A precision part production control method and system based on big data analysis

Abstract

The application discloses a precision part production control method and system based on big data analysis, and relates to the technical field of numerical control machining data processing. Multi-source data and laboratory data are acquired, migration is performed to obtain multi-source migration data; a signal segment is input into a channel attention mechanism module for processing to obtain a channel enhanced signal, and the channel enhanced signal is input into the module to obtain multi-scale deep features; a pooling operation is performed to obtain a reduced dimension feature vector, and a classification network is calculated to obtain a cutting state; if the cutting state result is correct, it is determined that the detection model training is qualified, real-time sensing signals during workpiece cutting are acquired, real-time cutting states are obtained by detecting the real-time sensing signals according to the detection model, and if the real-time cutting state is chatter, lathe control is performed. The channel attention and the module are used to improve the chatter recognition accuracy under a small sample, the lathe intervention can be performed after the chatter is recognized, the chatter is inhibited, the machining precision of the precision part is protected, and the abnormal wear of a tool is reduced.

Description

Technical Field

[0001] This invention relates to the field of CNC machining data processing technology, specifically to a precision parts production control method and system based on big data analysis. Background Technology

[0002] In the production of precision parts on lathes, workpiece vibration is a prominent factor affecting product quality. The lathe cutting process is influenced by a combination of factors, including spindle speed, feed rate, clamping conditions, tool wear, workpiece length-to-diameter ratio, and workshop environment, making workpiece vibration highly susceptible to occur. This vibration leads to decreased machining accuracy and exacerbates abnormal tool wear, increasing production scrap costs. Traditional production methods rely heavily on fixed process parameters and operator experience for production control, primarily using fixed vibration thresholds for vibration assessment.

[0003] Publication No. CN121559976A discloses a data processing method and system for multi-axis linkage CNC lathe machining; it acquires vibration information, performs time-frequency analysis, and extracts energy changes and instantaneous peak characteristics within a preset high-frequency band; it compares the characteristics with preset judgment criteria to identify early signs of chatter, and generates machining parameter adjustment commands for spindle speed, feed rate, and / or depth of cut when chatter is identified; it acquires the adjusted vibration information for time-frequency analysis, evaluates the chatter suppression effect, and further adjusts machining parameters and / or issues early warnings or shutdown alarms accordingly, thereby forming a closed-loop active suppression control for micro-cutting chatter.

[0004] Existing methods for detecting chatter in precision machining have limitations. Under limited laboratory sample conditions, different sensor channels contribute differently to chatter identification and are susceptible to noise interference. Traditional networks struggle to fully extract key local features of chatter in the temporal dimension of vibration signals, and their generalization ability is insufficient in small sample scenarios, leading to misjudgments. Furthermore, they only identify the chatter state and lack the production control logic to interface with the machine tool after identification. This prevents timely intervention after chatter occurs, resulting in decreased workpiece machining accuracy and abnormal tool wear, which fails to meet the actual control requirements of precision parts machining. Summary of the Invention

[0005] The purpose of this invention is to solve the problems mentioned in the background art, such as the susceptibility of existing cutting chatter detection to noise interference, poor generalization under small samples, limited recognition accuracy, and lack of subsequent machine tool intervention methods, which can easily lead to damage to the precision of precision parts and abnormal wear of tools. Therefore, this invention proposes a precision parts production control method and system based on big data analysis.

[0006] A first aspect of this invention provides a method for controlling the production of precision parts based on big data analysis, the method comprising: Acquire multi-source data and laboratory data, and perform migration on the multi-source data based on the laboratory data to obtain multi-source migrated data; the multi-source migrated data contains multiple signal segments. The signal segment is input into the channel attention mechanism module for processing to obtain the channel enhancement signal, and the channel enhancement signal is input into the BSR module to obtain multi-scale depth features; the BSR module consists of a BayesformerEncoder stacking submodule, an SE channel attention submodule, and a residual feature fusion submodule; Pooling operations are performed on multi-scale depth features to obtain dimensionality-reduced feature vectors. Classification network calculations are then performed on the dimensionality-reduced feature vectors to obtain the cutting state; the cutting state includes stability or chatter. If the cutting state result is correct, the detection model is deemed to have passed training, and the real-time sensing signal during workpiece cutting is acquired. The real-time cutting state is obtained by detecting the real-time sensing signal according to the detection model. If the real-time cutting state is chatter, lathe control is performed.

[0007] The acquired multi-channel sensor data undergoes data migration processing. Through channel attention mechanisms and a BSR module, feature mining is performed on the signals, enabling the differentiation of the contribution of different channel signals to chatter identification. This fully extracts deep features related to the cutting state from the signals, improving the model's ability to distinguish between stable cutting states and chatter states, and reducing recognition interference from noise. After iterative training and performance verification to obtain a qualified detection model, online identification of sensor signals during the workpiece cutting process can be performed. Once chatter is detected, the machine tool can be linked to execute corresponding process intervention operations.

[0008] Optionally, the process of acquiring the multi-source data includes: A differential equation for the dynamics of a precision component-tool is constructed. The equation is then assigned values ​​to obtain excitation conditions. These excitation conditions are then subjected to variable-step time-domain numerical iterative solution to obtain a time-series parameter response sequence. This time-series parameter response sequence includes a time-domain displacement response sequence and a time-domain velocity response sequence. The first basic data is obtained by performing calculations on the time-domain displacement response sequence and the time-domain velocity response sequence, the second basic data is obtained by performing calculations on the time-domain displacement response sequence, and the third basic data is obtained by performing calculations on the time-domain velocity response sequence. The first basic data, the first basic data and the first basic data are fused to obtain multi-source data.

[0009] Optionally, performing migration on the multi-source data based on the laboratory data to obtain multi-source migrated data includes: Physical rule hard filtering is performed on the multi-source data to obtain a clean signal. Z-score normalization and sliding window slicing are performed on the clean signal to obtain unlabeled data. A preset model is obtained, and the preset model is trained on the laboratory data to obtain an initial model. The unlabeled data is input into the initial model to perform forward inference to obtain multiple probability value sets. The mean and standard deviation are calculated based on each probability value set. After filtering each probability value set according to the mean and standard deviation and adding labels, the automatically labeled data is obtained. Based on the mean and standard deviation, each probability value set is filtered according to the second preset rule to obtain fuzzy samples. The fuzzy samples are manually labeled to obtain manually labeled data. The probability value set and the manually labeled data are fused to obtain multi-source migration data.

