AI fault prediction and diagnosis system and method based on numerical control machine tool

Abstract

The invention discloses an AI fault prediction and diagnosis system and method based on a numerical control machine tool, and belongs to the technical field of fault diagnosis. The technical problem that efficient fault early warning and health management cannot be implemented in the full life cycle of a numerical control machine tool in an existing scheme is solved. Monitoring data covering the full life cycle and multiple working conditions of the numerical control machine tool, and providing high-quality annotation data for subsequent model training through association of a high-dimensional feature matrix and a fault tag; an improved variational mode decomposition algorithm is utilized, fault information of each mode is quantified in combination with wavelet packet energy entropy, time migration of multi-sensor data is eliminated through space-time alignment, weights of different features are adaptively distributed by utilizing an attention mechanism, and discrimination of fusion feature vectors is effectively enhanced; a long-term dependency relationship of a feature sequence is captured based on a bidirectional gating circulation unit, a convolutional neural network is improved to reinforce local detail features, and the fitting capability of a model to a complex fault mode can be effectively improved through cooperation of the two.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of fault diagnosis, in particular to an AI fault prediction and diagnosis method based on a numerical control machine tool. BACKGROUND

[0002] The AI fault prediction and diagnosis of the numerical control machine tool refers to a process of detecting, diagnosing and predicting the faults of the numerical control machine tool by using artificial intelligence (AI) technology. This process identifies the working state of the equipment by collecting and analyzing various data from the machine tool, such as vibration signals, temperature, current waveform, etc., and predicts possible faults.

[0003] The existing AI fault prediction and diagnosis scheme of the numerical control machine tool mostly relies on a single sensor to collect data when implemented, ignores the multi-physical field coupling characteristics, and causes the early fault signals to be submerged in noise. The empirical mode decomposition is commonly used for non-stationary vibration signal decomposition, which is prone to modal aliasing, and the fault sensitive feature extraction accuracy is insufficient. In addition, the traditional method uses simple splicing or weighted average to fuse multi-source data, which cannot dynamically adapt to the importance of features under different working conditions, resulting in poor discriminability of the fused features. The single deep learning model of the existing technology cannot balance the time sequence dependence and local impact features, so the model accuracy under varying conditions is not good.

[0004] Therefore, there is an urgent need for a fault prediction and diagnosis method that integrates multi-modal fusion, intelligent feature extraction and multi-algorithm fusion to realize efficient fault early warning and health management of the numerical control machine tool throughout its life cycle. SUMMARY

[0005] The purpose of the present application is to provide an AI fault prediction and diagnosis method based on a numerical control machine tool, which solves the technical problem that the existing scheme cannot implement efficient fault early warning and health management of the numerical control machine tool throughout its life cycle.

[0006] The purpose of the present application can be achieved by the following technical solutions: An AI fault prediction and diagnosis method based on a numerical control machine tool, comprising: S1: monitoring the running state of the numerical control machine tool throughout its life cycle under different working conditions, collecting multi-source sensor data and fault record data of the machine tool, extracting feature parameters of different time sequences in the sensor data and constructing a high-dimensional feature matrix X(t), labeling the fault record data as fault type labels Y(t), and combining to form a labeled diagnosis data set; S2: decomposing the non-stationary vibration signal in the high-dimensional feature matrix X(t) by using an improved variational modal decomposition algorithm, obtaining a plurality of modal components IMF(i), and calculating the fault sensitive features of each component in combination with the wavelet packet energy entropy; S3: Temporally and spatially align the current, temperature, and main shaft speed features in the multi-source sensor data with the modal component features, and dynamically weight and fuse them into a fusion feature vector F(t) through an attention mechanism; S4: Take the fusion feature vector F(t) as input and the fault type label Y(t) as output, and construct a hybrid diagnostic model based on a bidirectional gated recurrent unit and an improved convolutional neural network, use the bidirectional gated recurrent unit to capture the temporal dependence, and use the improved convolutional neural network to extract local features, and output a preliminary fault prediction probability; S5: Real-time acquisition and preprocessing of multi-source sensor data in the operation of the numerical control machine tool, input of the preprocessed data into the constructed fault diagnosis model, and output of the fault diagnosis result.

