A numerical control machine tool machining state evaluation method based on multi-source sensor data

Abstract

The application discloses a numerical control machine tool processing state evaluation method based on multi-source sensor data, and particularly relates to the fields of regulation and evaluation, and comprises multi-source heterogeneous data acquisition and preprocessing, acquisition of physical sensor signals, CNC internal parameters, in-situ measurement and process parameters, fusion, synchronization, purification and dynamic windowing processing; then feature deep fusion is carried out, modes are divided according to characteristics, exclusive coding, GRU time sequence modeling and cross attention interaction are carried out, and then core features are obtained through dimension reduction quantization; and uncertainty quantization evaluation is carried out, an evidence deep learning model is constructed, and state values and uncertainty of chatter and tool wear are output; the results are applied, multi-protocol cross-end transmission is carried out, state and confidence level grading decisions are made, and reliability is improved by combining closed-loop optimization and fault guarantee; the application solves the problems of multi-source data asynchronization, shallow fusion and evaluation without confidence, is suitable for multiple working conditions, improves evaluation precision, and supports intelligent control.

Description

Technical Field

[0001] This invention relates to the field of control and evaluation technology, and more specifically, to a method for evaluating the machining status of CNC machine tools based on multi-source sensor data. Background Technology

[0002] As the core equipment of high-end manufacturing, the monitoring and evaluation of the processing status of CNC machine tools is a key link to ensure workpiece accuracy, production efficiency and equipment safety, and a crucial support for the implementation of intelligent manufacturing systems.

[0003] Currently, the single-sensor monitoring mode has gradually developed towards multi-source data fusion. The industry generally collects multi-dimensional data to enrich the state characterization, including vibration signals of the spindle and slide, acoustic emission signals of the tool holder and spindle box, current signals of the motor circuit, temperature signals of bearings and cutting areas, and other physical sensor data. At the same time, it combines the operating parameters of the CNC system, such as spindle load rate and position tracking error, as well as in-situ measurement data and preset process parameters from laser interferometers and tool probes.

[0004] In terms of data processing and evaluation modeling, deep learning technology is being used more and more widely. It is often used to extract features and determine the state through models such as convolutional neural networks and recurrent neural networks. Moreover, the industry continues to pay more attention to and demand more from the evaluation methods in terms of real-time performance, adaptability to multiple working conditions, and reliability of results, which is driving the technology to evolve towards a more accurate and intelligent direction.

[0005] However, it still has some drawbacks in practical use, such as:

[0006] 1. Superficial multi-source data fusion: lack of standardized preprocessing system, failure to perform credibility-weighted fusion of data from the same source, reliance on software interpolation for cross-source data synchronization leading to large deviations, fragmented modal feature extraction, failure to establish effective correlations, and difficulty in forming fusion features that accurately represent the processing state;

[0007] 2. The evaluation results lack quantification of uncertainty: The existing model only outputs a single state prediction value, without quantifying cognitive uncertainty. It cannot reflect the model's confidence level in the results. When faced with fluctuations in multiple operating conditions, it is prone to misjudgment due to insufficient confidence, making it difficult to support risk-controlled decision-making.

[0008] 3. Insufficient targeting of data preprocessing: Differentiated purification schemes were not designed according to the characteristics of signal modes, resulting in poor noise filtering effect; the data window division was fixed and could not adapt to the dynamic characteristics of signals under different processing speeds, leading to loss of effective information or interference from redundant data.

[0009] 4. Disconnect between assessment and control and lack of iterative optimization: Assessment results are mostly used only for monitoring and are not linked to the CNC system to achieve adaptive control; there is a lack of closed-loop feedback mechanism, and model parameters and decision rules cannot be iteratively updated with changes in working conditions, resulting in poor adaptability to multiple working conditions. Summary of the Invention

[0010] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for evaluating the machining status of CNC machine tools based on multi-source sensor data, which solves the problems mentioned in the background art through the following scheme.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a method for evaluating the machining status of a CNC machine tool based on multi-source sensor data, comprising:

[0012] S1: Multi-source heterogeneous data acquisition and preprocessing: Acquire multi-source data including physical sensor signals of the target equipment, CNC internal parameters, in-situ measurement data and process parameters, perform data-level pre-fusion on the same source data, use hardware trigger timestamp as a reference, achieve synchronous processing of multi-source data through interpolation, perform sub-modal signal purification according to modal characteristics, and then perform dynamic window segmentation according to the processing speed.

[0013] S2: Feature deep fusion based on GRU cross-attention: The preprocessed multi-source data is divided into modes according to signal characteristics. Local features are extracted from each mode through mode-specific feature encoding. Long-term dependencies are captured through temporal context modeling. Multi-view fusion is achieved by cross-attention mode interaction. Core fusion features are obtained through fusion feature optimization.

[0014] S3: Uncertainty Quantification State Assessment: Construct an evidence deep learning model, input fused features and output inverse gamma distribution evidence parameters, calculate the processing state prediction value and cognitive uncertainty, and perform online assessment after model training and calibration;

[0015] S4: Application and output of evaluation results: Transmit cross-end data through multiple protocols using evaluation results, implement adaptive control decisions based on state values ​​and uncertainties, and dynamically update the model and rules through closed-loop verification and feedback optimization.

[0016] The technical effects and advantages of this invention are as follows:

[0017] 1. Achieve deep multimodal fusion: fused source data with credibility weighting and achieved high-precision synchronization with hardware-triggered timestamps; established cross-modal associations through modal coding, GRU temporal modeling and cross-attention interaction to generate high-quality fusion features and make up for the defects of shallow fusion.