[0010] By performing hard filtering on raw industrial multi-source data using physical rules to remove obviously abnormal signal segments, and then performing normalization and slicing, the initial model trained on laboratory data is used to perform multiple inferences on unlabeled industrial data. The confidence level of samples is distinguished based on the mean and standard deviation of the inference results. High-confidence samples are automatically pseudo-labeled, while a small number of fuzzy samples with high uncertainty are manually reviewed and labeled. The two types of labeled data are then fused according to different weights to form multi-source transfer data. This approach can make full use of the large amount of unlabeled time-series signals in industrial scenarios, reduce the workload of manual labeling, and reduce label noise caused by automatic labeling. It balances data scale and labeling quality, realizes the transfer and adaptation from laboratory datasets to actual industrial data, and improves the model's recognition performance under real processing conditions.

[0011] Optionally, the signal segment is input into the channel attention mechanism module for processing to obtain the channel enhancement signal, including: Perform global average pooling on the signal segment to obtain the global average pooling feature; Perform global max pooling on the signal segment to obtain the global max pooling feature; The global average pooling feature and the global max pooling feature are concatenated to obtain the combined pooling feature. The combined pooling features are input into a two-layer MLP (Multilayer Perceptron) network, and ReLU activation and Sigmoid activation operations are performed respectively to obtain the channel weight coefficients. The channel weighting coefficients are multiplied one by one with the corresponding signal segments to obtain the channel-enhanced signal.

[0012] Statistical features of signal segments are extracted using two different pooling operations. The two types of features are then concatenated and fed into a multilayer perceptron network for nonlinear transformation to generate channel weight coefficients adapted to the flutter recognition task. The weights are then used to perform channel-by-channel weighting on the original signal segments, redistributing the signal contribution of each sensor channel. This strengthens the information of channels with strong correlation to flutter features and weakens the information of channels with more noise interference, thereby improving the ability of subsequent networks to mine effective features of cutting flutter.

[0013] Optionally, inputting the channel enhancement signal into the BSR module to obtain multi-scale depth features includes: Sliding window block processing is performed on each channel of the channel enhancement signal to obtain block signals. The block signals are then input into the encoder to perform variational self-attention (VSA) operation to obtain temporal attention weights. The first temporal feature is obtained by performing feature weighting on the block signal according to the temporal attention weight. The first temporal feature is then passed through the Add&Norm residual normalization layer and the FeedForward network layer to complete feature transformation and obtain the second temporal feature. The second temporal feature is then subjected to repeated multi-layer BayesformerEncoder stacking operation to obtain the deep temporal feature. The deep temporal features are input into the SE channel attention module to obtain intermediate features. The channel enhancement signal is then added to the intermediate features to obtain multi-scale depth features.

[0014] The channel enhancement signals of each channel are segmented and processed. Temporal attention weights are obtained by variational self-attention operation. Based on the weights, signal features are weighted and then feature transformation is performed through residual normalization and feedforward network. After stacking multiple encoders, key information in the temporal dimension is fully extracted. Channel features are filtered again by SE channel attention module. Finally, multi-scale depth features are output by combining residual addition. This can capture local temporal segments related to flutter in vibration signals, alleviate the degradation problem that is prone to occur during deep network training, and suppress redundant noise at the channel level.

[0015] Optionally, performing a classification network calculation on the reduced-dimensional feature vector to obtain the cutting state includes: Perform Flatten operation on the dimensionality-reduced feature vector to obtain a one-dimensional flattened feature vector, and input the one-dimensional flattened feature vector into a fully connected FC layer to obtain an intermediate feature vector; The intermediate feature vector is input into the Softmax activation function for normalization to obtain a two-dimensional probability vector. The two-dimensional probability vector is then analyzed to obtain the probability of stable cutting state and the probability of chattering state. The maximum value between the probability of a stable cutting state and the probability of a chattering state is taken, and the category corresponding to the maximum value is taken as the final judgment result of the signal segment. The cutting state is output based on the final judgment result.

[0016] Optionally, obtaining the real-time cutting state by detecting the real-time sensing signal according to the detection model includes: The real-time sensing signal is subjected to Z-score normalization to obtain a standard time-series signal, and the standard time-series signal is divided into multiple real-time signal segments by a sliding time window. Real-time signal segments are input into the detection model to obtain real-time judgment results. The real-time judgment results are aligned in chronological order to obtain continuous judgment results. If the number of chattering events in the continuous judgment results is greater than the alarm threshold, the real-time cutting state is determined to be chattering.

[0017] A second aspect of this invention provides a precision parts production control system based on big data analysis, the system comprising: A data migration module is used to acquire multi-source data and laboratory data, and perform migration on the multi-source data based on the laboratory data to obtain multi-source migrated data; the multi-source migrated data contains multiple signal segments. The feature extraction module is used to input signal segments into the channel attention mechanism module for processing to obtain channel enhancement signals, and input the channel enhancement signals into the BSR module to obtain multi-scale depth features; the BSR module consists of a BayesformerEncoder stacking submodule, an SE channel attention submodule, and a residual feature fusion submodule; The result determination module is used to perform pooling operations on multi-scale depth features to obtain dimensionality-reduced feature vectors, and to perform classification network calculations on the dimensionality-reduced feature vectors to obtain the cutting state; the cutting state includes stability or chatter. The production control module is used to determine that the detection model training is qualified if the cutting state result is correct, and to acquire the real-time sensing signal during workpiece cutting. The real-time cutting state is obtained by detecting the real-time sensing signal according to the detection model. If the real-time cutting state is chatter, the lathe is controlled.