[0007] Preferably, the feature vectors of consecutive time steps are intercepted with a sliding window to construct a high-dimensional feature matrix: ; wherein t is the time sample; is a single time step; is a feature vector; the window contains =100 time steps, each column corresponds to a feature dimension, a total of 18 dimensions, the number of rows of the matrix is the number of time steps in the window, and the number of columns is the feature dimension; When the matrix dimension is calibrated, if the total monitoring time is T, the total number of feature matrix samples N=(T-W) / S+1, and the final high-dimensional feature matrix dimension is , S is the sliding window step, and W is the sliding window length.

[0008] Preferably, when configuring the improved variational modal decomposition VMD parameters, first determine the adaptive number of decomposition layers, and dynamically optimize the number of decomposition layers K based on the kurtosis maximization criterion; configure the penalty factor a: ; wherein Energy() represents the modal energy concentration, and the optimization goal is to maximize the energy separation degree of each IMF component.

[0009] Preferably, a time-varying scaling term g(t) is introduced into the improved variational modal decomposition VMD objective function: ; wherein is a set of modal components to be decomposed; is a set of center frequencies of each modal component; i is the modal component index; is a time partial derivative operator; is a Dirac delta function; is an imaginary unit; * is a convolution operator; is a time-domain signal of the i-th modal component; is a frequency modulation term; is a time-varying scaling factor; is the square of the Frobenius norm.

[0010] Preferably, when performing modal component decomposition and screening, the central frequency is initialized as , the modal component ; is updated iteratively by the alternating direction multiplier method: ; wherein, is the (n+1)th iteration result of the i th modal component; is the minimum value search operator of ; is the time domain signal of the i th modal component; is the Lagrange multiplier; x is the original input signal; is the superposition of all modal components.

[0011] Preferably, when constructing the energy entropy and sensitive features, the energy of each sub-band is calculated as ; wherein, is the wavelet packet sub-band index; is the wavelet packet coefficient; the normalized energy distribution ; The wavelet packet energy entropy is calculated as: ; The fault sensitive feature vector is constructed as: Fsen = [H1, H2,..., Hm], m is the number of effective IMF components.

[0012] Preferably, the sensor feature matrix , the modal feature matrix are spliced by feature dimension: ; wherein, the row is the feature type; the column is the time sample; The spliced is subjected to maximum-minimum normalization processing, and the joint feature matrix is output.

[0013] Preferably, when constructing the hybrid diagnostic model, the training set , the validation set are input to the pre-constructed hybrid model structure; when performing feature fusion and classification head, the bidirectional feature splicing is compressed by global maximum pooling to ; The splicing and obtain the fusion feature .

[0014] Preferably, the loss function is configured as: ; wherein, L is the batch average loss value; B is the batch size; a one-hot encoding matrix of true labels; a model prediction probability, representing a probability that the bth sample is predicted to be the mth class.

[0015] An AI fault prediction and diagnosis system based on a numerical control machine tool, comprising: a diagnosis data processing module: monitoring the running state of the numerical control machine tool in different working conditions throughout its life cycle, collecting multi-source sensor data and fault record data of the machine tool, extracting feature parameters of different time sequences in the sensor data and constructing a high-dimensional feature matrix X(t), labeling the fault record data as fault type labels Y(t), and combining to form a labeled diagnosis data set; a fault sensitive processing module: decomposing non-stationary vibration signals in the high-dimensional feature matrix X(t) through an improved variational mode decomposition algorithm, obtaining multiple modal components IMF(i), and calculating fault sensitive features of each component in combination with wavelet packet energy entropy; a fusion feature vector module: time and space aligning the current, temperature, and spindle speed features in the multi-source sensor data with the modal component features, and dynamically weighting and fusing them into a fusion feature vector F(t) through an attention mechanism; a hybrid diagnosis model construction module: taking the fusion feature vector F(t) as input and the fault type label Y(t) as output, constructing a hybrid diagnosis model based on a bidirectional gate recurrent unit and an improved convolutional neural network, capturing time sequence dependencies using the bidirectional gate recurrent unit, extracting local features through the improved convolutional neural network, and outputting preliminary fault prediction probabilities; a fault prediction and diagnosis module: real-time collection and preprocessing of multi-source sensor data during the operation of the numerical control machine tool, input of the preprocessed data into the constructed fault diagnosis model, and output of fault diagnosis results.