[0018] 2. Quantitative assessment of uncertainty: Construct an evidence deep learning model, output inverse gamma distribution parameters to calculate state values ​​and variance, and classify high, medium and low confidence levels according to variance to provide a confidence basis for decision-making and solve the problem of misjudgment caused by the lack of uncertainty reference;

[0019] 3. Refined data preprocessing: Dedicated purification algorithms are designed for vibration, acoustic and electrical modes, and window parameters are dynamically adjusted in combination with processing speed; through data-level pre-fusion and cross-source alignment, noise is effectively filtered while retaining key information, improving the problem of insufficient preprocessing targeting;

[0020] 4. Construct a closed loop for evaluation, control, and optimization: Transmit the results to the CNC system via multiple protocols to achieve hierarchical decision-making. Optimize the model and rules by verifying the collected data through closed-loop verification, and combine this with fault emergency strategies to achieve linkage and dynamic iteration between evaluation and control, thereby improving adaptability to multiple working conditions. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0022] Figure 2 This is a schematic diagram of the S1 process of the present invention.

[0023] Figure 3 This is a schematic diagram of the S2 process of the present invention.

[0024] Figure 4 This is a schematic diagram of the S4 adaptive control process of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] refer to Figures 1-4 The method for evaluating the machining status of CNC machine tools based on multi-source sensor data, as shown, includes:

[0027] S1: Multi-source heterogeneous data acquisition and preprocessing

[0028] It achieves comprehensive coverage of multi-source data, accurate synchronization and alignment, and reliable signal quality. Through standardized acquisition and refined preprocessing, it provides high-quality time-series data for subsequent feature fusion, solving the core problems of asynchronous, noisy, and complex multi-source data in industrial scenarios.

[0029] S101: Implementation of Multi-Source Data Acquisition

[0030] A1: Physical sensor signal acquisition:

[0031] Vibration signal: One vibration sensor is installed on each of the front bearing housing of the spindle (one in the horizontal direction and one in the vertical direction), the X-axis slide, and the Z-axis slide. The sampling rate is set to 10kHz and the resolution to 16-bit. A time-domain waveform containing 100 data points is generated every 10ms. The acquisition is triggered by binding the CNC G01 and G02 cutting commands and is paused after the command ends.

[0032] Acoustic emission signal: One acoustic emission sensor is installed at the end of the tool holder and one on the side of the spindle box. The sampling rate is set to 2MHz to match the 150kHz resonant frequency. The preamplifier gain is adjusted to 40dB, and the output pulse signal has a rise time of less than or equal to 10μs. The hardware has a built-in 100-200kHz bandpass filter preprocessing.

[0033] Current signal: One closed-loop Hall sensor is installed on each of the power input lines of the spindle motor and the X, Y, and Z axis feed motors. The sampling rate is 50kHz, the range is 0-100A, and the output is a 4-20mA analog-to-digital signal. The sensor housing is grounded and the wires are shielded twisted-pair cables, kept away from high-voltage cables to prevent interference.

[0034] Temperature signal: Temperature sensors are attached to the outer rings of the front and rear bearings of the spindle using thermally conductive adhesive. Non-contact infrared probes are used in the cutting area at a distance of 10cm. Sensors are pre-embedded in the spindle motor windings. PT1000 platinum resistance thermometers are selected with a measurement accuracy of ±0.1℃ and a sampling rate of 10Hz to collect the temperature of the front and rear bearings and the spindle motor in the cutting area.

[0035] A2: CNC Internal Parameter Acquisition:

[0036] Establish a TCP connection via FOCAS API V3 (IP is the machine tool IP, port matches the preset port), and read parameters at 10ms intervals: address corresponding to spindle load rate, X, Y, and Z axis position tracking error, actual feed rate, actual spindle speed, and allocate a 1MB circular buffer locally to prevent data overflow.

[0037] A3: In-situ Measurement and Process Parameter Acquisition

[0038] In-situ measurement: The laser interferometer outputs one flatness error data every 50ms with an accuracy of ±0.001mm; the tool probe triggers measurement during the machining gap and outputs one tool length deviation value every 1s.

[0039] Process parameters: Before machining, the preset feed rate f0, spindle speed n0, and depth of cut ap are read from the CNC process library and stored in the system as static parameters.

[0040] S102: Pre-processing implementation:

[0041] B1: Data-level pre-fusion, employing credibility-weighted fusion for multi-sensor data of the same type:

[0042] First, the signal deviation rate of each sensor is calculated based on historical data: ,in These are the actual measured values ​​from each sensor. The standard value is used; then the credibility weight is calculated: Final fusion value: The weights are recalibrated every hour. It is then marked as low confidence and its weight is reduced to 0.05.

[0043] B2: Multi-source data synchronization processing

[0044] Synchronization reference establishment: The acquisition card is connected to the CNC's "cutting start" digital signal (DI0) as a global trigger source. The rising edge generates a timestamp with an accuracy of ±0.1μs. The physical sensor signals are directly stamped by the acquisition card. The CNC parameters and in-situ measurement data are accompanied by their respective device timestamps.

[0045] Cross-source alignment: Using the data acquisition card timestamp as the reference axis, CNC parameters are aligned using linear interpolation. ,in For CNC data timestamps, , For adjacent base timestamps, , The parameter values ​​correspond to the reference time; 100 sets of data were collected for verification, and the alignment error did not exceed 0.5ms.

[0046] B3: Modal signal cleanup

[0047] Vibration signal: Butterworth 4th order bandpass filter was used to remove 50Hz power frequency interference, and the signal-to-noise ratio after processing was greater than 25dB;

[0048] Acoustic emission signal: First, it is filtered by a 100-200kHz bandpass filter, and then adaptive threshold noise reduction is used. The threshold is set to 0.4 times the maximum value of the signal in the first 5 seconds and is updated every 1 second. At the same time, the waveform is shaped to ensure that the rise time is less than 10μs and the ringing count error is less than 5%.

