Intelligent fault diagnosis method for numerical control machine tool

Abstract

The application provides an intelligent fault diagnosis method of a numerical control machine tool, which is applied to a control system of the numerical control machine tool. The method comprises the following steps: collecting multi-source sensor data in a running process of the numerical control machine tool, wherein the multi-source sensor data comprises at least two of a vibration signal, a temperature signal, a current signal and a sound signal; pre-processing and extracting features of the multi-source sensor data to obtain pre-processed data; constructing and training a machine tool dynamic model based on the pre-processed data to obtain a target machine tool dynamic model; when detecting that the numerical control machine tool is abnormal, performing fault state estimation based on the target machine tool dynamic model and a model predictive control method to obtain a fault estimation result; and constructing a fault evolution path by using a Monte Carlo tree search algorithm based on the fault estimation result and outputting a fault diagnosis result. The application can improve the accuracy and efficiency of fault diagnosis of the numerical control machine tool.

Description

Technical Field

[0001] This application relates to the field of CNC machine tool technology, and in particular to an intelligent fault diagnosis method for CNC machine tools. Background Technology

[0002] CNC machine tools are key equipment in modern industrial production, and their operating status directly affects production efficiency and processing quality. With the improvement of industrial automation, higher requirements are placed on the intelligence and accuracy of CNC machine tool fault diagnosis.

[0003] Currently, common CNC machine tool fault diagnosis methods mainly include threshold monitoring based on a single sensor and rule-based judgment based on expert experience. For example, a vibration sensor monitors the vibration amplitude of a bearing, and a fault is determined when it exceeds a preset threshold; or a temperature sensor monitors the temperature of key components, and fault judgment is made in combination with expert experience.

[0004] More advanced diagnostic methods employ traditional machine learning algorithms, such as Support Vector Machines (SVM), to classify fault features. This method learns from historical fault data to establish a mapping relationship between fault features and fault types, thereby achieving fault identification.

[0005] However, existing technologies have the following problems: on the one hand, a single sensor or a single algorithm is difficult to fully reflect the complex operating state of a machine tool, resulting in insufficient accuracy of diagnostic results; on the other hand, traditional methods lack the ability to predict the development trend of faults and cannot provide timely warnings of potential fault risks. Summary of the Invention

[0006] In view of this, this application provides an intelligent fault diagnosis method and device for CNC machine tools, which solves the problem that in the prior art, a single sensor or a single algorithm is difficult to fully reflect the complex operating state of the machine tool and lacks the ability to predict the fault development trend.

[0007] This application provides an intelligent fault diagnosis method for CNC machine tools, applied to the control system of CNC machine tools, including:

[0008] The CNC machine tool collects multi-source sensor data during operation, including at least two of vibration signals, temperature signals, current signals, and sound signals.

[0009] The multi-source sensor data is preprocessed and features are extracted to obtain preprocessed data;

[0010] Based on the preprocessed data, a dynamic model of the machine tool is constructed and trained to obtain the dynamic model of the target machine tool.

[0011] When an abnormal operation of a CNC machine tool is detected, a fault state prediction is performed based on the target machine tool dynamic model and model predictive control method to obtain the fault prediction result.

[0012] Based on the fault prediction results, a fault evolution path is constructed using the Monte Carlo tree search algorithm, and fault diagnosis results are output.

[0013] The step of constructing and training a machine tool dynamic model based on the preprocessed data to obtain a target machine tool dynamic model includes:

[0014] A deep learning model architecture is constructed, which includes a convolutional neural network (CNN) for spatial feature extraction, a bidirectional long short-term memory network (LSTM) for temporal feature processing, and a fully connected layer for fusing the features extracted by the CNN and the LSTM.

[0015] Based on the outputs of the CNN and the LSTM, multiple loss functions are designed, including: prediction error loss for measuring prediction accuracy and L2 regularization constraint for preventing model overfitting, wherein different loss weights are set for different types of fault samples.

[0016] Based on the calculation results of the loss function, the Adam optimizer is used for model training, and the learning rate is dynamically adjusted based on the performance index of the validation set. When the loss value of the validation set does not decrease for three consecutive training cycles, the learning rate is adjusted to 0.8 times the original value.

[0017] The model evaluation mechanism based on K-fold cross-validation is used to determine the dynamic model of the target machine tool. This includes: in each training fold, a hierarchical sampling strategy is used to divide the dataset to ensure a balanced distribution of samples of different fault types, and the model performance is evaluated by statistically analyzing the accuracy, precision, and recall on the validation set of each fold, and the optimal model parameter configuration is selected.

[0018] In the deep learning model architecture:

[0019] The CNN consists of three convolutional layers connected in sequence. The first layer uses 32 3×3 convolutional kernels for feature extraction and normalization through a BatchNormalization layer. The second layer uses 64 5×5 convolutional kernels to expand the receptive field and enhances non-linear expression through the ReLU activation function. The third layer reduces dimensionality through a 2×2 max pooling window.

[0020] The LSTM consists of two bidirectional LSTM layers, each containing 128 neurons, with a Dropout layer with a dropout rate of 0.3 between the two layers;

[0021] The fully connected layer consists of three layers with corresponding node numbers of 512, 256 and 128 respectively. It is used to fuse spatial features and temporal features, and a BatchNormalization layer is added before the last layer to ensure stable feature distribution.

[0022] When an abnormal operation of a CNC machine tool is detected, a fault state prediction is performed based on the target machine tool's dynamic model and model predictive control method to obtain a fault prediction result, including:

[0023] Based on the target machine tool dynamic model, a state predictor with a recursive prediction structure is constructed. The state predictor uses a time window length of 20 sampling points as a prediction unit.

[0024] Using the state predictor, an adaptive rolling time-domain prediction strategy is designed. The rolling time-domain prediction strategy includes short-term prediction with 3 sampling points, medium-term prediction with 6 sampling points, and long-term prediction with 10 sampling points. The prediction step size is dynamically adjusted according to the working conditions of the CNC machine tool.

[0025] Based on the rolling time-domain prediction strategy, the future state of the CNC machine tool is predicted in multiple steps. The prediction strategy outputs the results and the deviation between the predicted state and the normal operating state of the CNC machine tool is calculated in real time to assess the fault risk level.

[0026] The deviation between the real-time calculated predicted state and the normal operating state of the CNC machine tool includes:

[0027] Construct a health status benchmark library for CNC machine tools that includes the distribution of normal operating characteristics under different cutting parameters;

[0028] Calculate the Mahalanobis distance between the predicted state sequence and the corresponding working condition reference state in the reference library to obtain the state deviation value;

[0029] Linear regression analysis was performed on the state deviation values ​​to obtain the trend characteristics of deviation changes;

[0030] The state transition matrix is ​​trained based on historical fault data, and the fault development trend is predicted by combining the current state deviation value and the deviation change trend.