[0016] Compared with the prior art, the present application has the following advantages: The present application can ensure the diversity of data by covering the full life cycle and multi-condition data monitoring of the numerical control machine tool, provide high-quality labeled data for subsequent model training through the association of the high-dimensional feature matrix and the fault label, and effectively improve the generalization ability of the diagnosis model; the improved variational mode decomposition algorithm can solve the traditional VMD modal aliasing problem, the fault information of each mode can be quantified by combining the wavelet packet energy entropy, the sensitivity of the early weak fault feature extraction can be enhanced, and the timeliness of fault recognition can be effectively improved; the time offset of multi-sensor data is eliminated through time and space alignment, the weights of different features are adaptively allocated through the attention mechanism, the redundant information interference can be reduced, and the discriminability of the fusion feature vector is effectively enhanced; the long-term dependence of the feature sequence is captured through the bidirectional gate recurrent unit, and the local detail features are strengthened through the improved convolutional neural network, which can effectively improve the fitting ability of the model to complex fault patterns and the accuracy of the preliminary prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described below with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of the steps for implementing an AI fault prediction and diagnosis method based on CNC machine tools in the present invention.

[0019] Figure 2 This is a module block diagram of an AI fault prediction and diagnosis system based on CNC machine tools in the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] Example 1, as Figure 1 As shown, the present invention is an AI fault prediction and diagnosis method based on CNC machine tools, comprising: S1: Monitor the operating status of CNC machine tools throughout their life cycle under different working conditions. Collect multi-source sensor data and fault record data from the machine tools. Extract characteristic parameters of different time series in the sensor data and construct a high-dimensional feature matrix X(t). Label the fault record data with fault type labels Y(t) to form a labeled diagnostic dataset. The specific steps include: When designing working conditions and developing monitoring plans, the CNC machine tool is monitored over its entire life cycle under three machining loads, four tool wear stages, and three ambient temperature combinations. Among them, there are three types of processing loads: no load, light load 50% rated power, and heavy load 80% rated power; Four stages of tool wear: new tool, slightly worn, moderately worn, and severely worn; Three ambient temperatures: 20±2℃, 35±2℃, 50±2℃; Under each working condition combination, the machine tool continuously operates to process typical workpieces, such as thin-walled aviation aluminum alloy parts. The single monitoring time is no less than 72 hours, and the cumulative monitoring period covers the complete life cycle of the tool from new to scrap, for example, no less than 500 hours; During the monitoring process, the cutting parameters were set to an orthogonal experimental combination of feed rate 50-200 mm / min, spindle speed 3000-8000 r / min, and cutting depth 0.1-1 mm; When collecting data through a distributed sensor array, install three-axis vibration acceleration sensors on the spindle box, feed guide, and servo motor to collect X, Y, and Z direction vibration signals; the sampling frequency is 10 kHz, and the range is ±50 g; Install a Hall current sensor on the power supply end of the servo motor to collect three-phase current signals; the sampling frequency is 5 kHz, and the accuracy is 0.1% FS; Install thermocouple temperature sensors at the spindle bearing seat and guide block; the sampling frequency is 1 kHz, and the measurement range is -40~150℃; Read the internal parameters of the numerical control system in real time through the machine tool PLC, including but not limited to spindle load rate, feed axis position deviation, and hydraulic system pressure; Synchronize all sensor data to the edge computing terminal through industrial Ethernet with a timestamp synchronization accuracy of ≤1 ms; When preprocessing the original sensor data, use the wavelet threshold denoising algorithm to eliminate cutting noise in the vibration signal, and use a 50 Hz notch filter to remove power frequency interference in the current signal; When extracting feature parameters from the processed time domain signal, extract 12 time domain features including root mean square value, peak factor, kurtosis, waveform factor, pulse factor, margin factor, variance, maximum value, minimum value, temperature change rate, current effective value, and energy value; Extract the frequency spectrum features in the 500-5000 Hz frequency band through fast Fourier transform (FFT), including FFT of 1024-point time domain signal, to obtain the frequency spectrum X(k) with frequency resolution Δf=fs / N=10kHz / 1024≈9.766Hz, where fs is the sampling frequency, k is the frequency point index, and k=0,1,...,N / 2; Extract 3 maximum peak frequencies fp1, fp2, and fp3 and their corresponding amplitudes Ap1, Ap2, and Ap3 in the 500-5000 Hz frequency band; Use db4 wavelet packet to decompose the vibration signal into 8 sub-bands, and calculate the energy proportion of each sub-band , Calculate the energy entropy H for the jth sub-band energy ; Combine the 12 time domain features and 6 frequency domain features of a single time step into a feature vector ; the frequency domain features include 3 peak frequencies, 3 amplitudes, and 1 energy entropy; Use a sliding window to extract the feature vectors of consecutive time steps, with a window size W=10s and a step size S=5s, to construct a high-dimensional feature matrix: ; where t is the time sample; the window contains =100 time steps, each column corresponds to a feature dimension, a total of 18 dimensions, the number of rows of the matrix is the number of time steps within the window, and the number of columns is the feature dimension.