[0049] Current signal: High-frequency interference is filtered out by a Butterworth 4th order low-pass filter (cutoff frequency 1kHz), and then kurtosis analysis is performed. Abnormal fluctuations are marked when the kurtosis is greater than 3.5. The harmonic amplitude extraction error is less than 2%.

[0050] Temperature signal: Smoothed using Kalman filtering, process noise is 0.01, measurement noise is 0.1, and fluctuation amplitude is controlled within 0.3℃;

[0051] CNC parameters: If three consecutive cycles are missing, an anomaly is marked, and forward padding is used to complete the data, ensuring data integrity is greater than 99.9%.

[0052] B4: Dynamic Window Segmentation

[0053] Read the actual CNC feed rate v. If v ≤ 1000 mm / min, the window duration is 1 second with an overlap rate of 50%. Each window contains 100 data points at 10 ms each, with 50 overlapping data points. If v > 1000 mm / min, the window is shortened to 0.5 seconds with an overlap rate of 70%. Each window contains 50 data points with 35 overlapping data points. If the percentage of missing data types in a single window is greater than 10%, it is marked as invalid and filled by the average of adjacent windows.

[0054] S2: Feature Deep Fusion Based on GRU Cross-Attention

[0055] S201: Multi-source data modality partitioning:

[0056] The structured preprocessing is classified according to its characteristics. The classification is based on the following criteria: the data after S1 preprocessing is divided into four modes: vibration, acoustic-electric, CNC and measurement, with signal dynamics (high frequency, low frequency), data form (waveform, value) and correlation with the processing state as the core criteria.

[0057] Modal definition:

[0058] Vibration modes: Only the fused vibration time-domain signal is included. The input is 1×10000 dimensions in a 1s window and 1×5000 dimensions in a 0.5s window. The features are mainly high-frequency fluctuations with strong time-series dependence.

[0059] Acousto-electric mode: Integrating acoustic emission pulses and current waveforms, the 1s window input is 2×10000 dimensions, and the 0.5s window is 2×5000 dimensions, which belongs to the hybrid characteristics of pulse and continuous waveforms;

[0060] CNC modality: encompasses parameters such as spindle load rate and feed rate. The 1s window input is 5×100 dimensions, and the 0.5s window is 5×50 dimensions. It features low-frequency stability and strong correlation with control state.

[0061] Measurement modes include flatness error, tool deviation and process parameters. The input is 5×10 dimensions in a 1s window and 5×5 dimensions in a 0.5s window, mainly focusing on static or quasi-static features.

[0062] Data integrity verification: Check the data missing rate modally. If the missing rate of a single modality is greater than 5%, use the average of adjacent windows in S1 to fill the missing data to ensure that the data within the modality is continuous and available.

[0063] S202: Modality-Specific Feature Encoding:

[0064] Vibration mode coding (3-layer 1D-CNN):

[0065] The input layer receives vibration signals. The first layer uses a convolutional kernel with a kernel size of 5×1, 32 output channels, and a stride of 2, activated by LeakyReLU, and then outputs a 32×2499 dimension after max pooling.

[0066] The second layer uses a convolutional kernel with a kernel size of 3×1, 64 output channels, and a stride of 1, with the same activation function. After pooling, the output is 64×1249 dimensions.

[0067] The third layer uses a convolutional kernel with a kernel size of 3×1, 128 output channels, and a stride of 1. It then uses global average pooling to compress the temporal dimension and outputs a 128-dimensional local feature vector. .

[0068] Acoustic-electric modal coding (2-layer 1D-CNN):

[0069] The input is a mixed acoustic and electrical signal. The first layer uses a convolution kernel with a kernel size of 5×1, an output channel size of 32, and a stride of 2. It is activated by LeakyReLU and then pooled to output a 32×2499 dimension.

[0070] The second layer uses a convolutional kernel with a kernel size of 3×1, 64 output channels, and a stride of 1. After pooling, it undergoes global average pooling to output a 64-dimensional feature vector. .

[0071] CNC modality coding (2-layer 1D-CNN):

[0072] The input CNC parameter sequence is used. The first layer uses a convolutional kernel with a kernel size of 3×1, 32 output channels, and a stride of 1. It is activated by LeakyReLU and then pooled to output a 32×49 dimension.

[0073] The second layer uses a convolutional kernel with a kernel size of 3×1, 64 output channels, and a stride of 1. After pooling, it undergoes global average pooling to output a 64-dimensional feature vector. .

[0074] Measurement mode coding (2 fully connected layers):

[0075] Input the flattened measurement parameters (50 or 25 dimensions), set 32 ​​nodes for the first layer, and activate LeakyReLU;

[0076] The second layer has 32 nodes, uses the same activation function, and outputs a 32-dimensional feature vector. .

[0077] Feature summary: The four modal feature vectors are concatenated, with a total dimension of 128+64+64+32=288, which are used as input for time series modeling.

[0078] S203: Temporal context modeling, long-term dependency capture in GRU networks:

[0079] Modal-specific GRU parameter configuration:

[0080] Vibrational modes: 2-layer GRU, 128 hidden nodes per layer, dropout rate 0.2, input sequence length 100 (1s window) or 50 (0.5s window);

[0081] Acousto-electric mode: 2-layer GRU, 64 hidden nodes per layer, dropout rate 0.2, input sequence length is the same as vibration mode;

[0082] CNC modality: 1-layer GRU, 64 hidden layer nodes, dropout rate 0.1, input sequence length as before;

[0083] Measurement modality: 1-layer GRU, 32 hidden layer nodes, dropout rate 0.1, input sequence length 10 (1s window) or 5 (0.5s window).