[0031] Based on the state deviation value, deviation change trend and fault development trend, the fuzzy comprehensive evaluation method is used to classify the operating status of CNC machine tools into four levels: normal, minor abnormality, moderate fault and severe fault.

[0032] The state predictor includes the following feedback correction mechanism:

[0033] An error tracker is used to calculate a sequence of deviations between predicted and actual observations.

[0034] An adaptive weight adjuster dynamically adjusts the correction weights within a range of 0.6 to 1.0 based on the prediction accuracy, where the prediction accuracy is calculated using the mean square error between the predicted and actual values.

[0035] The state corrector performs real-time correction on the prediction results based on the correction weights.

[0036] The step of constructing a fault evolution path based on the fault prediction results using the Monte Carlo tree search algorithm includes:

[0037] Based on the fault prediction results, an initial state node is established, wherein the attribute information of the initial state node represents the feature vector, fault risk level and confidence level of the current state of the CNC machine tool;

[0038] A state transition rule base is established based on this node. The state transition rule base includes: component-level fault transfer rules, system-level fault transfer rules, and environmental factor influence rules. Each rule is configured with a corresponding transition probability and time constraint.

[0039] Based on the state transition rule base, a heuristic method is used to dynamically determine the expansion depth. Specifically, a maximum expansion depth of 8 layers is used for important fault paths, and a maximum expansion depth of 4 layers is used for minor fault paths.

[0040] An improved UCT algorithm is used for node selection and expansion to determine the failure evolution path.

[0041] The use of the improved UCT algorithm for node selection and expansion to determine the fault evolution path includes:

[0042] Calculate the node score: Score = Q / N + c*sqrt(ln(N_parent) / N) + w*R,

[0043] Where Q represents the cumulative value of a node, N represents the number of times a node is visited, N_parent represents the number of times the parent node is visited, c represents the exploration coefficient, R represents the node risk coefficient, and w represents the risk weight.

[0044] Child nodes are generated based on the rules in the state transition rule base, and the transition probability is dynamically adjusted according to historical fault data during the expansion process.

[0045] Node simulations are performed using a fast stochastic simulation method, recording the cumulative risk value and time delay along the path;

[0046] Starting from the leaf node, update the number of visits and cumulative value of each node upwards along the path.

[0047] This application also performs at least one of the following operations:

[0048] The number of visits to each node in the fault evolution path is counted, and the path with more than 30% of the total number of visits is identified as a high-probability fault evolution path.

[0049] A hierarchical fault diagnosis report is generated based on the high-probability fault evolution path, including a description of the current fault status, early warning of potential fault evolution, and corresponding maintenance suggestions;

[0050] The fault evolution tree structure is displayed in a visual way, with different colors indicating the fault level, node size representing the state transition probability, and the thickness of the connecting edges representing the path traversal frequency.

[0051] This application embodiment also provides an intelligent fault diagnosis device for CNC machine tools, including:

[0052] The data acquisition unit is used to acquire multi-source sensor data during the operation of the CNC machine tool. The multi-source sensor data includes at least two of the following: vibration signal, temperature signal, current signal, and sound signal.

[0053] The preprocessing unit is signal-connected to the data acquisition unit and is used to receive the multi-source sensor data and perform preprocessing and feature extraction to obtain preprocessed data.

[0054] The model training unit is signal-connected to the preprocessing unit and is used to receive the preprocessed data and, based on the preprocessed data, construct and train the machine tool dynamic model to obtain the target machine tool dynamic model.

[0055] The state prediction unit is signal-connected to the model training unit and is used to perform fault state prediction based on the target machine tool dynamic model and model predictive control method when an abnormal operation of the CNC machine tool is detected, so as to obtain the fault prediction result.

[0056] The diagnostic unit is signal-connected to the state prediction unit and is used to construct a fault evolution path based on the fault prediction result using a Monte Carlo tree search algorithm, and output the fault diagnosis result.

[0057] This application embodiment also provides a computer device, the computer device comprising:

[0058] At least one processor; and,

[0059] A memory communicatively connected to the at least one processor; wherein,

[0060] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the intelligent fault diagnosis method for the CNC machine tool described above.

[0061] This application also provides a computer-readable storage medium that stores computer instructions for causing a computer to execute the above-described intelligent fault diagnosis method for CNC machine tools.

[0062] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the intelligent fault diagnosis method for CNC machine tools described above.

[0063] This application has the following technical effects:

[0064] 1. By collecting and fusing data from multiple sources of sensors, the ability to perceive the operating status of machine tools and the efficiency of information utilization have been improved.

[0065] 2. The CNN-LSTM hybrid deep learning architecture is adopted, which effectively combines spatial and temporal features, improving the accuracy of fault feature extraction and classification.

[0066] 3. By introducing model predictive control and adaptive feedback mechanisms, accurate prediction and dynamic correction of fault states are achieved.

[0067] 4. The innovative application of Monte Carlo tree search to construct fault evolution paths improves the interpretability and early warning capability of fault diagnosis. Attached Figure Description

[0068] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0069] Figure 1 This is a flowchart illustrating the intelligent fault diagnosis method for CNC machine tools provided in the embodiments of this application;

[0070] Figure 2 This is a schematic diagram of the multi-source sensor data acquisition and preprocessing process provided in the embodiments of this application;

[0071] Figure 3 This is a schematic diagram of the deep learning model architecture provided in the embodiments of this application;

[0072] Figure 4 This is a schematic diagram of the state predictor structure provided in an embodiment of this application. Detailed Implementation

[0073] This application provides an intelligent fault diagnosis method and device for CNC machine tools, the preliminary technical solution of which is as follows:

[0074] S1: Collect multi-source sensor data during the operation of the CNC machine tool, wherein the multi-source sensor data includes at least two of the following: vibration signal, temperature signal, current signal, and sound signal.

[0075] This application uses multi-source sensors deployed in key parts of CNC machine tools to collect machine tool operating status data.

[0076] Specifically, combined Figure 2 Vibration sensors are installed on key functional components such as the machine tool spindle, feed system, and tool system to collect vibration signals during machine tool operation; temperature sensors are installed on heat-prone parts such as the spindle box, motor, and bearing housing to monitor temperature changes in real time; current sensors are installed on the spindle motor and feed motor to collect motor operating current; and sound sensors are installed around the machine tool workspace to collect acoustic signals during machining.