[0018] The beneficial effects of this invention are: This invention proposes a precision parts production control method based on big data analysis. It obtains multi-source migrated data by migrating multi-source data from laboratory data. Relying on a channel attention mechanism, it re-weights and recalibrates the signals of each channel, strengthening effective channels related to chatter and suppressing redundant noise channels. Then, a BSR module extracts multi-scale deep features that consider both channel and time dimensions, fully exploring time-series segments within the vibration signal strongly correlated with chatter, improving the model's generalization performance under small sample conditions, and enhancing the classification accuracy of chatter and stable cutting states. After the model is successfully trained, online detection is performed on the real-time sensing signals of the workpiece cutting process. Once chatter is identified, lathe control intervention can be directly executed to promptly suppress chatter development, reduce workpiece precision degradation and abnormal tool wear caused by chatter, and achieve a closed loop from chatter identification to machining control, adapting to the machining control requirements of precision parts. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a precision parts production control method based on big data analysis, provided as an embodiment of the present invention. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] This invention provides a method for controlling the production of precision parts based on big data analysis. See also... Figure 1 The method includes the following steps: S101, acquire multi-source data and laboratory data, and perform migration on the multi-source data based on the laboratory data to obtain multi-source migrated data; S102, the signal segment is input into the channel attention mechanism module for processing to obtain the channel enhancement signal, and the channel enhancement signal is input into the BSR module to obtain multi-scale depth features; S103, perform pooling operation on multi-scale depth features to obtain dimensionality-reduced feature vectors, and perform classification network calculation on the dimensionality-reduced feature vectors to obtain the cutting state; S104, if the cutting state result is correct, the detection model is deemed to be qualified and the real-time sensing signal during workpiece cutting is obtained. The real-time cutting state is obtained by detecting the real-time sensing signal according to the detection model. If the real-time cutting state is chatter, the lathe is controlled.

[0022] Multi-source migration data contains multiple signal segments; the BSR module consists of a BayesformerEncoder stacking submodule, an SE channel attention submodule, and a residual feature fusion submodule; the cutting state includes stable or chattering; In one implementation, laboratory data is acquired by multi-channel sensor signals to obtain raw multi-channel signals P; multiple types of sensors are arranged on the milling test bench to acquire 7 channels of raw signals: triaxial acceleration signal, triaxial cutting support force signal, and acoustic signal; the sampling frequency is set to 10kHz to acquire time-series sensor raw signals under multiple cutting conditions such as different spindle speeds and axial depths of cut.

[0023] In one implementation, signal preprocessing yields a normalized time-series signal segment set: ① Z-score standardization is performed on all original multi-channel signals to eliminate differences in amplitude dimensions among sensors; ② A sliding time window is used to segment the long time-series signal, with each segment containing 1000 sampling points and a window step size of 0.05s, resulting in a large number of signal segments; ③ Combining the processing spectrum characteristics and workpiece surface processing morphology, each signal segment is manually labeled: 0 represents a stable cutting state, and 1 represents a chattering state; training set, validation set, and test set are obtained to acquire a labeled time-series signal segment dataset.

[0024] In one implementation, pooling operations are performed on multi-scale deep features to obtain dimensionality-reduced feature vectors: pooling dimensionality reduction yields dimensionality-reduced feature vectors; global max pooling and global average pooling are performed simultaneously on multi-scale deep features; the pooling results are concatenated to complete feature dimensionality reduction and obtain dimensionality-reduced feature vectors.

[0025] In one implementation, if the cutting state result is correct, the detection model is deemed to have passed training: The model is iteratively trained to obtain a converged ARENET complete model. ① The training set samples are fed into the complete network in steps 3 and 6 above; the loss function includes both cross-entropy loss and variational attention KL divergence loss; ② The Adam optimizer is used to backpropagate and update all network parameters; StepLR learning rate decay and dropout=0.3 are set to prevent overfitting; ③ The classification accuracy is evaluated on the validation set, and the iteration is repeated until the loss converges, and the trained ARENET model is saved; ④ The model is input using an independent test set to evaluate the accuracy and recall of flutter recognition, and the model performance is verified.

[0026] In one implementation, after obtaining the chatter identification result through online monitoring, the time-series chatter judgment result and the corresponding chatter occurrence probability are transmitted to the machine tool production control unit. When the chatter alarm triggering condition is met, the production control unit retrieves the pre-stored process adjustment strategy and intervenes in the cutting parameters of the lathe or milling machine tool, including reducing or adjusting the spindle speed, changing the feed rate, adjusting the cutting depth, or stopping the lathe machining program, so that the machining condition jumps out of the chatter excitation range and suppresses the further development of workpiece chatter. At the same time, the time of chatter occurrence, the corresponding sensor data, and the adjusted process parameters are recorded. If no chatter alarm is triggered, the current cutting process parameters are maintained and machining continues. After a single machining task is completed, the chatter event is linked back to the dimensional accuracy, surface quality, and tool wear detection results of the parts for subsequent process parameter optimization. This achieves closed-loop production control of the precision parts machining process and reduces the deterioration of part accuracy and abnormal tool wear caused by chatter.

[0027] In one embodiment, the process of acquiring multi-source data includes: A differential equation for the dynamics of precision components and cutting tools is constructed. The excitation conditions are obtained by assigning values ​​to the differential equation for the dynamics of precision components and cutting tools. The time-domain numerical iterative solution of the excitation conditions with variable step size is performed to obtain the time-series parameter response sequence. The time-series parameter response sequence includes: time-domain displacement response sequence and time-domain velocity response sequence. The first basic data is obtained by performing calculations on the time-domain displacement response sequence and the time-domain velocity response sequence. The second basic data is obtained by performing calculations on the time-domain displacement response sequence and the third basic data is obtained by performing calculations on the time-domain velocity response sequence. The first basic data, the second basic data, and the third basic data are fused to obtain multi-source data.

[0028] In one implementation, tool chatter is a major problem affecting workpiece dimensional accuracy and surface quality, and accelerating abnormal tool wear in precision turning. To achieve real-time data-driven chatter state determination, it is necessary to construct a multi-source sensor signal system mainly composed of acceleration, cutting force, and acoustic vibration as model training input. However, due to the highly specialized nature of this machining object, effective samples under abnormal working conditions are extremely scarce. Furthermore, general sensor data collected during conventional workpiece machining cannot be transferred and used due to significant differences in machining mechanisms, process windows, and dynamic characteristics.