[0084] GRU initialization: The weights are initialized using a Xavier normal distribution, and the bias term is set to 0 to ensure gradient stability in the early stages of training.

[0085] Temporal Context Vector Output: The hidden state of the last time step of the GRU is taken as the global temporal vector.

[0086] Vibration modes: 128 dimensions, characterizing the temporal evolution features of vibration signals;

[0087] Acousto-electric modes: 64-dimensional, representing the temporal characteristics of acoustic-electric correlation;

[0088] CNC mode: 64 dimensions, representing the gradual change trend of processing parameters;

[0089] Measurement modes: 32-dimensional, representing the measurement benchmark and temporal changes.

[0090] S204: Cross-attention modal interaction, multi-view active feature fusion

[0091] Cross-attention core definition: Employing a scaled dot product attention mechanism, the core formula is:

[0092]

[0093] Where Q is the core modality context vector, and K and V are other modality vectors. For the dimension of Q, Used to avoid gradient vanishing.

[0094] Multi-view fusion computing:

[0095] Vibration perspective :Will , , The data is concatenated into a 160-dimensional K / V array, then linearly transformed to 128 dimensions; the score is calculated. Weight Generate fusion vector 128 dimensions;

[0096] Sound TV Corner :Will , , The concatenation is a 224-dimensional K / V array, which is then linearly transformed to 64 dimensions; similarly, this process is repeated to generate... ,64 dimensions.

[0097] CNC perspective :Will , , The concatenation is a 224-dimensional K / V array, which is then linearly transformed to 64 dimensions; similarly, this process is repeated to generate... ,64 dimensions.

[0098] Measurement perspective :Will , , The concatenation is a 256-dimensional K / V array, which is then linearly transformed to 32 dimensions; similarly, this process is repeated to generate... 32 dimensions.

[0099] Viewpoint weight calibration: Accuracy of each viewpoint was calculated using a validation set of 1200 sets.

[0100] Vibration perspective: , Audio-visual corner: CNC perspective: Measurement perspective: Then follow the formula: Calculate the weights for each perspective: , It is an ordinal number.

[0101] Final Fusion: , , Zero-padding to 128 dimensions, and generating 128-dimensional features according to the formula: .

[0102] S205: Feature Optimization and Engineering-Adaptive Dimensionality Reduction and Quantization

[0103] Redundant feature removal: calculation The Pearson correlation coefficient of internal features is calculated using the following formula: , Features with small variance are removed, and 64-dimensional features are retained.

[0104] PCA dimensionality reduction: Principal component analysis was performed on the 64-dimensional features. Principal components were selected based on their cumulative variance contribution rate exceeding 95%. The dimension is reduced to 48, where W is a 48×64 dimension load matrix;

[0105] Feature normalization: Min-Max normalization is used to map to the interval [0, 1]. The specific mathematical formula is as follows: ,in , These are the extreme values ​​of the features in the training set.

[0106] Model quantization: TensorFlow Lite is used to quantize float32 features into int8, using the following formula: Where S is the scaling factor and Z is the zero offset, the inference latency is compressed from 50ms to less than 15ms;

[0107] S3: State Assessment Based on Uncertainty Quantification

[0108] S301: Model Input / Output and Architecture Design, Basic Framework Construction for Uncertainty Quantification:

[0109] Input / output definitions:

[0110] Input: S2-optimized 48-dimensional core fusion feature vector It has been normalized to the [0, 1] interval, covering the correlation information of vibration, acoustics, CNC, and measurement multimodal;

[0111] Output: For each of the two status indicators, chatter and tool wear, output four evidence parameters. There are a total of 8 parameters, which correspond to the position, degrees of freedom, shape, and scale parameters of the inverse gamma distribution, and are used to calculate the state prediction value and uncertainty.

[0112] Network architecture design:

[0113] Input layer: Receives 48-dimensional feature vectors, whose dimensions directly match the S2 output, requiring no additional dimension transformation;

[0114] Hidden layers: There are 3 fully connected layers, with 128 nodes in the first layer, 64 nodes in the second layer, and 32 nodes in the third layer. All layers use the ReLU activation function, and a batch normalization layer is added after each layer.

[0115] Output layer: 8 nodes, no activation function, directly outputs the evidence parameter vector.

[0116] S302: Loss Function Design, Core Constraint Construction in Evidence-Based Deep Learning:

[0117] Loss function decomposition:

[0118] Total loss function: , where λ is the balance coefficient, which was set to 0.5 after validation set testing to balance uncertainty modeling and prediction accuracy.

[0119] Calculation of evidence loss:

[0120]

[0121] in, The likelihood log probability of the predicted value versus the true value reflects how well the model fits the known samples.

[0122] , For gamma function, is the Dirichlet distribution parameter, corresponding to three states: normal, warning, and fault. K=3 is the number of state categories. For strength of evidence, KL divergence measures the difference between the model's current distribution and the uniform distribution, avoiding model overconfidence.

[0123] Loss calculation under supervision:

[0124] Using mean squared error, the formula is: ,in , where is the position parameter output by the model, i.e., the state prediction value, and y is the true label, which directly constrains the deviation between the predicted value and the true state;

[0125] Loss function implementation: The computation logic is written using the TensorFlow custom loss function interface to ensure... , Constraints;

[0126] S303: Training dataset construction and preprocessing, ensuring data source for model generalization:

[0127] Dataset size and composition:

[0128] Total sample size: 6000 groups, covering 3 states: 4000 normal groups, 1000 warning groups, and 1000 fault groups, with a sample ratio of 4:1:1;

[0129] Operating conditions covered: Includes 9 typical operating conditions with three materials: 45# steel, aluminum alloy, and titanium alloy, spindle speed of 2000-4000rpm, feed rate of 0.1~0.3mm / r, and depth of cut of 1~3mm, ensuring sample diversity.