[0077] It should be noted that vibration signals are mainly used to reflect the dynamic characteristics and mechanical faults of machine tools, such as bearing faults and gear faults; temperature signals can reflect the thermal state of machine tool components and help identify faults such as overload and poor lubrication; current signals can reflect the motor load state and electrical system faults; and sound signals can assist in judging the processing status and mechanical faults.

[0078] In practical applications, at least two types of sensor signals can be combined depending on the machine tool type and monitoring requirements. For example, for precision machining centers, vibration and temperature signals can be collected as the main focus; for heavy-duty cutting machine tools, vibration, temperature, and current signals can be collected simultaneously. This multi-source data acquisition method can comprehensively reflect the operating status of the machine tool from different perspectives, providing a reliable data foundation for subsequent fault diagnosis.

[0079] In addition, the sampling frequency of various sensors needs to be configured appropriately according to the characteristics of the signal. For example, the sampling frequency of vibration signals is usually set to above 10kHz to capture high-frequency fault characteristics; the sampling frequency of temperature signals can be relatively low, set to around 1Hz; and the sampling frequency of current signals needs to be determined according to the motor control cycle, generally around 1kHz.

[0080] In conclusion, by properly configuring a multi-source sensor network, a comprehensive perception of the machine tool's operating status can be achieved, laying a solid data foundation for subsequent intelligent fault diagnosis.

[0081] It is important to emphasize that the sensor's installation location, sampling parameters, and other configurations directly affect data quality. Therefore, in practical applications, it is necessary to optimize the design by combining the characteristics of the machine tool and professional experience.

[0082] S2: Preprocess and extract features from the multi-source sensor data to obtain preprocessed data.

[0083] Preprocessing and feature extraction of multi-source sensor data are key steps in achieving accurate fault diagnosis.

[0084] Combination Figure 2 First, the collected raw data undergoes data cleaning, primarily including outlier detection and noise removal. Specifically, the 3σ criterion based on statistical characteristics is used to detect outliers, and data points exceeding the normal range are marked. Then, median filtering is used to eliminate impulse noise and random interference in the signal. The filter window size is dynamically adjusted according to the signal characteristics; for example, a 5-9 point window is selected for vibration signals, and a 3-5 point window is selected for temperature signals.

[0085] Secondly, to address the issue of missing values ​​that may occur during data acquisition, a piecewise linear interpolation method is used for data repair. In practical applications, data loss may occur due to temporary sensor malfunctions or communication interruptions. In such cases, by analyzing the time span of the missing data and the changing trends of adjacent data, an adaptive interpolation strategy is adopted: for short-term missing data (less than 3 sampling periods), linear interpolation is used; for longer-term missing data (3-10 sampling periods), cubic spline interpolation is used; and for even longer-term missing data, it needs to be marked and specially handled in subsequent analysis.

[0086] Furthermore, considering the significant differences in the data dimensions and distribution characteristics of different types of sensors, data standardization is necessary.

[0087] This application employs an improved Min-Max normalization method to uniformly map various types of sensor data to the [0,1] interval. In particular, when calculating the mapping range, not only are the statistical characteristics of historical data considered, but a certain margin is also reserved to accommodate possible data fluctuations, typically by extending the range of the maximum and minimum values ​​by 10% respectively.

[0088] After data preprocessing is completed, time synchronization and feature extraction are required.

[0089] First, multi-source sensor data is aligned based on timestamps to ensure consistency across different sources in the time dimension. Then, corresponding feature parameters are extracted for different signal types: for vibration signals, time-domain features (mean, variance, peak value, kurtosis, etc.) and frequency-domain features (power spectral density, dominant frequency component, etc.) are extracted; for temperature signals, statistical features (mean, rate of change, etc.) and trend features are mainly extracted; for current signals, in addition to basic statistical features, spectral features are also extracted to reflect the motor's operating state; for sound signals, energy features and frequency band features are the focus.

[0090] Finally, the extracted features are combined to form a feature vector.

[0091] It should be noted that the feature extraction process adopts a sliding window mechanism. The window length is determined according to the signal characteristics and the time scale of the fault features. Typically, the window length for vibration signals is 1024 points, and the window length for temperature signals is 100 points. The window overlap rate is set to 50% to ensure the continuity of features.

[0092] In summary, systematic data preprocessing and feature extraction can significantly improve data quality, providing reliable input features for subsequent model training. Furthermore, the preprocessing scheme presented in this application exhibits strong adaptability and robustness, effectively addressing various data problems encountered in practical engineering.

[0093] S3: Based on the preprocessed data, construct and train the machine tool dynamic model to obtain the target machine tool dynamic model.

[0094] Based on the preprocessed data, the process of constructing and training the dynamic model of the machine tool in this application includes several key steps.

[0095] First, an innovative deep learning model architecture was designed, which combines the advantages of CNN and LSTM and can process spatial and temporal features simultaneously.

[0096] Specifically, the CNN part is mainly used to extract local features and spatial correlations of various sensor signals, while the LSTM part is responsible for capturing long-term temporal dependencies. The outputs of the two networks are fused through a fully connected layer to achieve comprehensive modeling of the dynamic characteristics of the machine tool.

[0097] During model training, multiple loss functions were designed to guide model optimization.

[0098] The prediction error loss measures the degree of matching between the model output and the actual fault type, and uses the cross-entropy loss function; the L2 regularization constraint is used to prevent the model from overfitting, and is achieved by adding a sum of squares of the model parameters to the loss function.

[0099] In particular, different loss weights are set for different types of fault samples, with higher weights assigned to severe faults and critical component fault samples to improve the model's ability to identify important faults.

[0100] The training process employs the Adam optimizer for parameter updates. Compared to traditional stochastic gradient descent, the Adam optimizer features adaptive learning rate adjustment, better addressing gradient sparsity issues. Simultaneously, a dynamic learning rate adjustment mechanism based on validation set performance is introduced: if the validation set loss value does not decrease for three consecutive training epochs, the learning rate is adjusted to 0.8 times its original value. This strategy helps the model escape local optima and achieve better generalization performance.

[0101] To ensure the reliability of the model, a K-fold cross-validation evaluation mechanism is adopted.

[0102] First, a stratified sampling strategy is used to partition the dataset, ensuring that the distribution of various fault samples in the training and validation sets is basically consistent. During each training fold, model performance is evaluated using metrics such as accuracy, precision, and recall on the validation set. Finally, based on the results of multiple rounds of cross-validation, the model parameter configuration with the best overall performance is selected as the dynamic model for the target machine tool.

[0103] See Figure 3 In the specific network structure design, the CNN part adopts a three-layer convolutional structure. The first layer uses 32 3×3 convolutional kernels for initial feature extraction and normalizes the feature distribution through a BatchNormalization layer; the second layer uses 64 5×5 convolutional kernels to expand the receptive field and enhances the nonlinear expressive ability of the model through the ReLU activation function; the third layer uses a 2×2 max pooling window for feature dimensionality reduction and extracts the most significant feature information.

