the internal cavity machining area of ​​the cabin, and blue indicates the low vibration area of ​​the end face machining area of ​​the cabin.

[0080] The NC code bar highlights the current line of execution in real time. When the operator clicks on any line of code, the 3D view automatically rotates to the tool position corresponding to that line of code, facilitating quick positioning.

[0081] 3.4 Learning Model Construction Stage.

[0082] For this cabin machining program (program name "1CFA103005_001", tool number "T01"), the operator selected 5 historical qualified machining records for model training: In the anomaly monitoring interface, click "Create New Model" and enter the model name "1CFA103005"; Select the program name "1CFA103005_001" and the tool number "T01", and the system will automatically filter out 20 historical records; Manually remove records containing abnormal events and select 5-8 high-quality records as training samples; Hyperparameter settings: Time window width W = 0.5s, step size = 0.25s, alarm boundary bias value Δ alarm =0.2g, wear boundary deviation value Δ wear =0.1g, lower limit of load bias Δ low =0.05×(P2-P1), boundary coefficient δ=0.2; Click "Start Training" to initiate the model training task in a non-real-time service process.

[0083] The training process is as follows: A dynamic time warping algorithm is used to align the timelines of the five records to ensure that each record corresponds to the same processing position. The sliding window iterates through each record, performing calculations within each window: Merge 5 records and calculate the mean and standard deviation for each window position: Generate three envelope curves: B alarm (t)=U env (t)+0.2×σ U (t) B wear (t)=U env (t) 0.1×σ U (t) B low (t)=L env (t) 0.05×σ L (t) After training, the model is saved to the model_version table in MySQL with the version number "V1.0". At the same time, the envelope curve is displayed in the 3D view as a semi-transparent curved surface for operators to visually confirm.

[0084] 3.5 Adaptive optimization stage.

[0085] The core algorithm layer's real-time service process reads the current program name "1CFA103005_001" and tool number "T01," and matches the pre-calculated parameters from the optimization parameter database. No-load load P1: By analyzing historical no-load travel data, extract the loads that satisfy P1≥P min +0.1×(P max -P min The minimum power value with a duration ≥ 5s, in this example P min =0.03kW, P max =0.15kW, calculated to P1=0.05kW; Target load P2: Based on the statistical processing data during the stable period, select the minimum power value that is lower than 95% of the data. In this example, P2 = 0.12kW. Vibration reference V ref : The average vibration intensity corresponding to the P2 load. In this example, V ref = 0.8g.

[0086] The real-time service process reads the current spindle load P(t) at a 50ms cycle and compares it with P1 and P2: If P(t) < P1 (idle state), output the maximum feed rate multiple of 150% and quickly move to the next cutting position; If P(t) > P2 (overload state), output the minimum feed rate multiple of 30% to prevent tool damage; If P1 ≤ P(t) ≤ P2 (normal cutting), run the fuzzy PID control algorithm.

[0087] At a certain moment in this example, P(t) = 0.1kW, the calculation error E = 0.1 - 0.12 = -0.02kW, and the error differential Ec = 0.005kW / s. The quantization factors are set as follows:

[0088]

[0089] Map to the fuzzy domain:

[0090]

[0091] Query the fuzzy rule table, output U = ZO, and after defuzzification, get ΔU = 0. The feed rate multiple remains the current value. At the same time, introduce a vibration feedback correction term. The current vibration intensity V(t) = 0.85g, and the vibration suppression coefficient η = 0.2:

[0092] The final optimized value is written into the R1 parameter of the Siemens 840D system and activated through the synchronous action instruction. The hardware interaction layer writes and checks through the DB21 data block, and returns a successful status after confirming that the multiple takes effect.

[0093] 3.6. Abnormal recognition stage.