[0029] This results in severe distribution gaps and class imbalances in the feature space of multi-source data, making it difficult for the model to establish a reliable mapping relationship from sensor signals to tremor states.

[0030] In one implementation, the tool-workpiece coupled dynamic characteristics of the precision lathe machining system are modeled using time-domain state-space equations based on regenerative chatter theory to obtain differential equations of precision component-tool dynamics that include structural modal mass, stiffness, damping, and cutting force coefficients.

[0031] The established differential equations of precision component-tool dynamics are processed by assigning values ​​to process parameters such as continuously varying spindle speed, depth of cut, feed rate, and tool overhang length to obtain deterministic system excitation conditions corresponding to different machining conditions.

[0032] For the excitation conditions of the deterministic system, a variable-step time-domain numerical iterative solution based on the Runge-Kutta method is performed to obtain the time-domain displacement response sequence and the time-domain velocity response sequence of the tool system at the cutting point.

[0033] The time-domain displacement response sequence and the time-domain velocity response sequence are substituted into the dynamic cutting force calculation formula and the average cutting force component is superimposed to obtain the first basic data characterizing the load fluctuation at the cutting interface, namely the multi-source time series signal of cutting force.

[0034] The second-order central difference derivative of the time-domain displacement response sequence is processed by multiplying it by the machine tool structure acceleration transmission coefficient to obtain the second basic data characterizing the inertial load of the tool or workpiece structure, namely the multi-source acceleration time series signal.

[0035] The time-domain velocity response sequence is processed by converting the near-field sound pressure level based on the air acoustic dipole radiation theory and superimposing the propagation distance attenuation factor to obtain the third basic data characterizing the air acoustic wave disturbance in the processing area, namely the acoustic vibration multi-source time series signal.

[0036] The cutting force time series signal, acceleration time series signal and acoustic vibration time series signal acquired simultaneously are processed by adding Gaussian white noise with a specific signal-to-noise ratio corresponding to the sensor type and calibrating it to the actual sensor range amplitude, so as to obtain a three-channel synchronous analog front-end output sequence with real sensor physical characteristics.

[0037] The output sequence of the three-channel synchronous analog front end is processed by anti-aliasing filtering and digital resampling at a uniform sampling frequency to obtain a multi-source sensor raw time series dataset (multi-source data) with consistent format and strict time alignment.

[0038] In one implementation, the differential equation for the dynamics of precision components and cutting tools is:

[0039] Where x(t) is the dynamic displacement in the x-feed direction and y(t) is the dynamic displacement in the y-depth of cut direction. Let x be the velocity in the feed direction. Let be the velocity in the y-direction of tangent. Let x be the acceleration in the x-feed direction. Let be the acceleration in the y-direction of the tangent. Equivalent modal mass in the x-direction (unit: kg). Equivalent modal mass in the y-direction (unit: kg). Let be the equivalent damping coefficient in the x-direction. Let be the equivalent damping coefficient in the y-direction. Let be the equivalent stiffness coefficient in the x-direction. is the equivalent stiffness coefficient in the y-direction. The dynamic cutting force acting on the tool in the x-direction is... The dynamic cutting force acting on the tool in the y-direction; Equivalent modal mass: refers to the inertial parameter that is equivalently converted to the vibration degrees of freedom according to a specific vibration mode, based on the complex distribution of mass of the actual lathe tool system. It characterizes the system's ability to resist changes in motion state during vibration.

[0040] Equivalent damping coefficient: refers to the linear damping parameter that is only related to speed, which is a comprehensive conversion of various complex energy-consuming factors such as internal material friction, connection surface loss and air resistance of machine tool structure. It characterizes the degree of vibration energy consumed by the system per unit vibration speed.

[0041] Equivalent stiffness coefficient: refers to the magnitude of the force required to produce a unit static displacement in a specific force direction of the tool system. It characterizes the overall stiffness of the system in resisting elastic deformation under force.

[0042] In one implementation, the Runge-Kutta method employs a variable-step-size time-domain numerical iterative solution: The complex second-order vibration differential equation is transformed into a state vector composed of displacement and velocity, allowing for recursive computation by the computer. Starting from the current state, within a preset time micro-element, the rate of change of the state is calculated sequentially at the four positions: the starting point, the two midpoints, and the ending point of the micro-element. These four rates of change are assigned different weights, and a weighted average is calculated as the standard rate of change within that time micro-element. This standard rate of change is then multiplied by the time micro-element and superimposed onto the current state, thereby accurately predicting the vibration state at the next moment. Simultaneously, this step employs a variable-step-size strategy to balance accuracy and efficiency. After each step, a lower-precision algorithm is used to estimate the computational error. If the error exceeds an acceptable threshold, the current time micro-element is automatically halved and the step is recalculated. If the error is significantly less than the threshold, the time micro-element is automatically doubled in the next step, thus ensuring both accuracy and saving computational resources.

[0043] In one implementation, the dynamic cutting force calculation formula and the average cutting force components are superimposed: The cutting force is divided into two parts: static average cutting force (generated by the nominal feed rate) and dynamic fluctuating cutting force (generated by the cutting thickness fluctuation caused by tool vibration). Dynamic cutting thickness expression:

[0044] in, The nominal cutting thickness (determined by the feed per revolution, unit: mm), T is the spindle rotation period (unit: s, N is the spindle speed (in rpm). This represents the displacement of the tool in this direction at the moment of the previous cutting action.

[0045] Formula for calculating dynamic cutting force:

[0046] Wherein, F(t) is the dynamic cutting force (unit: N, direction along the resultant cutting force direction). Specific cutting force coefficient (unit: N / mm) 2 , which is determined by the workpiece material and tool geometry and is calibrated through orthogonal cutting tests), b is the cutting width (unit: mm, i.e. the axial length of the tool edge participating in the cutting).

[0047] Superimposed average cutting force components: The actual force on the blade tip should be the sum of the dynamic fluctuation and the static average, that is:

[0048] in, The constant cutting force component, independent of vibration, is calculated using empirical formulas from the cutting handbook. The average tangential and radial forces are then combined. Substitute this into the right-hand side of the dynamic equation in Section 1.