[0104] The LSTM portion employs a bidirectional structure, consisting of two LSTM layers, each containing 128 neurons. This bidirectional structure allows for the simultaneous consideration of forward and backward dependencies in the sequence, enhancing the model's ability to understand temporal patterns. A dropout layer with a dropout rate of 0.3 is placed between the two LSTM layers to prevent overfitting by randomly deactivating some neurons.

[0105] The feature fusion part employs a three-layer fully connected network with 512, 256, and 128 nodes respectively, achieving progressive feature dimensionality reduction and high-level semantic extraction. A BatchNormalization layer is added before the last fully connected layer to ensure the stability of the feature distribution, which is beneficial for model convergence and generalization.

[0106] In summary, through a carefully designed deep learning architecture and systematic training strategy, this application constructs an intelligent model that accurately reflects the dynamic characteristics of machine tools. This model possesses strong feature extraction and fault identification capabilities, providing a reliable foundation for subsequent fault state prediction.

[0107] S3 specifically includes:

[0108] S3.1: Construct a deep learning model architecture, which includes a convolutional neural network (CNN) for spatial feature extraction, a bidirectional long short-term memory network (LSTM) for temporal feature processing, and a fully connected layer for fusing the features extracted by the CNN and the LSTM.

[0109] This application designs a deep learning model architecture that integrates CNN and LSTM, which can simultaneously process the spatial and temporal features of machine tool operation data. The CNN part is primarily responsible for extracting the spatial features of multi-source sensor data, including the correlation between signals from different sensors and local feature patterns. The LSTM part focuses on processing the temporal features of the data, capturing long-term dependencies through a bidirectional structure, thus enabling the understanding of the evolution of the machine tool's state over time.

[0110] The feature extraction results from the two networks are fused through a fully connected layer. This fusion mechanism allows the model to learn the optimal combination of spatial and temporal features, thereby forming a comprehensive representation of the machine tool's state. The fully connected layer maps the fused features to the final fault diagnosis space through layer-by-layer dimensionality reduction and nonlinear transformation.

[0111] S3.2: Based on the outputs of the CNN and the LSTM, design multiple loss functions, including: prediction error loss for measuring prediction accuracy and L2 regularization constraint for preventing model overfitting, wherein different loss weights are set for different types of fault samples.

[0112] Two key loss functions were designed during model training. The prediction error loss uses the cross-entropy loss function to measure the difference between the model's predictions and the actual fault types. This loss function assigns differentiated weights to different types of fault samples; for example, severe fault samples such as bearing faults and spindle faults are given higher weights (typically 1.5-2.0), while general fault samples are given standard weights (1.0). This weighting strategy ensures that the model focuses more on identifying critical faults.

[0113] L2 regularization is implemented by adding a sum of squares of the model parameters to the loss function, with coefficients typically set between 0.001 and 0.01. This constraint mechanism can suppress excessive growth of model parameters, prevent the model from overfitting the training data, and improve its generalization ability in practical applications.

[0114] S3.3: Based on the calculation results of the loss function, the Adam optimizer is used to train the model, and the learning rate is dynamically adjusted based on the performance index of the validation set. When the loss value of the validation set does not decrease for three consecutive training cycles, the learning rate is adjusted to 0.8 times the original value.

[0115] The model was trained using the Adam optimizer, an adaptive moment estimation optimization algorithm. This optimizer combines the advantages of momentum and RMSprop, and can adaptively adjust the learning rate of each parameter. In actual training, the initial learning rate of the Adam optimizer was set to 0.001, the β1 parameter was set to 0.9, and the β2 parameter was set to 0.999.

[0116] At the same time, a dynamic learning rate adjustment mechanism based on the performance of the validation set was introduced.

[0117] Specifically, after each training epoch, the model's loss value on the validation set is evaluated. If the validation set loss value does not decrease for three consecutive training epochs, the current learning rate is reduced to 0.8 times its original value. This strategy helps the model converge better in the later stages of training, avoiding oscillations around local optima.

[0118] S3.4: A model evaluation mechanism based on K-fold cross-validation is adopted to determine the dynamic model of the target machine tool, including: in each fold of training, a hierarchical sampling strategy is used to divide the dataset to ensure a balanced distribution of samples of different fault types, and the model performance is evaluated by statistically analyzing the accuracy, precision, and recall on the validation set of each fold, and the optimal model parameter configuration is selected.

[0119] This application employs a K-fold cross-validation mechanism to evaluate model performance, typically with K set to 5 or 10. When partitioning the dataset, a stratified sampling strategy is used to ensure that the proportion of each type of fault sample in each subset is essentially consistent with the original dataset. For example, if a certain type of fault sample accounts for 15% in the original dataset, then the proportion of that type of fault sample in each partitioned subset should also remain around 15%.

[0120] During each training fold, the model is evaluated by calculating multiple performance metrics on the validation set. These metrics include accuracy (overall prediction accuracy), precision (the proportion of true positives among those predicted as positives), and recall (the proportion of true positives correctly predicted). Based on the comprehensive score of these metrics, the model parameter configuration with the best performance is selected as the final model.

[0121] In the deep learning model architecture:

[0122] The CNN consists of three convolutional layers connected in sequence. The first layer uses 32 3×3 convolutional kernels for feature extraction and normalization through a BatchNormalization layer. The second layer uses 64 5×5 convolutional kernels to expand the receptive field and enhances non-linear expression through the ReLU activation function. The third layer reduces dimensionality through a 2×2 max pooling window.

[0123] The LSTM consists of two bidirectional LSTM layers, each containing 128 neurons, with a Dropout layer between the two layers having a dropout rate of 0.3.

[0124] The fully connected layer consists of three layers with corresponding node numbers of 512, 256 and 128 respectively. It is used to fuse spatial features and temporal features, and a BatchNormalization layer is added before the last layer to ensure stable feature distribution.

[0125] The CNN part employs a three-layer convolutional structure for spatial feature extraction. The first layer uses 32 3×3 convolutional kernels for initial feature extraction with a stride of 1, and performs feature normalization through a Batch Normalization layer, which helps accelerate network training and improve model stability. The second layer uses 64 5×5 convolutional kernels. These larger kernels can capture a wider range of spatial correlations, and the ReLU activation function introduces a non-linear transformation to enhance the model's expressive power. The third layer uses a 2×2 max-pooling window for dimensionality reduction, preserving the most salient feature information.