[0094] The abnormal recognition sub-module running in the real-time service process compares the current vibration intensity V(t) with the envelope curve model at a frequency of 20Hz: At time t1: V(t1) = 1.2g, B alarm (t1) = 1.1g → V(t) > B alarm (t1), it is determined as a serious abnormality; Trigger a three-level emergency stop: Output the E-stop signal through the hard-wired IO, and at the same time trigger the emergency stop instruction by software; The alarm and backtracking unit records the complete waveform 5 seconds before and after the abnormal moment, and automatically takes a snapshot of the 3D scene. The UI interface flashes red and beeps continuously, while a pop-up window displays "Collision Warning - Emergency Stop Triggered"; Time t2: V(t2) = 0.95g, B alarm (t2) = 1.0g, B wear (t2) = 0.9g → B wear <V(t)≤B alarm ; If the problem is determined to be tool wear or a minor abnormality, a level two alarm is triggered. The UI interface is constantly lit in red, with intermittent buzzing alarm sounds, and data is automatically saved for 2 seconds before and after the event. The suggested message reads "Increased tool wear, inspection recommended." At time t3: V(t3) = 0.3g, Blow(t3) = 0.35g → V(t) <Blow(t); If the system is determined to be loose or has cut off, a Level 1 warning is triggered. The UI interface flashes yellow, records event logs, and does not trigger a shutdown. 3.7. Anomaly backtracking and historical query.

[0095] When an alarm occurs, the operator can click the "Anomaly Retrospective" button: In the 3D view, the tool position and trajectory segment at the time of the anomaly are highlighted in red. The abnormal periods are marked with red background bars in the graph, and detailed parameters (time, alarm type, V(t), threshold) are displayed when the mouse is hovered over them. Supports 5-second playback before and after, with adjustable playback speed.

[0096] The historical search function supports searching by conditions: Time range: 08:00 on October 15, 2023 to 20:00 on October 15, 2023; Program name: 1CFA103005_001; Alarm type: Level 2 alarm; Query results: 3 records; Clicking on any record will automatically jump to the corresponding time point for playback, simultaneously displaying the 3D model status, real-time curves, and alarm information.

[0097] 3.8. End Phase.

[0098] After processing is completed, the state machine detects the program end signal and transitions to the end state (S3): The indicator light in the top UI bar is grayed out. Real-time curves stop updating, and the 3D scene reverts to static state. Automatically save the current machining record to MySQL, including: start / end time, program name, tool number, operator, alarm event list, and optimization parameter usage. If the processing time is less than 10 seconds (due to accidental triggering or debugging), the system will automatically delete the invalid record.

[0099] Clean up temporary data in Redis to free up memory resources.

[0100] The present invention features decoupling of each layer and high data transmission efficiency, which effectively improves the real-time monitoring of the processing process, the accuracy of optimization, and the reliability of anomaly identification, thereby reducing losses caused by abnormal operating conditions.

[0101] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A system for adaptive optimization of processing and identification of abnormal working conditions, characterized in that, A layered and decoupled architecture is adopted, which includes: a front-end presentation layer, a core algorithm layer, a hardware interaction layer, and a data transfer and storage support layer; The front-end display layer is the display and interaction interface of the recognition system, and its display content is obtained from the core algorithm layer through the data transfer and storage support layer. The core algorithm layer adopts a separate architecture including real-time services and non-real-time services. The real-time services are used to complete real-time data processing, and the non-real-time services are used to execute time-consuming tasks. The real-time services and the non-real-time services are decoupled from each other. The hardware interaction layer communicates bidirectionally with the machine tool CNC system and sensors via a CNC system communication adapter and an external sensor acquisition card. The bidirectional data interaction includes data acquisition during the machining process and feedback control of the machining process. The data relay and storage support layer adopts the Redis publish-subscribe mechanism, and the data relay and storage support layer is divided into multiple fixed data transmission channels to achieve data transmission decoupling; The identification system uses a three-source data acquisition method to obtain full-dimensional data of the processing process.

2. The adaptive optimization and abnormal working condition identification system for the processing procedure according to claim 1, characterized in that, The real-time service uses a fixed calculation cycle of 50ms. And / or, the plurality of fixed data transmission channels include: a high-frequency real-time channel, a low-frequency real-time channel, a command issuance channel, and a command feedback channel; the high-frequency real-time channel has a data transmission frequency greater than or equal to 20Hz and is used for continuous data update and transmission; the low-frequency real-time channel has a data transmission frequency less than 20Hz and is used for continuous data update and transmission; the command issuance channel is used to issue request commands to the front-end interactive interface; the command feedback channel is used to return the back-end processing results; And / or, the data acquired using the three-source data acquisition method includes: internal machine tool data acquired from the machine tool CNC system via the HNC adapter or NC-link protocol, external machining data acquired from external vibration sensors via the sensor acquisition card, and NC code execution data during the machining process.