[0130] Labeling method:

[0131] Flutter label : 0 = no chatter (peak vibration < 0.5g), 0.5 = slight chatter (0.5g ≤ peak value < 1.0g), 1 = severe chatter (peak value ≥ 1.0g), based on the joint labeling of vibration sensor peak value and machined surface roughness;

[0132] Tool wear label : 0 = new tool (wear amount < 0.1mm), 0.5 = intermediate wear (0.1mm ≤ wear amount < 0.2mm), 1 = scrap wear (wear amount ≥ 0.2mm). The wear amount on the back face is marked by measuring it with a tool microscope.

[0133] Uncertainty auxiliary label: label confidence level: high, medium, low, corresponding to σ² thresholds of less than 0.1, [0.1, 0.3], and greater than 0.3, respectively, for subsequent model calibration.

[0134] Dataset preprocessing:

[0135] The data was divided into three sets: a training set of 4200 groups, a validation set of 1200 groups, and a test set of 600 groups, in a ratio of 7:2:1, to ensure that the proportion of samples for each working condition was consistent across the three sets.

[0136] Feature normalization: using statistics from the training set , Perform Min-Max normalization on the entire dataset to maintain consistency with the S2 optimization process;

[0137] Sample augmentation: Time axis shifting was used to augment faulty samples by ±5 data points, expanding the sample size to 1500 groups to alleviate the sample imbalance problem;

[0138] S304: Model training and calibration, accuracy optimization for uncertainty quantification:

[0139] Training parameter configuration:

[0140] Optimizer: The Adam optimizer is used, with an initial learning rate of 0.001, which decays by 10% every 50 epochs; the batch size is 32, and the training epochs are 200.

[0141] Early Stop Mechanism: Monitoring and Verifying Total Loss If there is no decrease after 15 consecutive rounds, training is stopped, the optimal model is saved, overfitting is avoided, and the accuracy on the validation set is improved by more than 5%.

[0142] Training process monitoring:

[0143] Key performance indicator (KPI) tracking: Record the training and validation sets every 10 rounds. , Flutter prediction accuracy, wear prediction accuracy;

[0144] Convergence criteria: The training is considered successful if the accuracy of flutter and wear prediction on the validation set is greater than 96% and the uncertainty assessment matches the label confidence level with a rate greater than 90%.

[0145] Model calibration implementation:

[0146] Evidence parameter calibration: Adjust the initial value of β on the test set (iterate from 0.1 to 0.5) to achieve the normal state. , Where zzz represents chatter and mmm represents wear;

[0147] Threshold fine-tuning: Based on the distribution of fault samples in the test set, fine-tune the uncertainty classification thresholds (high confidence σ² < 0.1, medium confidence 0.1 ≤ σ² < 0.3) to make the false positive rate of fault identification less than 2% and the false negative rate ≤ 1%.

[0148] Model saving and export:

[0149] Save format: Save the trained model weights in HDF5 format, and store... , Calibration parameters such as initial value of β;

[0150] Quantization Export: The model is quantized to int8 precision using TensorFlow Lite, resulting in a compressed file size of less than 200MB and an inference time of less than 10ms.

[0151] S305: Online evaluation process, real-time status and uncertainty output:

[0152] Real-time feature input:

[0153] Receive the 48-dimensional normalized fusion features output in real time from S2 and transmit them with low latency via shared memory;

[0154] Input validation: Check if the feature value is in the range [0, 1]. If it is outside the range, mark the feature as abnormal and use the valid feature from the previous period as a substitute.

[0155] Model inference calculation:

[0156] The quantized evidence deep learning model is invoked, inputting F_norm for inference, and outputting 8 evidence parameters: , , , , , , , Inference time is less than 10ms;

[0157] State value and uncertainty calculation:

[0158] State prediction value: directly take the γ parameter. , ;

[0159] Cognitive uncertainty has the following specific function:

[0160] ,

[0161] The larger the variance, the lower the reliability of the model's prediction results.

[0162] Confidence level grading and result output:

[0163] Grading standards: For high confidence, For medium confidence level, Low confidence level;

[0164] Output format: Output in JSON format, including timestamp, status value, uncertainty, confidence level and calculation basis;

[0165] Output frequency: consistent with the S2 feature output frequency, 0.5s / 1s, to ensure synchronization with the processing rhythm;

[0166] S4: Application and Output of Evaluation Results

[0167] S401: Evaluation results are published via multiple protocols, ensuring reliable cross-platform data transmission.

[0168] Communication protocol selection and configuration:

[0169] The MQTT protocol is designed for workshop cloud platforms and monitoring terminals. The service port, client connection address, and client ID are all preset based on the target device, and the QoS level is set to 2 to ensure that messages are delivered only once, avoiding duplication or loss.

[0170] OPC UA protocol, for CNC systems: reuses the OPC UA server of the CNC system in S1, with preset port, and creates a set of status evaluation result nodes in the server, with node paths corresponding to data types;

[0171] Publication content and frequency:

[0172] MQTT publish: The published content is the complete result of the S3 output, including timestamp, status value, uncertainty, and calculation basis. The publishing frequency is consistent with the evaluation cycle, 0.5s / 1s, and is synchronized with the S3 output.