[0126] The LSTM part consists of two bidirectional LSTM layers, each containing 128 neurons. The bidirectional structure can simultaneously consider forward and backward information of the sequence, providing a more complete temporal feature representation. The Dropout layer (dropout rate 0.3) placed between the two LSTM layers effectively prevents overfitting by randomly deactivating some neurons.

[0127] The fully connected layer employs a three-layer structure with 512, 256, and 128 nodes respectively, achieving progressive dimensionality reduction and abstraction of features. A BatchNormalization layer added before the last layer ensures stable feature distribution, which is beneficial for subsequent fault diagnosis. This progressive dimensionality reduction design retains important feature information while reducing computational complexity.

[0128] S4: When an abnormal operation of the CNC machine tool is detected, the fault state is predicted based on the target machine tool dynamic model and the model predictive control method to obtain the fault prediction result.

[0129] This step primarily addresses the problem of fault prediction under abnormal machine tool operating conditions. When the monitoring system detects that the machine tool's operating parameters deviate from the normal range, the system immediately initiates a fault state prediction process based on the target machine tool's dynamic model. This process combines the predictive capabilities of deep learning models with the optimization characteristics of model predictive control, enabling not only the assessment of the current fault state but also the prediction of the fault's development trend. The prediction process fully utilizes historical and real-time monitoring data, employing a sliding window mechanism to achieve continuous state assessment and prediction updates.

[0130] Combination Figure 1 and Figure 4 S4 specifically includes the following steps:

[0131] S4.1: Based on the target machine tool dynamic model, a state predictor with a recursive prediction structure is constructed. The state predictor uses a time window length of 20 sampling points as a prediction unit.

[0132] Based on a pre-trained dynamic model of the target machine tool, a state predictor with a recursive prediction structure was constructed. This predictor uses a time window of 20 sampling points as the basic prediction unit; the selection of this length is based on a comprehensive consideration of multiple factors:

[0133] First, considering the characteristic evolution cycle of common machine tool faults, the time span of 20 sampling points is sufficient to capture the initial features of most fault modes. Second, this window length achieves a good balance between computational complexity and prediction accuracy, ensuring timely prediction without introducing excessive historical noise due to an excessively long window. The recursive structure of the predictor allows new observation data to continuously update the prediction basis, thereby enabling continuous tracking of the machine tool status.

[0134] S4.2: Using the state predictor, design an adaptive rolling time-domain prediction strategy. The rolling time-domain prediction strategy includes short-term prediction with 3 sampling points, medium-term prediction with 6 sampling points, and long-term prediction with 10 sampling points. The prediction step size is dynamically adjusted according to the working conditions of the CNC machine tool.

[0135] The rolling time-domain prediction strategy employs a multi-scale prediction scheme, designing three prediction levels to meet prediction needs at different time scales. Short-term prediction (3 sampling points) focuses on the immediate evolution of faults, with a short sampling interval, typically 50-100ms, enabling timely capture of sudden anomalies. Medium-term prediction (6 sampling points) focuses on the development trend of faults, with the sampling interval appropriately extended to 200-300ms, aiding in the identification of gradual faults. Long-term prediction (10 sampling points) is used to assess the potential impact of faults, with sampling intervals exceeding 500ms. The prediction step size is dynamically adjusted according to the machine tool's operating conditions. For example, during high-speed cutting, the prediction step size is shortened to 70% of the baseline value, while during low-speed stable machining, it can be extended to 130% of the baseline value to adapt to the evolution speed of fault characteristics under different operating conditions.

[0136] S4.3: Based on the rolling time-domain prediction strategy, perform multi-step prediction of the future state of the CNC machine tool, output the prediction strategy, and calculate the deviation between the predicted state and the normal operating state of the CNC machine tool in real time to assess the fault risk level.

[0137] Based on a rolling time-domain prediction strategy, the system continuously predicts the future state of the machine tool in multiple steps. The prediction process not only outputs the predicted values ​​of state parameters at each time point but also calculates the deviation between these predicted states and the machine tool's normal operating state in real time. The deviation calculation uses a weighted Euclidean distance, assigning different weights to different parameters; for example, vibration characteristics have a weight of 0.4, temperature characteristics 0.3, and current characteristics 0.3. The system categorizes fault risk into four levels based on the magnitude of the deviation: a safe state when the deviation is less than the threshold T1 (usually set to 1.2 times the normal value), a warning state when the deviation is between T1 and T2 (1.5 times), a cautionary state when the deviation is between T2 and T3 (2 times), and a dangerous state when the deviation is greater than T3. This real-time assessment mechanism can promptly identify potential fault risks, providing decision support for preventative maintenance.

[0138] Furthermore, in one embodiment, the deviation between the real-time calculated predicted state and the normal operating state of the CNC machine tool in S4.3 includes:

[0139] A1. Construct a health status benchmark library for CNC machine tools that includes the distribution of normal operating characteristics under different cutting parameters.

[0140] The construction of the health status benchmark library is a systematic process that requires collecting and organizing normal operating data of machine tools under different cutting conditions. The benchmark library is classified according to different combinations of cutting parameters, mainly including: speed range (usually divided into low speed 800-2000rpm, medium speed 2000-4000rpm, and high speed above 4000rpm), feed rate (divided into 5 levels within the range of 0.1-0.5mm / rev), and depth of cut (divided into 4 levels within the range of 0.5-2.5mm), etc.

[0141] For each parameter combination, its characteristic distribution during normal operation is collected and statistically analyzed, including vibration characteristics (mean, standard deviation, peak value, etc.), temperature characteristics (temperature at each measuring point and its rate of change), and current characteristics (RMS value, harmonic components, etc.). These characteristics are modeled using probability density functions to form a standardized reference benchmark. The benchmark library is updated using a sliding window mechanism, updating at fixed intervals (e.g., once a month) to adapt to the slow changes in machine tool performance.

[0142] A2: Calculate the Mahalanobis distance between the predicted state sequence and the corresponding working condition reference state in the reference library to obtain the state deviation value.

[0143] This step uses Mahalanobis distance to calculate the deviation between the predicted state sequence and the baseline state. Compared to Euclidean distance, Mahalanobis distance considers the correlation between features and is more suitable for handling distance metrics in multi-dimensional feature spaces. Specifically, the baseline state distribution corresponding to the current operating condition is first extracted from the baseline database, and then the standardized distance of the predicted state sequence relative to this baseline distribution is calculated. The calculation formula is: D = sqrt((x-μ)^T Σ^(-1) (x-μ)), where x is the predicted state vector, μ is the baseline state mean vector, and Σ is the baseline state covariance matrix. The state deviation value obtained in this way can more accurately reflect the degree of difference between the current state and the normal state.

[0144] A3: Perform linear regression analysis on the state deviation values ​​to obtain the trend characteristics of deviation changes.