[0049] In one implementation, the second-order central difference is differentiated and multiplied by the machine tool structure acceleration transmission coefficient: Displacement response sequence It cannot be directly output as a sensor acceleration signal; the acceleration sequence must be obtained through numerical differentiation, and the dynamic difference between the sensor installation position and the tool tip point must be compensated.

[0050] Second-order central difference derivative formula (for discrete sequences) Apply):

[0051] in, , and The first , , Displacement value at each sampling time, The sampling time interval (unit: seconds, determined by the preset sampling frequency) Decide, ), For the first Numerical acceleration at each sampling point (unit: mm / s²) 2 ).

[0052] In one implementation, near-field sound pressure level conversion and distance attenuation are based on the theory of airborne acoustic dipole radiation: The mechanical vibration velocity at the tool-workpiece coupling point is converted into an air sound pressure signal received by a microphone or acoustic sensor.

[0053] Near-field sound pressure calculation formula:

[0054] in, The distance is the straight-line distance from the center of the sound source to the microphone measurement point (unit: m). The resultant force of dynamic cutting force (unit: N). This indicates that the first-order central difference derivative is taken from its discrete sequence. The speed of sound in air (unit: m / s, take...) ), The angle between the dipole axis (i.e., the principal direction of the cutting force) and the orientation of the measuring point is given by the angle between the axis and the point. (In near-field free space,) To determine the radiation directivity, if the measuring points are located directly in front of the principal axis, then take... (to be simplified) : Geometric divergence attenuation term (sound pressure decreases inversely with distance).

[0055] In one embodiment, performing migration on multi-source data based on laboratory data to obtain multi-source migrated data includes: Physical rule hard filtering is performed on multi-source data to obtain a clean signal. Z-score normalization and sliding window slicing are performed on the clean signal to obtain unlabeled data. A preset model is obtained, and the preset model is trained on laboratory data to obtain an initial model. Unlabeled data is input into the initial model to perform forward inference to obtain multiple probability value sets. The mean and standard deviation are calculated based on each probability value set. After filtering each probability value set according to the mean and standard deviation and adding labels, the automatically labeled data is obtained. Based on the mean and standard deviation, each probability value set is filtered according to the second preset rule to obtain fuzzy samples. The fuzzy samples are manually labeled to obtain manually labeled data. The probability value set and the manually labeled data are then merged to obtain multi-source migration data.

[0056] In one implementation, raw multi-source industrial big data (multi-source data) is acquired, which includes 7 channels of sensor signals (triaxial acceleration, triaxial support force, and monophonic sound pressure), with a sampling rate of 10kHz and a data format of long-term continuous time-series signal; Perform operation 1 (physical rule hard filtering) on ​​multi-source data: Based on the physical boundaries of the processing (such as the acceleration amplitude usually not exceeding 50g, and the spindle frequency energy must be concentrated within the octave band), remove abnormal segments such as sensor disconnection, saturation clipping, and obvious jumps to obtain a clean signal (a clean but unlabeled long-time signal). Perform operation 2 (Z-score normalization and sliding window slicing) on ​​the clean signal: Divide the entire signal into numerous sample segments of 1000 sampling points (corresponding to 0.1 seconds) with a step size of 500 points (0.05 seconds), and perform Z-score normalization on each segment to align its amplitude magnitude with laboratory data, resulting in unlabeled data (dataset format: (No label).

[0057] In one implementation, a preset model is obtained, and an initial model is obtained by training the preset model on laboratory data: the ARENET pre-trained model trained on laboratory data is denoted as the initial model (this model has Bayesian hierarchical structure and has the ability to output uncertainty). Perform multiple random forward inferences on the unlabeled data: input each sample from the unlabeled data into the initial model. The initial model contains variational self-attention, and the number of samplings is set. Repeat the inference 10 times for the same sample, record the class probability (a continuous value between 0 and 1) of each output, and obtain a set of probability values ​​(with a set of 10 predicted probability values).

[0058] Calculate confidence and uncertainty indices for the probability value set: calculate the mean (Mean) of 10 predictions, which represents the model's bias; calculate the standard deviation (Std) of 10 predictions, which represents the model's cognitive uncertainty (the larger the standard deviation, the less reliable the model is for this sample). Add the following labels to the probability value set: Filtering rules: Keep those that meet the criteria. and The samples are classified as stable states (label 0); samples that satisfy the following conditions are retained. and The samples were classified as flutter states (label 1). These two types of samples and their assigned pseudo-labels are extracted to obtain automatically labeled data (high-confidence fully automatic labeled big data).

[0059] In one implementation, an operation is performed on the probability value set (difficult example mining and minimal-scale manual review): samples with the largest standard deviations are selected. ) or the mean is in the middle of a fuzzy zone ( From the sample, 500 to 1000 samples were randomly selected, and combined with the measured values ​​of the surface roughness of the processed surface or the on-site noise monitoring records corresponding to that time period, a very small-scale manual annotation was performed to obtain the manually annotated data; The process involves merging and balancing automatically labeled and manually labeled data: A massive amount of automatically labeled data is combined with a small amount of manually labeled data. To prevent pseudo-label noise from the automatically labeled data from contaminating the manually labeled data, sample weights are set during subsequent training (0.6 for automatically labeled data and 1.5 for manually labeled data). The merged data results in multi-source transfer data (a processed multi-source large-scale training set).

[0060] In one embodiment, inputting a signal segment into the channel attention mechanism module for processing to obtain a channel-enhanced signal includes: Perform global average pooling on the signal segment to obtain the global average pooling feature; Perform global max pooling on the signal segment to obtain the global max pooling feature; The global average pooling feature and the global max pooling feature are concatenated to obtain the combined pooling feature. The combined pooling features are input into a two-layer MLP (Multilayer Perceptron) network, and ReLU activation and Sigmoid activation operations are performed respectively to obtain the channel weight coefficients. The channel weighting coefficients are multiplied one by one with the corresponding signal segments to obtain the channel-enhanced signal.