[0173] OPC UA push: It adopts a change-triggered and periodic fallback mode. When the evaluation result is updated, it is pushed immediately with a delay of less than 50ms. If there is no update within 500ms, the current value is pushed periodically to ensure that the CNC system obtains the latest status in real time.

[0174] Transmission reliability guarantee:

[0175] Local caching: The computing unit allocates a 512KB circular buffer locally to cache the 100 most recent evaluation results. If the network is interrupted, the data that was not successfully published will be automatically resent after the network is restored.

[0176] Transmission verification: The MQTT client listens for the on_publish callback. If no acknowledgment is received, it retryes up to 3 times, with an interval of 100ms. The OPC UA reads the server node value within 100ms after push. If the deviation from the pushed value is greater than 0.01, it re-pushes.

[0177] S402: Adaptive control decision-making, precise adjustment driven by uncertainty:

[0178] Deployment of control decision-making agents:

[0179] An adaptive control agent is deployed on the CNC system side or in the computing unit. The agent subscribes to the evaluation result node through the OPC UA client. The decision logic is triggered immediately upon receiving a new result, and the decision response time is less than 20ms.

[0180] Hierarchical decision rule base design:

[0181] The rule base is structured around state conditions, actions, and parameter calculations, and the following rules are designed for key states:

[0182] Flutter state decision:

[0183] High confidence anomaly, and :

[0184] Actions: Reduce feed rate, slow down spindle speed, and trigger audible and visual alarms;

[0185] The adjustment amounts are shown below. For feed rate adjustment, This refers to the spindle speed reduction adjustment amount, and the labeling is the same throughout the text.

[0186] , ;

[0187] Execution priority: Level 1, highest, interrupts non-urgent processing tasks.

[0188] Medium confidence level abnormality and :

[0189] Actions: Fine-tune the feed rate and continuously monitor chatter status;

[0190] Parameter calculation: Meanwhile, the feed smoothing coefficient is corrected based on the CNC position tracking error, and the smoothing coefficient is adjusted to 0.8 when the error is >0.005mm;

[0191] Execution priority: Level 2, executed intermittently during normal processing.

[0192] Low confidence anomaly and :

[0193] Actions: Issue early warnings, enhance monitoring and data recording;

[0194] Specific operation: Trigger the acoustic emission sensor sampling rate to increase to 4MHz, save the vibration and current data of the most recent 10 windows, push early warning information to the operation and maintenance terminal, and do not perform parameter adjustment;

[0195] Execution priority: Level 3, only provides warnings, does not interfere with processing.

[0196] Tool wear condition decision:

[0197] High confidence anomaly, and :

[0198] Actions: Immediate emergency stop, tool change prompt, and marking of wear location;

[0199] Parameter calculation: Based on the acoustic and electrical modal characteristics, the wear area is located, and the appropriate tool model is pushed in combination with the process parameters;

[0200] Execution priority: Level 1, forcibly interrupts processing to avoid scrapping the workpiece.

[0201] Medium confidence anomaly and :

[0202] Actions: Tool length compensation, feed rate reduction, and in-situ measurement verification;

[0203] Parameter calculation: , After compensation, the laser interferometer is triggered to remeasure the flatness; For tool length compensation;

[0204] Execution priority: Level 2, compensation is performed during processing gaps.

[0205] Low confidence anomaly and :

[0206] Actions: Early warning notification, manual inspection, and data archiving;

[0207] Specific operation: The CNC system pops up a tool wear warning (low confidence), along with the dimensional error and temperature change curves of the last 5 windows, marks the period of CNC load rate fluctuation, and prompts the operator to stop the machine to check the tool;

[0208] Execution priority: Level 3, prompt only, relies on human decision-making;

[0209] S403: Closed-loop verification and feedback optimization, dynamic calibration of control performance:

[0210] Command execution effect verification:

[0211] Parameter consistency verification: 100ms after the command is issued, the actual parameters of the CNC system are read via OPC UA, and the relative deviation from the command value is calculated. ;

[0212] like Verification passed; the recorded instruction executed normally.

[0213] like Trigger a retry, up to 2 times, with an interval of 200ms. If the limit is still exceeded, mark the parameter drift warning. Combine the CNC load rate to determine if there is mechanical jamming. If the load rate is greater than 90%, indicate that the transmission mechanism is abnormal.

[0214] like Immediately suspend control actions, initiate joint vibration and current detection. If the vibration amplitude is greater than 1g or the current surge is greater than 50%, the sensor is determined to be faulty, and fault location information is pushed to the maintenance terminal.

[0215] Machining accuracy verification: After the compensation command is executed, the laser interferometer is triggered to measure the flatness of the workpiece. If the error is less than 0.003mm, the compensation is deemed effective; otherwise, the compensation is marked as ineffective and the compensation amount needs to be recalculated.

[0216] Feedback data collection and modeling:

[0217] Data Acquisition: For each control command executed, the entire data chain, including evaluation results, control commands, actual parameters, and accuracy results, is collected and stored in a historical database according to timestamp, status, command, and effect formats. Approximately 10,000 new records are added daily, and the database is retained for 3 months.

[0218] Effectiveness evaluation modeling: Historical data is statistically analyzed weekly to calculate the effectiveness of each decision rule under different operating conditions.

[0219] Model and rule base optimization:

[0220] Monthly short-term optimization: Fine-tune the parameter calculation coefficients for rules with an effectiveness rate of less than 90%;

[0221] Quarterly long-term optimization: Collect more than 200 sets of new fault data, retrain the S2 cross-attention model and the S3 evidence deep learning model, update the modality contribution weight and uncertainty threshold, improve evaluation accuracy, and indirectly optimize decision-making performance;

[0222] Optimization and verification: After each optimization, 100 sets of working conditions are verified on the test bench to ensure that the fault identification accuracy is greater than 97% and the control effectiveness is greater than 92% before it can be deployed to the production environment.