3. The adaptive optimization and abnormal working condition identification system for the processing procedure according to claim 1, characterized in that, The core algorithm layer includes: a multi-dimensional monitoring and visualization module, an adaptive optimization control module, and an abnormal operating condition identification and prediction module; The multi-dimensional monitoring and visualization module constructs a spatiotemporally coupled digital twin monitoring environment, dynamically correlates the time-domain monitoring signals of the processing process with the three-dimensional spatial position, and realizes multi-dimensional visualization of the CNC machining process in the workshop through color mapping and state machine. The adaptive optimization control module is based on an optimization parameter database constructed from real-time load data and offline analysis. It uses a fuzzy control algorithm to dynamically calculate the feed rate optimization value and writes and executes the optimization parameters in real time through the CNC system interface. The abnormal working condition identification and prediction module establishes a learning and monitoring mechanism based on the spindle power signal. It generates an envelope curve model by performing sliding window statistics and boundary fusion on multiple historical qualified machining records of the same CNC machining task. During the machining process, it monitors in real time whether the vibration intensity exceeds the alarm boundary, wear boundary or load lower limit, thereby realizing the identification and early warning of abnormal working conditions including chipping, tool collision and clamping loosening.

4. The adaptive optimization and abnormal working condition identification system for the processing procedure according to claim 3, characterized in that, The multi-dimensional monitoring and visualization module includes: a spatiotemporal coupling monitoring unit, a spatial location monitoring unit, a state machine and UI interaction unit, and a UI interface partition design unit. The spatiotemporal coupling monitoring unit uses the ECharts dynamic line chart component to display the spindle load, triaxial vibration intensity, synthetic vibration intensity and time domain signal in real time at a preset refresh rate, and supports switching between multiple data sources and filtering of data items. Real-time communication between the front-end and back-end is achieved through a WebSocket push mechanism. Historical average curves and alarm threshold lines are overlaid on the chart, and the deviation between the real-time status and the historical baseline is calculated using the following formula: in, The signal value at the current moment. This represents the average signal value at the corresponding position in the same historical code segment. For standard deviation, when An exception flag is triggered when the value is greater than 3; The spatial position monitoring unit constructs a 3D rendering scene based on the Python VTK engine, and sets up a CuttingTool class and a Part class; the CuttingTool class includes tool type, tool diameter, tool cutting edge length, and tool overhang attributes; the Part class includes workpiece name, workpiece material, workpiece STL model vertex array, and workpiece normal vector attributes; The spatial position monitoring unit reads the tool and workpiece models in STL format, constructs an octree spatial index to accelerate collision detection, and realizes real-time visualization of the tool motion trajectory through a 4×4 homogeneous coordinate transformation matrix. The homogeneous transformation matrix is ​​as follows: in, It is a 3×3 rotation matrix. The vector is a 3×1 translation vector, and T is a 4×4 homogeneous transformation matrix; simultaneously, the data size is processed by mapping the color depth to the corresponding position, and the color mapping function is: in, For position Vibration intensity or load value at the location, , These are the minimum and maximum values ​​of the currently displayed data across the entire domain. A function for a rainbow-colored lookup table; The state machine and UI interaction unit construct a finite state machine model, defining stable states and state transition events; the stable states include S0 standby, S1 start, S2 processing, and S3 end; the UI top bar status indicator color is updated in real time through the state transition events: blue for standby, yellow flashing for start, green for processing, and gray for end, and the corresponding business logic is triggered when the state changes, and the optimization parameter library is automatically loaded when entering the processing state; The UI interface partition design unit adopts a Flex flexible layout to divide the interface into multiple areas, including: top toolbar, left status bar, central main view area, bottom NC code bar, and bottom status bar. The top toolbar is used to switch between spatiotemporal monitoring views and abnormal monitoring views, as well as to enable adaptive function switches and system settings. The left status bar displays the machine tool status description, current program name, operator information, and current alarm list. The central main view area displays a 3D rendering canvas and real-time curve graphs. The bottom NC code bar highlights the currently executing code line in real time and supports clicking to navigate to the corresponding position in the 3D view. The bottom status bar displays the database connection status, the activation status of each functional module, system time, and network latency.