[0145] Linear regression analysis is performed on the state deviation value sequence using a sliding window mechanism, typically with a window size of 100 sampling points. The regression analysis employs the least squares method to calculate the regression coefficient β and intercept α. The sign and magnitude of the regression coefficient β reflect the trend of the deviation: a positive value indicates the fault may be worsening, while a negative value indicates the state may be improving; the absolute value of the coefficient reflects the rate of change. Simultaneously, the goodness of fit R² of the regression equation is calculated to assess the reliability of the trend prediction. When R² is below 0.6, the system reduces the window size and re-performs the regression analysis to improve the accuracy of the trend characteristics.

[0146] A4: Train the state transition matrix based on historical fault data, and combine the current state deviation value and the deviation change trend to predict the fault development trend.

[0147] The state transition matrix is ​​trained based on a large amount of historical fault data, and a Markov chain model is used to describe the transition probabilities between different fault states.

[0148] The training process first categorizes historical fault data according to severity, counts the frequency of transitions between different fault states, and then obtains a transition probability matrix through frequency normalization. For example, for a fault state i, the probability of transitioning to state j is Pij = Nij / Ni, where Nij is the number of samples transitioning from state i to state j, and Ni is the total number of samples in state i. Combining the currently observed state deviation values ​​and deviation trends, the state transition matrix can be used to predict future possible fault states and their probabilities.

[0149] A5: Based on the aforementioned state deviation value, deviation change trend, and fault development trend, the fuzzy comprehensive evaluation method is used to classify the operating status of CNC machine tools into four levels: normal, minor abnormality, moderate fault, and severe fault.

[0150] The fuzzy comprehensive evaluation method is based on three evaluation factors: state deviation value U1, deviation change trend U2, and fault development trend U3. First, a set of evaluation factors is established.

[0151] U={U1,U2,U3}

[0152] and rating scale

[0153] V = {Normal, Minor Abnormality, Moderate Fault, Severe Fault}.

[0154] Then, a fuzzy relation matrix R is constructed, where each element rij represents the membership degree of the i-th factor to the j-th level. The weight vector A is determined using the analytic hierarchy process (AHP), typically with larger weights for state deviation values ​​(0.4-0.5) and relatively smaller weights for deviation change trends and fault development trends (0.2-0.3). Finally, a comprehensive evaluation result is obtained through the fuzzy synthesis operation B=A○R, and the level with the highest membership degree is selected as the evaluation result for the current state.

[0155] In addition, the state predictor includes the following feedback correction mechanism:

[0156] An error tracker is used to calculate the sequence of deviations between predicted and actual observations.

[0157] An adaptive weight adjuster dynamically adjusts the correction weights within a range of 0.6 to 1.0 based on the prediction accuracy, where the prediction accuracy is calculated using the mean square error between the predicted and actual values.

[0158] The state corrector performs real-time correction on the prediction results based on the correction weights.

[0159] Specifically, the feedback correction mechanism of the state predictor comprises three core components. The error tracker calculates the deviation sequence between predicted and actual observations using a sliding window, typically set to 20 sampling points. This deviation sequence includes not only absolute error but also relative and trend errors; this error information is used to evaluate the accuracy of the prediction. The adaptive weight adjuster dynamically adjusts the correction weights based on the prediction accuracy.

[0160] ACC = 1 - RMSE / Range

[0161] Where RMSE is the root mean square error and Range is the range of observations. When ACC > 0.9, the correction weight is close to 1.0; when ACC < 0.7, the correction weight is close to 0.6. The state corrector then corrects the prediction results in real time according to the correction weight w. The corrected prediction value is equal to the weighted average of the original prediction value and the actual observation value.

[0162] x_corrected = w*x_predicted + (1-w)*x_observed.

[0163] This adaptive correction mechanism can dynamically adjust the prediction strategy according to changes in prediction performance, thereby improving the stability and reliability of predictions.

[0164] S5: Based on the fault prediction results, the Monte Carlo tree search algorithm is used to construct the fault evolution path and output the fault diagnosis results.

[0165] This step uses the Monte Carlo Tree Search (MCTS) algorithm to construct the evolution path of machine tool faults.

[0166] MCTS is a sampling-based search algorithm, particularly suitable for handling decision problems with large state spaces. In machine tool fault diagnosis, by repeatedly simulating different fault evolution possibilities, this algorithm can effectively identify the most likely fault paths. The search process includes four main steps: selection, expansion, simulation, and backtracking. Each iteration refines the evaluation of the fault evolution path.

[0167] S5 specifically includes:

[0168] S5.1: Based on the fault prediction results, establish an initial state node, wherein the attribute information of the initial state node represents the feature vector, fault risk level and confidence level of the current state of the CNC machine tool.

[0169] An initial state node is established based on the fault prediction results. This node contains three types of key attribute information: feature vector, fault risk level, and confidence level.

[0170] The feature vector consists of the machine tool's current monitoring indicators, including multi-dimensional data such as vibration characteristics (e.g., root mean square value, peak factor, spectral characteristics), temperature characteristics (e.g., temperature values ​​at key points and their rate of change), and current characteristics (e.g., RMS value, harmonic content). The fault risk level is divided into four levels (normal, minor, moderate, and severe), with specific numerical scores (0-100 points). The confidence level reflects the reliability of the judgment on the current state, and is determined through historical data verification and comprehensive evaluation of expert experience, with a value range of 0-1.

[0171] S5.2: Establish a state transition rule base based on this node. The state transition rule base includes: component-level fault transfer rules, system-level fault transfer rules, and environmental factor influence rules. Each rule is configured with a corresponding transition probability and time constraint.

[0172] The state transition rule base contains three categories of rules: component-level fault transition rules describe the fault evolution pattern within a single component, such as the evolution of a bearing from normal to micro-wear, moderate wear, and then to severe wear, with each stage having corresponding characteristic index ranges and transition conditions; system-level fault transition rules describe the fault association and propagation mechanism between multiple components, such as the association rule that abnormal spindle vibration leads to accelerated tool wear; and environmental factor influence rules consider the degree of influence of external factors such as temperature, humidity, and dust on fault evolution. Each rule is configured with a transition probability (based on historical data statistics) and time constraints (including the shortest transition time and the longest duration). For example, the transition probability of a bearing from micro-wear to moderate wear is 0.3, requiring a minimum of 24 hours and a maximum of 72 hours.

[0173] S5.3: Based on the state transition rule base, a heuristic method is used to dynamically determine the expansion depth, wherein a maximum expansion depth of 8 layers is used for important fault paths, and a maximum expansion depth of 4 layers is used for minor fault paths.

[0174] A heuristic method is used to dynamically determine the expansion depth of the search tree.