[0223] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0224] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the machining status of CNC machine tools based on multi-source sensor data, characterized in that, include: S1: Multi-source heterogeneous data acquisition and preprocessing: Acquire multi-source data including physical sensor signals of the target equipment, CNC internal parameters, in-situ measurement data and process parameters, perform data-level pre-fusion on the same source data, use hardware trigger timestamp as a reference, achieve synchronous processing of multi-source data through interpolation, perform sub-modal signal purification according to modal characteristics, and then perform dynamic window segmentation according to the processing speed. S2: Feature deep fusion based on GRU cross-attention: The preprocessed multi-source data is divided into modes according to signal characteristics. Local features are extracted from each mode through mode-specific feature encoding. Long-term dependencies are captured through temporal context modeling. Multi-view fusion is achieved by cross-attention mode interaction. Core fusion features are obtained through fusion feature optimization. The modality-specific feature encoding includes: Design dedicated extraction networks for four modalities: Vibration modes are encoded using a 3-layer 1D-CNN. Features are extracted layer by layer using convolutional kernels of different sizes and the LeakyReLU activation function. After max pooling and global average pooling, a 128-dimensional local feature vector is output. The acoustic-electric modal is encoded using a 2-layer 1D-CNN, and mixed features are extracted by convolution, activation and pooling operations. A 64-dimensional feature vector is output after global average pooling. The CNC modality uses a 2-layer 1D-CNN encoding to meet the requirements of low-frequency parameter feature extraction. After pooling and global average pooling, a 64-dimensional feature vector is output. The measurement mode is encoded using a 2-layer fully connected layer. The flattened parameters are then transformed to output a 32-dimensional feature vector. After encoding, the feature vectors of the four modalities are concatenated to form a 288-dimensional feature, which is used as the input for time series modeling. S3: Uncertainty Quantification State Assessment: Construct an evidence deep learning model, input fused features and output inverse gamma distribution evidence parameters, calculate the processing state prediction value and cognitive uncertainty, and perform online assessment after model training and calibration; S4: Application and output of evaluation results: Transmit cross-end data through multiple protocols using evaluation results, implement adaptive control decisions based on state values ​​and uncertainties, and dynamically update the model and rules through closed-loop verification and feedback optimization.

2. The method for evaluating the machining status of a CNC machine tool based on multi-source sensor data according to claim 1, characterized in that: The multi-source data includes: Physical sensor signals, including vibration signals, acoustic emission signals, spindle and feed motor current signals, and spindle bearing and cutting zone and motor winding temperature signals; CNC internal parameters include spindle load rate, X, Y, and Z axis position tracking error, actual feed rate, and actual spindle speed; In-situ measurement data, including workpiece flatness error acquired by laser interferometer and tool length deviation acquired by tool probe; Machining process parameters, including static parameters such as preset feed rate, spindle speed, and depth of cut read from the CNC process library before machining.

3. The method for evaluating the machining status of a CNC machine tool based on multi-source sensor data according to claim 1, characterized in that: The multi-source data modality partitioning includes: Based on signal dynamics, data format, and correlation with processing state, the preprocessed data is divided into four independent modes: Vibration modes: Only the fused vibration time-domain signal is included. The input is 1×10000 dimensions in a 1s window and 1×5000 dimensions in a 0.5s window. Acousto-electric mode: Integrating acoustic emission pulses and current waveforms, the 1s window input is 2×10000 dimensions, and the 0.5s window is 2×5000 dimensions, which belongs to the hybrid characteristics of pulse and continuous waveforms; CNC modality: covers spindle load rate and feed rate parameters. The input is 5×100 dimensions in a 1s window and 5×50 dimensions in a 0.5s window. Measurement modes: including flatness error, tool deviation and process parameters, 1s window input is 5×10 dimensions, 0.5s window is 5×5 dimensions; After partitioning, the data missing rate is checked modally. When the missing rate exceeds 5%, the average of adjacent windows is used to fill the missing data to ensure data continuity.

4. The method for evaluating the machining status of a CNC machine tool based on multi-source sensor data according to claim 1, characterized in that: The temporal context modeling includes: A modality-adaptive GRU network is used to capture long-term temporal dependencies of each modality: Two-layer GRUs are configured for vibration modes, with 128 hidden nodes in each layer and a dropout rate of 0.

2. The length of the input sequence matches the data window, with 100 feature points in a 1-second window and 50 feature points in a 0.5-second window. The acoustic-electric mode uses a 2-layer GRU with 64 hidden nodes in each layer and a dropout rate of 0.

2. The input sequence length is the same as that of the vibration mode. The CNC modal configuration has a 1-layer GRU, 64 hidden layer nodes, a dropout rate of 0.1, and the input sequence length is consistent with the vibration mode. The measurement modality uses a 1-layer GRU with 32 hidden layer nodes, a dropout rate of 0.1, and an input sequence length of 10 1s windows or 5 0.5s window feature points. The GRU network weights are initialized with a Xavier normal distribution, the bias term is set to 0, and the hidden state of the last time step is taken as the global temporal context vector of each mode, which respectively characterizes the evolution of vibration signal, acoustic-electric correlation, processing parameter trend and measurement reference and temporal change characteristics.