5. The adaptive optimization and abnormal working condition identification system for the processing procedure according to claim 3, characterized in that, The adaptive optimization control module includes: an optimization parameter database construction unit, an intelligent control algorithm unit, and a parameter execution interface unit; The optimization parameter database construction unit analyzes historical machining data offline and extracts the no-load load P1 and target load P2 for each combination of program name and tool number; it then uses a sliding window algorithm with a preset window width and preset step size to scan the no-load segment and extract a stable no-load segment that meets the conditions. The no-load load P1 is defined as satisfying P1≥P min +0.1×(P max P min The minimum power value for which the continuous data duration is not less than 5 seconds and the load value is less than P1 is selected, and transient fluctuation data points at the moment the cutting fluid is turned on or off are excluded; P max This refers to the maximum load power collected during the machining process corresponding to the same program name and tool number combination; P min This refers to the minimum load power collected during the machining process corresponding to the same program name and tool number combination; The target load P2 is defined as the minimum power value in the processing stability section where the proportion of data points with load values ​​less than P2 is not less than 95%, used to characterize the typical load level in the processing stability section; at the same time, the vibration intensity under the corresponding load is recorded as a reference benchmark V. ref ; The intelligent control algorithm unit adopts a two-dimensional fuzzy PID control algorithm, with the error E(t) between the real-time load P(t) and the target load P2 being equal to P2. P(t) and the error derivative Ec(t) = dE / dt are the input quantities, quantized by factor K. e K ec Mapping to the fuzzy domain: E * =round(K e E),K e =6 / P2×0.5 Ec * =round(K ec Ec),K ec =6 / P2×0.1 Among them, E * and Ec * The input fuzzy subsets are all {NB, NM, NS, ZO, PS, PM, PL}; the universe of discourse for error E(t) is {-6, -4, -2, 0, 2, 4, 6}; the universe of discourse for error differential Ec is {-6, -4, -2, 0, 2, 4, 6}; the output is the feed rate increment ΔU, with the same fuzzy subset as the input, and the universe of discourse is {-3, -2, -1, 0, 1, 2, 3}. The membership function adopts a triangular distribution. A fuzzy rule table is constructed, fuzzy inference is performed using the Mamdani synthesis inference method, and fuzziness is resolved using a weighted average method. Where, μ i Let C be the activation degree of the i-th rule. i The center value of the fuzzy set is output for the corresponding rule; the final output control quantity ΔU=Ku ΔU*, Ku is the scaling factor; the actual feed rate optimization value F opt (t+1)=F opt (t)+ΔU, and limited to the range of 30%~150%, F opt (t) refers to the actual optimized feed rate at the current time t; a vibration feedback correction term is also introduced: Where V(t) is the real-time vibration intensity, The target vibration reference is η, which is the vibration suppression coefficient; when the vibration intensity exceeds 1.5 times... At that time, active speed reduction control is executed; The parameter execution interface unit is designed for CNC systems. It writes R parameters through the OPC UA method node and calls synchronous action instructions to activate the magnification coverage. It adopts a dual-channel redundant writing mechanism, using dual writing through PLC data block channel 1 magnification DB21 and axis magnification DB31, to ensure that the magnification activation success rate is not less than 99.9%. Final feed rate F final =F opt ×F panel / 100%, of which F panel Set the control panel knob value to achieve a fusion control of optimized magnification and manual settings.