[0175] For critical failure paths that could lead to machine tool downtime or significant losses (such as spindle system failures, transmission system failures, etc.), a maximum expansion depth of 8 layers is used, with each layer representing a critical state transition point, enabling prediction of the failure evolution process over a longer period. For minor failure paths with relatively small impacts (such as minor lubrication system anomalies, cooling system efficiency declines, etc.), a maximum expansion depth of 4 layers is used to balance computational efficiency. The dynamic adjustment of the expansion depth is based on the current node's failure level (8 layers for severity scores greater than 70), risk assessment results (8 layers for potential losses exceeding 30% of the equipment value), and computational resource limitations.

[0176] S5.4: Use the improved UCT algorithm to perform node selection and expansion to determine the failure evolution path.

[0177] The improved UCT (Upper Confidence Bound for Trees) algorithm introduces a failure risk weight factor during node selection and expansion. Node selection comprehensively considers the number of visits, cumulative reward value, and risk weight, choosing the node with the highest score for expansion. Specifically, the node selection score is calculated as follows:

[0178] Score = Q / N + c*sqrt(ln(N_parent) / N) + w*R,

[0179] Where Q represents the cumulative value of a node, N represents the number of visits, c is the exploration coefficient (usually set to 1.414), w is the risk weight (dynamically adjusted according to the fault level, ranging from 0.1 to 0.5), and R is the node risk coefficient. The node expansion process strictly follows the rules in the state transition rule base, generates child nodes through probability sampling, and dynamically adjusts the transition probability based on historical data.

[0180] In one embodiment, S5.4 specifically includes:

[0181] B1. Calculate the node score: Score = Q / N + c*sqrt(ln(N_parent) / N) + w*R

[0182] Where Q represents the cumulative value of a node, N represents the number of times a node is visited, N_parent represents the number of times the parent node is visited, c represents the exploration coefficient, R represents the node risk coefficient, and w represents the risk weight.

[0183] Node scores are calculated using a modified UCB formula, comprising three key components: a value utilization term (Q / N) reflecting the average benefit of a node, an exploration term (csqrt(ln(N_parent) / N)) ensuring opportunities are given to underexplored nodes, and a risk weighting term (wR) reflecting the impact of fault severity. Specifically, the Q value is accumulated through node simulation and ranges from 0 to 100; the N value records the frequency of node visits; the exploration coefficient c is typically set between 0.7 and 1.4 and dynamically adjusted according to the search phase (larger in the early stages to promote exploration, and smaller in the later stages to enhance utilization); the node risk coefficient R is determined based on the fault level and ranges from 0 to 1; and the risk weight w is set according to the severity of the fault and is typically between 0.3 and 0.8.

[0184] B2: Generate child nodes based on the rules in the state transition rule base, and dynamically adjust the transition probability according to historical fault data during the expansion process.

[0185] The child node generation process strictly follows the rules in the state transition rule base, employing a probabilistic sampling method. First, based on the current node's state characteristics, all possible transition rules are selected from the rule base. Then, based on the transition probabilities of these rules, multiple candidate child nodes are generated. During the expansion process, the system dynamically adjusts the transition probabilities based on recent fault data. For example, if a certain fault mode has occurred frequently recently, the probability of the relevant transition rule will increase by 10%-30%. Furthermore, the system considers time constraints to ensure that the generated child node state transitions occur within a reasonable timeframe.

[0186] B3: Perform node simulations using a fast random simulation method, and record the cumulative risk value and time delay on the path;

[0187] Fast stochastic simulation employs a lightweight Monte Carlo method, starting from the current node and randomly selecting a successor state according to state transition rules until a termination condition is met (such as reaching the maximum depth or encountering a terminating state). During the simulation, the system records two key metrics: cumulative risk value (the weighted sum of risk coefficients for each node) and time delay (the total time required for state transitions). To improve simulation efficiency, a fast evaluation strategy is adopted, focusing only on the state transitions of critical nodes and simplifying calculations for intermediate states. Each node typically undergoes 100-200 simulations to obtain statistically reliable results.

[0188] B4: Starting from the leaf node, update the number of visits and cumulative value of each node upwards along the path.

[0189] The backpropagation process employs a bottom-up update strategy. Starting from the leaf nodes, the statistics of each node are updated upwards along the search path. The number of visits N increases simply, while the update of the cumulative value Q considers the node's risk level and time delay. The specific update formula is: Q_new = Q_old + (R_value * T_discount), where R_value is the simulated reward value, and T_discount is a discount factor based on time delay. This update mechanism ensures that the search tree gradually converges towards the most likely and most dangerous failure evolution direction.

[0190] This application also includes the following embodiments:

[0191] C1: Count the number of visits to each node in the fault evolution path, and identify the path with more than 30% of the total number of visits as a high-probability fault evolution path.

[0192] The system performs statistical analysis on all explored failure evolution paths, calculating the access frequency of each node. Paths with access frequencies exceeding 30% of the total simulations are identified as high-probability failure evolution paths. This threshold is selected based on extensive experimental data and effectively filters out the most likely failure evolution sequences. Simultaneously, the system also considers the path's risk level; paths with access frequencies close to the threshold (25%-30%) but higher risk levels are also included in the high-probability path set.

[0193] C2: Generate a hierarchical fault diagnosis report based on the high-probability fault evolution path, including a description of the current fault status, early warning of potential fault evolution, and corresponding maintenance suggestions.

[0194] The hierarchical fault diagnosis report adopts a three-layer structure: the first layer is a description of the current fault status, detailing abnormal indicators, fault levels, and confidence levels; the second layer is a potential fault evolution warning, predicting possible fault development over the next 24 hours, 72 hours, and one week based on high-probability fault paths; the third layer is maintenance recommendations, providing specific maintenance plans for different fault evolution scenarios, including suggested maintenance times, required spare parts, and operating procedures. The report uses a structured format, facilitating system parsing and human reading.

[0195] C3: The fault evolution tree structure is displayed in a visual way, with different colors used to indicate the fault level, node size representing the state transition probability, and the thickness of the connecting edges representing the path traversal frequency.

[0196] The fault evolution tree visualization employs a force-directed graph layout algorithm to ensure both spatial efficiency and visual appeal in node distribution. Nodes are color-coded to represent fault levels (green for normal, yellow for minor faults, orange for moderate faults, and red for severe faults), and node size is proportional to the state transition probability (the higher the probability, the larger the node). The thickness of connecting edges represents the path traversal frequency; the more frequent the traversal, the thicker the edge. Furthermore, the system provides interactive functionality, allowing users to click on nodes to view detailed information or adjust the view layout by dragging and dropping.