5. The method for evaluating the machining status of a CNC machine tool based on multi-source sensor data according to claim 1, characterized in that: The cross-attention modal interaction includes: The scaling dot product attention mechanism is used to achieve multi-view active feature fusion. The core is to mine modal associations through Query-Key-Value interaction logic: the global temporal context vectors of four modalities, vibration, acoustics, CNC and measurement, are used as the core query. The context vectors of the other three modalities are concatenated as key and value. After linear transformation, the dimensions of key and value are matched with the query. Attention scores are calculated and weights are generated through softmax to obtain the fusion vector of each view. The remaining three modalities include: When vibration mode is used as the core query, acoustic-electric, CNC, and measurement modes are used as the remaining modes; When using acoustic-electric modes as the core query, vibration, CNC, and measurement modes are used as the remaining modes; When using CNC mode as the core query, vibration, acoustic-electric, and measurement modes are used as the remaining modes; When the measured mode is used as the core query, the vibration, acoustic-electric, and CNC modes are used as the remaining modes; The accuracy of fault identification from each perspective is calculated based on the validation set. The perspective weights are calibrated according to the accuracy ratio. The fusion vectors of acoustic, CNC, and measurement perspectives are extended to the vibration perspective vector dimension through zero padding. The fusion vectors of the four perspectives are weighted and summed according to the calibrated weights to generate a 128-dimensional fusion feature that combines local features, temporal information, and cross-modal correlation. The four types of perspective fusion vectors include: fusion vectors generated by using vibration, acoustic-electric, CNC, and measurement modes as queries respectively.

6. The method for evaluating the machining status of a CNC machine tool based on multi-source sensor data according to claim 1, characterized in that: The fusion features include: For the 128-dimensional fusion features generated by cross-attention modal interaction, the Pearson correlation coefficient between features is first calculated, and redundant features with an absolute value greater than 0.9 are removed, retaining 64 effective features. Then, through PCA principal component analysis, principal components are selected according to the principle that the cumulative variance contribution rate is greater than 95%, reducing the dimensionality to 48 features. Min-Max normalization is used to map the features to the [0, 1] interval, and the normalization parameters are determined based on the extreme values ​​of the training set features. Finally, the features are quantized from float32 precision to int8 precision using TensorFlow Lite to generate core fusion features that are adapted for engineering. After quantization, the inference latency is compressed to less than 15ms, providing high-quality input for uncertainty quantification evaluation.

7. The method for evaluating the machining status of a CNC machine tool based on multi-source sensor data according to claim 1, characterized in that: The evidence deep learning model includes: The model input is the 48-dimensional normalized fusion feature optimized in step S2, which integrates multimodal correlation information of vibration, acoustics, CNC, and measurement. The output generates four evidence parameters for each of the two core processing state indicators, chatter and tool wear, which correspond to the position, degree of freedom, shape, and scale characteristics of the inverse gamma distribution, respectively, and are used to calculate the state prediction value and cognitive uncertainty in the subsequent calculation. The model architecture consists of an input layer, hidden layers, and an output layer: the input layer directly adapts to the 48-dimensional feature dimension; the hidden layer has 3 fully connected layers with 128, 64, and 32 nodes respectively, all using the ReLU activation function, and each layer is followed by a normalization layer; the output layer contains 8 nodes, has no activation function, and during training, the shape parameter is constrained to be greater than 1 and the scale parameter to be greater than 0. The loss function is a weighted combination of evidence loss and supervision loss, with a balance coefficient of 0.

5. The evidence loss includes the likelihood probability term and regularization term of the predicted value and the true value, and the supervision loss uses the mean square error to constrain the prediction bias. The model is trained on 6,000 sets of samples covering multiple materials and working conditions. The samples are divided into training set, validation set and test set in a ratio of 7:2:

1. An optimizer is used and an early stopping mechanism is configured. After calibration of evidence parameters and fine-tuning of uncertainty threshold, the prediction accuracy of chatter and tool wear on the validation set is no less than 96%. During online inference, the model outputs state values, uncertainty variance and confidence level results.

8. The method for evaluating the machining status of a CNC machine tool based on multi-source sensor data according to claim 1, characterized in that: The online assessment includes: The 48-dimensional normalized fusion features optimized in step S2 are received in real time and transmitted to the model through a low-latency transmission method. At the same time, it is checked whether the feature values ​​are in the range of [0, 1]. If they are outside the range, the feature is marked as abnormal and the valid features of the previous cycle are used as a temporary substitute. The quantized evidence deep learning model is invoked to infer the input features and output four evidence parameters for each of the two states: chatter and tool wear. The inference time is controlled within 10ms. The position parameters output by the model are used as the state prediction values. The cognitive uncertainty variance is calculated by combining the degrees of freedom, shape, and scale parameters. The larger the variance, the lower the reliability of the model's prediction results. The confidence level is divided according to the variance of uncertainty: a variance of less than 0.1 indicates high confidence, 0.1 to 0.3 indicates medium confidence, and a variance of not less than 0.3 indicates low confidence. The final output includes timestamp, state prediction value, uncertainty, confidence level, and calculation basis. The output frequency is consistent with the feature output frequency in step S2.

9. The method for evaluating the machining status of a CNC machine tool based on multi-source sensor data according to claim 1, characterized in that: The application and output of the evaluation results include: Cross-end transmission through multiple industrial communication protocols: push full results to monitoring terminals using lightweight protocols, transmit core parameters to CNC systems using standardized protocols, and allocate buffers and retransmissions to ensure reliability. The deployment control agent makes decisions based on status values ​​and confidence levels. For issues such as chatter and tool wear, emergency adjustments are performed at high confidence levels, fine-tuning compensation is made at medium confidence levels, and only early warnings are issued at low confidence levels, with execution priorities assigned accordingly. Verify the effectiveness of control commands, collect end-to-end data to evaluate the efficiency of decision-making, and periodically optimize rule parameters and retrain the model.

Images