6. The adaptive optimization and abnormal working condition identification system for the processing procedure according to claim 3, characterized in that, The abnormal operating condition identification and prediction module includes: a learning model construction unit, a collision detection unit, and an abnormal alarm and backtracking unit; The learning model construction unit is built based on the recording, learning and monitoring mechanism of the spindle power signal; the learning model construction unit is used to analyze the temporal characteristic changes of the spindle power signal in order to realize real-time judgment of abnormal working conditions including chipping, tool collision, and loose clamping. The learning model building unit supports training corresponding independent learning models for different CNC programs. Each model is associated with the program name and tool number and is used for anomaly monitoring and identification in the repeated machining process of the same program. The process of generating the learning model includes the following steps: Sample selection steps: Select one or more qualified machining records with the same program name and the same tool number as the training sample set; Parameter initialization steps: Set the time window length W to 0.5s and the alarm boundary bias value Δ alarm 0.2g, wear boundary deviation value Δ wear The load lower limit bias value Δ is 0.1g. low It is 0.05 times the load fluctuation range; Time axis alignment step: The dynamic time warping algorithm is used to uniformize the timestamps of each processing record in the training sample set to align the time axes of multiple processing records. Early warning boundary calculation steps: For each processing record, traverse the entire processing process with a time window length W, and calculate the boundary value at each time window position t. i At this point, the statistical characteristics of the monitoring signal within the calculation window are used to generate the upper and lower warning boundary sequences for this record; The upper warning boundary point is defined as: The lower warning boundary point is defined as: in, This is the spindle power signal sequence within this window. This represents the maximum power within the window. This represents the minimum power value within the window. The average power within the window. These are boundary coefficients; Envelope curve fusion steps: For a sample set containing K learning records, at position t in the i-th time window... i At this point, the upper and lower warning boundaries generated from each record are merged to obtain the final envelope curve model: Upper envelope curve: , Lower envelope curve: , in, , These represent the upper and lower warning boundaries of the k-th record in the i-th window. , The paranoia value set for the user. The value of Δ corresponds to alarm , The value of Δ corresponds to wear or Δ low ; Dynamic boundary generation steps: based on the fused envelope curve , Generate three boundary curves for real-time monitoring: Alarm boundary: B alarm (t)=U env (t)+Δ alarm ×σ U (t) Wear boundary: B wear (t)=U env (t) Δ wear ×σ U (t) Lower load limit: B low (t)=L env (t) Δ low ×σ L (t) Where, σ U (t), σ L (t) represents the standard deviation of each learning record at time t corresponding to the upper and lower boundaries, which is used to characterize the data fluctuation of the sample set; During the machining process, the collision detection unit is used to monitor the vibration intensity signal V(t) in real time and compare it with the three boundary curves to determine the tool condition. When V(t) > B alarm When (t) occurs, it is determined to be a serious abnormality, including chipping or collision, triggering an emergency stop command; When B wear (t) <V(t)≤B alarm If the error is detected as tool wear or a minor abnormality, a level two alarm will be triggered. When V(t) low When (t) occurs, it is determined that the clamping is loose or the air cutting is abnormal, triggering a level one warning.​ The learning model building unit supports model version management and iterative optimization. The specific steps are as follows: Model evaluation steps: Select the optimal model by comparing the prediction errors of multiple learning models on the same validation set; the prediction errors include root mean square error (RMSE) and mean absolute percentage error (MAPE). The formula for calculating the root mean square error (RMSE) is as follows: The formula for calculating the Mean Absolute Percentage Error (MAPE) is as follows: Incremental update steps: Periodically use newly accumulated qualified processing data to incrementally learn and update the selected optimal model to adapt to changes in production conditions.

7. The adaptive optimization and abnormal working condition identification system for the processing procedure according to claim 6, characterized in that, The collision detection unit is used to monitor the vibration intensity of the turning / milling channel in real time. A collision judgment threshold is set by a preset formula. When the real-time monitored vibration intensity exceeds the collision judgment threshold, a machine tool stop signal is triggered. The formula for calculating the collision determination threshold is as follows: T collision =V ref ×(1+a)+b s ref Among them, V ref For reference vibration intensity, α is the base percentage coefficient, β is the dynamic margin coefficient, and σ ref For reference vibration standard deviation; The reference vibration intensity is obtained by the learning model construction unit through learning qualified processing records; the reference vibration standard deviation corresponds to the reference vibration intensity and is obtained synchronously by the learning model construction unit through learning qualified processing records.

8. The adaptive optimization and abnormal working condition identification system for the processing procedure according to claim 6, characterized in that, The abnormal alarm and backtracking unit triggers graded alarms and shutdown commands based on the comparison results between the vibration intensity and the preset boundary: When the vibration intensity exceeds the wear boundary or the lower limit of the load, the abnormal alarm and backtracking unit triggers a graded alarm. When the vibration intensity exceeds the alarm threshold, the abnormal alarm and backtracking unit triggers a machine tool stop command. The abnormal alarm and backtracking unit supports the abnormal event backtracking function, and can retrieve and analyze the parameters, waveforms and status records related to the alarm during the current and historical processing.