[0061] In one implementation, a signal segment is used as the input signal R, and the dimension of the input signal R is R. C×1×D Where C is the number of sensor channels and D is the sampling point length of a single channel signal.

[0062] In one implementation, a global average pooling (GAP) operation is performed on the input signal R, and the average value of all sampling points in each channel is calculated to obtain the global average pooling feature. Perform global max pooling (GMP) operation on the input signal R, and calculate the maximum value of all sampling points in each channel to obtain the global max pooling feature. The global average pooling feature and the global max pooling feature are concatenated to obtain the combined pooling feature. The combined pooling features are input into a two-layer MLP multilayer perceptron network; the first MLP uses the ReLU activation function for nonlinear transformation, and the second MLP uses the Sigmoid activation function for mapping operation, outputting the channel weight coefficients with values ​​between 0 and 1. The channel weighting coefficients are multiplied by the original input signal R channel by channel to weight and calibrate the signal amplitude of each channel, thus obtaining the enhanced signal.

[0063] In one implementation, when the original multi-channel signal is directly input into the model, all channels are treated equally. The effective channels carrying strong flutter features are not enhanced, and redundant channels containing environmental noise and irrelevant interference information are not suppressed. This increases the burden of subsequent network feature extraction and reduces the accuracy of flutter recognition.

[0064] By mining the statistical features of each channel through global average pooling and global max pooling, and then learning the weight coefficients corresponding to each channel through a multilayer perceptron, the original multi-channel signals are recalibrated using the weights to amplify the signal representation of channels strongly correlated with flutter, reduce the influence of noisy and redundant channels, and achieve signal attention enhancement at the channel scale. Without increasing the amount of additional hardware overhead, this improves the ability of subsequent networks to extract effective features of flutter and reduces the negative impact of noise interference on classification results.

[0065] In one embodiment, inputting the channel enhancement signal into the BSR module to obtain multi-scale depth features includes: Sliding window block processing is performed on each channel of the channel enhancement signal to obtain block signals. The block signals are then input into the encoder to perform variational self-attention (VSA) operation to obtain temporal attention weights. The first temporal feature is obtained by performing feature weighting on the block signal according to the temporal attention weight. The first temporal feature is then passed through the Add&Norm residual normalization layer and the FeedForward network layer to complete feature transformation and obtain the second temporal feature. The second temporal feature is then subjected to repeated multi-layer BayesformerEncoder stacking operation to obtain the deep temporal feature. The deep temporal features are input into the SE channel attention module to obtain intermediate features. The channel enhancement signal is then added to the intermediate features to obtain multi-scale depth features.

[0066] In one implementation, the channel enhancement signal is acquired and used as the input signal of the BSR module.

[0067] In one implementation, a sliding window block processing is performed on each channel signal of the input channel enhancement signal to divide the long time-series signal into multiple signal sub-segments; a learnable position code is superimposed on each signal sub-segment to supplement the signal temporal position information, thereby obtaining a block signal carrying position information; The block signal carrying the position information is input into the BayesformerEncoder encoder; the variational self-attention (VSA) operation is performed inside the encoder to convert the attention weights from the deterministic numerical point estimate into a probability distribution in the form of a log-normal distribution, calculate the KL divergence between the prior distribution and the variational posterior distribution as a constraint term, and then reparameterize the attention weight distribution to obtain the temporal attention weights. The block signal is weighted using the obtained temporal attention weights, and then the feature transformation is completed by passing the Add&Norm residual normalization layer and the FeedForward network layer in sequence. The multi-layer BayesformerEncoder stacking operation is repeated to extract the local signal segments that are strongly correlated with flutter in the temporal dimension and output the deep temporal features corresponding to each channel. The depth temporal features of each channel output by BayesformerEncoder are input into the SE channel attention module; channel statistical features are obtained through global average pooling and global max pooling, and new channel weights are generated through two layers of MLP. These weights are then used to perform channel-by-channel weighting and calibration of the depth temporal features, thereby strengthening the flutter correlation features and suppressing noise features, resulting in recalibrated intermediate features. The ResNet residual connection operation (residual feature fusion submodule) is performed, which adds the original input of the BSR module (i.e., the channel enhancement signal) to the recalibrated intermediate features output by the SE module element by element. This alleviates the gradient vanishing and network degradation problems during deep network training, and obtains the multi-scale deep features processed by the BSR block.

[0068] In one implementation, traditional CNNs and ordinary Transformers have the following drawbacks when processing milling vibration time-series signals: ordinary convolutional networks are difficult to capture long-distance global temporal dependencies of signals; standard Transformers, which use deterministic attention weights, are prone to overfitting on small-sample laboratory datasets; and deep networks with multiple layers stacked are prone to gradient vanishing and network degradation. In addition, after multi-channel features are processed by channel attention, it is necessary to further filter out key vibration segments in the time dimension and filter out redundant channel information again. Temporal positional information is preserved by using a sliding window block overlay positional encoding. Variational self-attention is used to model attention weights as a probability distribution, and KL divergence constraints are introduced to improve the model's generalization ability with limited samples. This allows the model to focus on capturing local segments in the temporal signal that are strongly correlated with flutter. The SE module then performs a second round of channel-dimensional feature filtering, and ResNet residual connections are combined to ensure the trainability of the deep network. Finally, the output is a multi-scale deep feature that takes into account both channel and temporal scales, improving the ability to distinguish between stable cutting and flutter states and reducing recognition errors caused by noise interference.

[0069] In one embodiment, performing a classification network calculation on the dimensionality-reduced feature vector to obtain the cutting state includes: Perform Flatten operation on the dimensionality-reduced feature vector to obtain a one-dimensional flattened feature vector, and input the one-dimensional flattened feature vector into a fully connected FC layer to obtain an intermediate feature vector; The intermediate feature vector is input into the Softmax activation function for normalization to obtain a two-dimensional probability vector. The two-dimensional probability vector is then analyzed to obtain the probability of stable cutting state and the probability of chattering state. The maximum value between the probability of a stable cutting state and the probability of a chattering state is taken, and the category corresponding to the maximum value is taken as the final judgment result of the signal segment. The cutting state is output based on the final judgment result.