Claims

1. An intelligent fault diagnosis method of a numerical control machine tool, applied to a control system of the numerical control machine tool, characterized in that the method comprises the steps of: The method comprises the following steps: ​ Collecting multi-source sensor data during the operation of the numerical control machine tool, wherein the multi-source sensor data comprises at least two of vibration signals, temperature signals, current signals and sound signals; Preprocessing and feature extraction are performed on the multi-source sensor data to obtain preprocessed data; Based on the preprocessed data, a machine tool dynamic model is constructed and trained to obtain a target machine tool dynamic model; When an abnormality of the numerical control machine tool is detected, a fault state estimation is performed based on the target machine tool dynamic model and a model predictive control method to obtain a fault estimation result; Based on the fault estimation result, a Monte Carlo tree search algorithm is used to construct a fault evolution path, and a fault diagnosis result is outputted; The method comprises the following steps: A deep learning model architecture is constructed, which comprises a convolutional neural network (CNN) for spatial feature extraction, a bidirectional long short-term memory network (LSTM) for time series feature processing, and a fully connected layer for fusing the features extracted by the CNN and the LSTM; Based on the outputs of the CNN and the LSTM, a plurality of loss functions are designed, which comprise a prediction error loss for measuring prediction accuracy and an L2 regularization constraint for preventing model overfitting, wherein different loss weights are set for different types of fault samples; According to the calculation results of the loss functions, an Adam optimizer is used for model training, and the learning rate is dynamically adjusted based on the performance indicators of the validation set. When the validation set loss value does not decrease for three consecutive training periods, the learning rate is adjusted to 0.8 times the original value; A K-fold cross-validation based model evaluation mechanism is used to determine the target machine tool dynamic model, which comprises the following steps: in each fold training, a stratified sampling strategy is used to divide the data set to ensure the balanced distribution of different fault types, and the accuracy, precision and recall rates on each validation set are calculated to evaluate the model performance and select the optimal model parameter configuration; The method comprises the following steps: According to the fault estimation result, an initial state node is established, wherein the attribute information of the initial state node represents the feature vector of the current state of the numerical control machine tool, the fault risk level and the confidence level; Based on the node, a state transition rule library is established, which comprises component-level fault transition rules, system-level fault transition rules and environmental factor influence rules, wherein each rule is configured with a corresponding transition probability and a time constraint; Based on the state transition rule library, a heuristic method is used to dynamically determine the expansion depth, wherein the maximum expansion depth of an important fault path is 8 layers, and the maximum expansion depth of a secondary fault path is 4 layers; An improved UCT algorithm is used for node selection and expansion to determine the fault evolution path; The method comprises the following steps: Calculate the node score: Score = Q / N + c sqrt(ln(N_parent) / N) + w R, Wherein Q represents the node cumulative value, N represents the node access frequency, N_parent represents the parent node access frequency, c represents the exploration coefficient, R represents the node risk coefficient, and w represents the risk weight; Generate a sub-node based on the rules in the state transition rule library, and dynamically adjust the transition probability during the expansion process according to historical fault data; Perform node simulation using a fast random simulation method, and record the cumulative risk value and time delay on the path; From the leaf node, update the access frequency and cumulative value of each node along the path.

2. The diagnostic method according to claim 1, characterized in that, In the deep learning model architecture: The CNN includes a three-layer convolutional layer structure connected in turn, the first layer uses 32 3*3 convolutional kernels for feature extraction and is normalized by a BatchNormalization layer, the second layer uses 64 5*5 convolutional kernels to expand the receptive field and enhances the non-linear expression ability through a ReLU activation function, and the third layer reduces the dimension through a 2*2 maximum pooling window; The LSTM includes two layers of bidirectional LSTM layers, each layer contains 128 neurons, and a Dropout layer with a dropout rate of 0.3 is set between the two layers; The fully connected layer includes three layers, and the corresponding node numbers are 512, 256 and 128 in turn, for fusing spatial features and time sequence features, and a BatchNormalization layer is added before the last layer to ensure stable feature distribution.

3. The diagnostic method according to claim 1, characterized in that, When the abnormal operation of the numerical control machine tool is detected, the fault state is estimated based on the target machine tool dynamic model and the model predictive control method to obtain a fault estimation result, including: Based on the target machine tool dynamic model, a state predictor of recursive prediction structure is constructed, and the state predictor uses 20 sampling points as the time window length of a prediction unit; Through the state predictor, an adaptive rolling horizon prediction strategy is designed, which includes short-term prediction of 3 sampling points, medium-term prediction of 6 sampling points and long-term prediction of 10 sampling points, and dynamically adjusts the prediction step according to the working condition of the numerical control machine tool; Based on the rolling horizon prediction strategy, the future state of the numerical control machine tool is predicted, the output result of the prediction strategy is obtained, and the deviation between the predicted state and the normal operation state of the numerical control machine tool is calculated in real time to evaluate the fault risk level.

4. The diagnostic method according to claim 3, characterized in that, The real-time calculation of the deviation between the predicted state and the normal operation state of the numerical control machine tool includes: A numerical control machine tool health state benchmark library containing normal operation feature distributions under different cutting parameters is constructed; The Mahalanobis distance between the predicted state sequence and the corresponding working condition benchmark state in the benchmark library is calculated to obtain a state deviation value; Linear regression analysis is performed on the state deviation value to obtain a deviation trend feature; A state transition matrix is trained based on historical fault data, and the fault development trend is predicted based on the current state deviation value and the deviation trend; According to the state deviation value, the deviation trend and the fault development trend, a fuzzy comprehensive evaluation method is used to divide the numerical control machine tool operation state into four levels: normal, slight abnormality, moderate fault and serious fault.

5. The diagnostic method according to claim 3, characterized in that, The state predictor includes the following feedback correction mechanism: an error tracker for calculating a deviation sequence between the predicted value and the actual observed value; an adaptive weight adjuster for dynamically adjusting the correction weight in the range of 0.6 to 1.0 according to the prediction accuracy, wherein the prediction accuracy is calculated by the mean square error of the predicted value and the actual value; a state corrector for real-time correction of the prediction result based on the correction weight.

6. The diagnostic method according to claim 1, characterized in that, At least one of the following operations is also performed: counting the access times of each node in the failure evolution path, and identifying a path with an access time exceeding 30% of the total simulation times as a high-probability failure evolution path; generating a hierarchical failure diagnosis report based on the high-probability failure evolution path, including a current failure state description, a potential failure evolution warning, and corresponding maintenance suggestions; adopting a visual method to display the failure evolution tree structure, using different colors to identify the failure level, using the node size to represent the state transition probability, and using the thickness of the connection edge to represent the path access frequency.

Images