9. A method for adaptive optimization of processing and identification of abnormal working conditions, characterized in that, The adaptive optimization and abnormal working condition identification system for the processing procedure as described in any one of claims 1 to 8 includes the following steps: System initialization steps: Start the system's layered decoupled architecture, load the CNC machine tool configuration parameters, sensor acquisition parameters and the system's stored learning model library, complete the initial deployment of each functional module, and ensure normal linkage between the front-end display layer, core algorithm layer, hardware interaction layer and data transfer and storage support layer; Learning model construction steps: For a preset CNC machining program, select the qualified machining records corresponding to the history of the CNC machining program as training samples. Through the learning model construction unit of the abnormal working condition identification and prediction module of the core algorithm layer, the envelope curve model is generated by the sliding window statistics and boundary fusion algorithm. The envelope curve model is associated with and stored with the CNC machining program and the corresponding tool number. Real-time machining monitoring steps: The machining process data is acquired in all dimensions through the system’s three-source data acquisition method, forming three-source data. The three-source data includes internal data acquired from the machine tool CNC system through the HNC adapter or NC-link protocol, external data acquired from the external vibration sensor through the sensor acquisition card, and NC code execution data. The acquired data is distributed to the corresponding functional modules through multiple fixed channels of the data transfer and storage support layer. Multi-dimensional monitoring and visualization steps: Through the multi-dimensional monitoring and visualization module of the core algorithm layer, based on the real-time data distributed in the real-time processing monitoring steps, the three-dimensional digital twin monitoring scene is updated and the real-time dynamic curve is drawn, so as to realize the multi-dimensional visualization of the processing process. Adaptive optimization control steps: The adaptive optimization control module of the core algorithm layer, based on the fuzzy control algorithm and combined with real-time load data and optimization parameter database, calculates the feed rate optimization value in real time, and writes it into the CNC system for execution through the parameter execution interface unit; Abnormal working condition identification steps: Through the collision detection unit and learning model construction unit of the abnormal working condition identification and prediction module of the core algorithm layer, the spindle power signal and vibration signal are collected. The real-time vibration signal is compared with the alarm boundary, wear boundary and load lower limit corresponding to the envelope curve model generated in the learning model construction step to determine the abnormal processing state. The abnormal alarm and backtracking unit of the abnormal working condition identification and prediction module of the core algorithm layer triggers the corresponding graded alarm, shutdown or early warning processing. Data storage and model iteration steps: Processing data collected in the real-time processing monitoring step, optimization parameter data in the adaptive optimization control step, and abnormal condition identification data in the abnormal condition identification step are stored in the database of the data transfer and storage support layer. Newly accumulated qualified processing data are used periodically to incrementally train the learning model through the learning model building unit to achieve model iterative optimization.

10. The method for adaptive optimization of the processing process and identification of abnormal working conditions according to claim 9, characterized in that, In the real-time processing monitoring step, the three-source data include: First source data: Spindle load, axis coordinates, and feed rate are acquired from the machine tool CNC system via an HNC adapter or NC-link protocol; Second source data: Three-dimensional vibration signals of the turning / milling channel are acquired from an external vibration sensor via a sensor acquisition card; Third source data: Obtain the current execution line number and code content through NC code parsing; The data relay and storage support layer uses a Redis publish-subscribe mechanism to divide the data into four fixed data transmission channels, which are as follows: The high-frequency real-time channel transmits monitoring data during the processing and data required for updating the digital twin scene at a data transmission frequency of greater than or equal to 20Hz. The low-frequency real-time channel transmits model-generated and interface-requested data at a data transmission frequency of less than 20Hz. The command issuing channel is used to receive request instructions issued by the front-end presentation layer and forward the request instructions to the corresponding back-end functional modules; The command feedback channel is used to receive the processing results of various backend functional modules and forward the processing results to the frontend display layer to realize data interaction between the frontend and backend.

Images