[0070] In one implementation, the Flatten operation is performed on the dimensionality-reduced feature vector to convert the multidimensional structured feature tensor into a one-dimensional feature vector, thus obtaining a one-dimensional flattened feature vector. The one-dimensional flattened feature vector is input into a fully connected FC layer. The nonlinear mapping transformation of the feature is completed through the weight matrix and bias parameters of the fully connected layer to obtain the mapped intermediate feature vector. The intermediate feature vector is input into the Softmax activation function for normalization. The elements in the vector are then normalized to make the sum of the output values ​​equal to 1, resulting in a two-dimensional probability vector. The two-dimensional probability vector is analyzed, and the two output dimensions of the vector correspond to the probability that the current signal segment belongs to a stable cutting state and the probability that it belongs to a fluttering state, respectively. Find the maximum probability of the two-dimensional probability vector, take the category corresponding to the maximum value as the final judgment result of the signal segment, and output the cutting state: stable cutting or chattering.

[0071] In one implementation, the multi-scale depth features obtained through channel attention and BSR block extraction cannot directly output interpretable cutting state judgment results. This is because depth features belong to high-dimensional abstract latent space data and cannot directly express the two physical processing states of stable cutting or chattering. Therefore, by Flattening, the multi-dimensional features are transformed into one-dimensional vectors suitable for processing by fully connected layers. These fully connected layers then map the high-level features to the task category space. Finally, Softmax activation converts the output into a probability distribution with a sum of 1. This not only directly outputs the final classification label but also simultaneously outputs the confidence probability of each state. The probability results can intuitively reflect the model's confidence in the current recognition result, thus achieving binary classification of milling signal segments for stability and chattering.

[0072] In one implementation, the maximum probability of the two-dimensional probability vector is obtained, and the category corresponding to the maximum value is taken as the final judgment result of the signal segment; when the two probabilities are equal, the segment is preferentially judged as chattering state, and the cutting state is output as chattering. In chatter detection, chatter is a hazardous condition. Missing it can damage the workpiece surface finish and cause abnormal tool wear. False alarms (mistaking stable cutting for chatter) only trigger an early warning and do not directly cause physical damage to the workpiece or tool. When the probabilities of the two states are equal, it indicates extremely low model confidence, and the sample is at the decision boundary between the two states. In this case, a conservative decision strategy biased towards the hazardous class is adopted, prioritizing the output of the chatter state to avoid missed chatter detection and reduce the risk of part scrapping and tool damage due to undetected chatter.

[0073] In one embodiment, obtaining the real-time cutting state by detecting the real-time sensing signal according to the detection model includes: Z-score normalization is performed on the real-time sensing signal to obtain a standard time-series signal. The standard time-series signal is then segmented using a sliding time window to obtain multiple real-time signal segments. Real-time signal segments are input into the detection model to obtain real-time judgment results. The real-time judgment results are aligned in chronological order to obtain continuous judgment results. If the number of chattering events in the continuous judgment results is greater than the alarm threshold, the real-time cutting state is determined to be chattering.

[0074] In one implementation, the online raw multi-channel real-time sensing signal is acquired in real time from the sensors at the milling site. The signal includes a triaxial acceleration signal, a triaxial cutting support force signal, and an acoustic signal. Zscore normalization processing is performed on the online raw multi-channel real-time sensing signals to eliminate the dimensional differences between different sensors and obtain standardized online time-series signals. The standardized online time-series signal was segmented using a sliding time window with a window length of 1000 sampling points and a window step of 0.05s, resulting in several online signal segments.

[0075] Read the network model parameters of the detection model, load the detection model, and input each online signal segment into the loaded detection model in sequence. Then, perform channel attention processing, BSR block feature extraction, pooling dimensionality reduction, and classification network operation in sequence. Output the stable cutting probability, chatter probability, and cutting state determination result corresponding to each online signal segment. According to the time sequence of the signal segments, the probability values ​​of each signal segment are time-series restored and aligned with the cutting state determination results to obtain the online chatter identification results that change continuously with the machining time, thus completing the online chatter state identification of the milling process.

[0076] In one implementation, the continuous identification results after time sequence restoration and alignment are statistically analyzed along the time axis using a sliding window to count the number of signal segments that are continuously identified as flutter.

[0077] Pre-configure the alarm threshold for continuous flutter segments; compare the statistically obtained number of continuous flutter segments with the alarm threshold.

[0078] If the number of consecutive chatter segments is greater than or equal to the alarm threshold, a chatter alarm signal is output (the real-time cutting state is determined to be chatter); if the number of consecutive chatter segments is less than the alarm threshold, no alarm signal is output, it is determined to be an instantaneous noise disturbance, and subsequent online signal segments are monitored.

[0079] In one implementation, the problem of inconsistent contributions of different sensor channels to the flutter detection task in multi-channel sensing signals is addressed. When the original multi-channel signal is directly input into the model, all channels are treated equally. Effective channels carrying strong flutter features are not enhanced, and redundant channels containing environmental noise and irrelevant interference information are not suppressed. This increases the burden of subsequent network feature extraction and reduces the accuracy of flutter recognition.

[0080] By mining the statistical features of each channel through global average pooling and global max pooling, and then learning the weight coefficients corresponding to each channel through a multilayer perceptron, the original multi-channel signals are recalibrated using the weights to amplify the signal representation of channels strongly correlated with flutter, reduce the influence of noisy and redundant channels, and achieve signal attention enhancement at the channel scale. Without increasing the amount of additional hardware overhead, this improves the ability of subsequent networks to extract effective features of flutter and reduces the negative impact of noise interference on classification results.

[0081] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should still fall within the scope of the claims of the present invention.

Claims

1. A method for controlling the production of precision parts based on big data analysis, characterized in that, The method includes: Acquire multi-source data and laboratory data, and perform migration on the multi-source data based on the laboratory data to obtain multi-source migrated data; the multi-source migrated data contains multiple signal segments. The signal segment is input into the channel attention mechanism module for processing to obtain the channel enhancement signal, and the channel enhancement signal is input into the BSR module to obtain multi-scale depth features; the BSR module consists of a BayesformerEncoder stacking submodule, an SE channel attention submodule, and a residual feature fusion submodule; Pooling operations are performed on multi-scale depth features to obtain dimensionality-reduced feature vectors. Classification network calculations are then performed on the dimensionality-reduced feature vectors to obtain the cutting state; the cutting state includes stability or chatter. If the cutting state result is correct, the detection model is deemed to have passed training, and the real-time sensing signal during workpiece cutting is acquired. The real-time cutting state is obtained by detecting the real-time sensing signal according to the detection model. If the real-time cutting state is chatter, lathe control is performed.

2. The precision parts production control method based on big data analysis according to claim 1, characterized in that, The process of acquiring the multi-source data includes: A differential equation for the dynamics of a precision component-tool is constructed. The equation is then assigned values ​​to obtain excitation conditions. These excitation conditions are then subjected to variable-step time-domain numerical iterative solution to obtain a time-series parameter response sequence. This time-series parameter response sequence includes a time-domain displacement response sequence and a time-domain velocity response sequence. The first basic data is obtained by performing calculations on the time-domain displacement response sequence and the time-domain velocity response sequence, the second basic data is obtained by performing calculations on the time-domain displacement response sequence, and the third basic data is obtained by performing calculations on the time-domain velocity response sequence. The first basic data, the first basic data and the first basic data are fused to obtain multi-source data.

3. The precision parts production control method based on big data analysis according to claim 1, characterized in that, Multi-source migrated data obtained by performing migration on the multi-source data based on the laboratory data includes: Physical rule hard filtering is performed on the multi-source data to obtain a clean signal. Z-score normalization and sliding window slicing are performed on the clean signal to obtain unlabeled data. A preset model is obtained, and the preset model is trained on the laboratory data to obtain an initial model. The unlabeled data is input into the initial model to perform forward inference to obtain multiple probability value sets. The mean and standard deviation are calculated based on each probability value set. After filtering each probability value set according to the mean and standard deviation and adding labels, the automatically labeled data is obtained. Based on the mean and standard deviation, each probability value set is filtered according to the second preset rule to obtain fuzzy samples. The fuzzy samples are manually labeled to obtain manually labeled data. The probability value set and the manually labeled data are fused to obtain multi-source migration data.

4. The precision parts production control method based on big data analysis according to claim 1, characterized in that, The signal segment is input into the channel attention mechanism module for processing to obtain the channel enhancement signal, including: Perform global average pooling on the signal segment to obtain the global average pooling feature; Perform global max pooling on the signal segment to obtain the global max pooling feature; The global average pooling feature and the global max pooling feature are concatenated to obtain the combined pooling feature. The combined pooling features are input into a two-layer MLP multilayer perceptron network, and ReLU activation and Sigmoid activation operations are performed respectively to obtain the channel weight coefficients. The channel weighting coefficients are multiplied by the corresponding signal segments one by one to obtain the channel-enhanced signal.

5. The precision parts production control method based on big data analysis according to claim 1, characterized in that, The channel enhancement signal is input into the BSR module to obtain multi-scale depth features, including: Sliding window block processing is performed on each channel of the channel enhancement signal to obtain block signals. The block signals are then input into the encoder to perform variational self-attention (VSA) operation to obtain temporal attention weights. The first temporal feature is obtained by performing feature weighting on the block signal according to the temporal attention weight. The first temporal feature is then passed through the Add&Norm residual normalization layer and the FeedForward network layer to complete feature transformation and obtain the second temporal feature. The second temporal feature is then subjected to repeated multi-layer BayesformerEncoder stacking operation to obtain the deep temporal feature. The deep temporal features are input into the SE channel attention module to obtain intermediate features. The channel enhancement signal is then added to the intermediate features to obtain multi-scale depth features.

6. The precision parts production control method based on big data analysis according to claim 1, characterized in that, The cutting state obtained by performing classification network calculations on the reduced-dimensional feature vector includes: Perform Flatten operation on the dimensionality-reduced feature vector to obtain a one-dimensional flattened feature vector, and input the one-dimensional flattened feature vector into a fully connected FC layer to obtain an intermediate feature vector; The intermediate feature vector is input into the Softmax activation function for normalization to obtain a two-dimensional probability vector. The two-dimensional probability vector is then analyzed to obtain the probability of stable cutting state and the probability of chattering state. The maximum value between the probability of a stable cutting state and the probability of a chattering state is taken, and the category corresponding to the maximum value is taken as the final judgment result of the signal segment. The cutting state is output based on the final judgment result.

7. The precision parts production control method based on big data analysis according to claim 1, characterized in that, The real-time cutting state is obtained by detecting the real-time sensing signal according to the detection model, including: The real-time sensing signal is subjected to Z-score normalization to obtain a standard time-series signal, and the standard time-series signal is divided into multiple real-time signal segments by a sliding time window. Real-time signal segments are input into the detection model to obtain real-time judgment results. The real-time judgment results are aligned in chronological order to obtain continuous judgment results. If the number of chattering events in the continuous judgment results is greater than the alarm threshold, the real-time cutting state is determined to be chattering.

8. A precision parts production control system based on big data analysis, characterized in that, The system includes: A data migration module is used to acquire multi-source data and laboratory data, and perform migration on the multi-source data based on the laboratory data to obtain multi-source migrated data; the multi-source migrated data contains multiple signal segments. The feature extraction module is used to input signal segments into the channel attention mechanism module for processing to obtain channel enhancement signals, and input the channel enhancement signals into the BSR module to obtain multi-scale depth features; the BSR module consists of a BayesformerEncoder stacking submodule, an SE channel attention submodule, and a residual feature fusion submodule; The result determination module is used to perform pooling operations on multi-scale depth features to obtain dimensionality-reduced feature vectors, and to perform classification network calculations on the dimensionality-reduced feature vectors to obtain the cutting state; the cutting state includes stability or chatter. The production control module is used to determine that the detection model training is qualified if the cutting state result is correct, and to acquire the real-time sensing signal during workpiece cutting. The real-time cutting state is obtained by detecting the real-time sensing signal according to the detection model. If the real-time cutting state is chatter, the lathe is controlled.

Images