[0022] When the matrix dimension is calibrated, if the total monitoring time is T, the unit is s, the total number of feature matrix samples N=(T-W) / S+1, and the final high-dimensional feature matrix dimension is , S is the sliding window step; W is the sliding window length; wherein, the three-dimensional tensor, the first dimension is the number of samples, the second dimension is the time step, and the third dimension is the feature dimension; When performing fault label labeling and dataset division, fault record data is obtained in three ways: (1) Fault codes recorded by the alarm system of the machine tool, such as 3005 spindle overload; (2) Hidden faults found by manual inspection, such as gear box abnormal noise and guide rail crawling; (3) Artificially implanted faults in the accelerated aging experiment, such as bearing inner ring crack and screw nut gap too large; The fault types are divided into 12 typical faults, such as bearing wear, motor demagnetization, and guide rail lubrication deficiency. Each fault sample is labeled with the fault occurrence time t0, fault level and maintenance record. The fault level includes slight, moderate and severe. The division rule of the fault level can be customized according to the application requirements of the actual application scene; The fault label is defined as a 12-dimensional vector Y(t) = [y1, y2,..., y12], wherein yk represents the kth fault, and the fault occurrence time is accurately labeled by cross-correlation analysis; The sliding window method is used to align the feature matrix X(t) and the fault label Y(t) to form a labeled time series dataset, and the dataset is divided into training set, validation set and test set according to 7:2:1. The window size is 10s, and the step size is 5s; Abnormal samples are removed by the Laplace criterion (3σ rule), and missing data is completed by using the time series interpolation method based on LSTM; To solve the problem of unbalanced fault samples, the SMOTE-ENN algorithm is used to oversample the minority class fault samples, so that the difference between the number of samples of each class is ≤10%; The final diagnostic dataset includes feature matrix X(t), fault label Y(t) and corresponding working condition parameter metadata. The working condition parameter metadata includes but is not limited to load and temperature.

[0023] In the embodiment of the application, through data monitoring covering the whole life cycle and multiple working conditions of the numerical control machine tool, the diversity of the data can be ensured, the association of the high-dimensional feature matrix and the fault label can be provided for subsequent model training, and high-quality labeled data can be provided, so that the generalization ability of the diagnostic model can be effectively improved.

[0024] S2: The non-stationary vibration signal in the high-dimensional feature matrix X(t) is decomposed by the improved variational modal decomposition algorithm, a plurality of modal components IMF(i) are obtained, and the fault sensitive features of each component are calculated in combination with the wavelet packet energy entropy; the specific steps include: The vibration signal components in the high-dimensional feature matrix X(t) are preprocessed by adaptive wavelet threshold denoising and empirical mode decomposition EMD detrending; both are existing conventional technical solutions, and the specific implementation steps will not be described here; The preprocessed vibration signal is segmented according to a single time step Each segment contains 1024 sampling points, and is uniformly resampled to 10 kHz by linear interpolation to ensure time domain resolution consistency; When configuring the improved variational modal decomposition VMD parameters, first determine the adaptive decomposition layer number, and dynamically optimize the decomposition layer number K based on the kurtosis maximization criterion: Initialize K=2, calculate the kurtosis value of each modal component ; When <0.1, stop iteration, and finally K takes a value range of 3~8; Configure the penalty factor a: ; wherein Energy() represents the modal energy concentration degree, and the optimization goal is to maximize the energy separation degree of each IMF component; Introduce a time-varying scaling term γ(t) into the improved variational modal decomposition VMD objective function: ; wherein is the set of modal components to be decomposed; is the set of center frequencies of each modal component; i is the modal component index; is the time partial derivative operator; is the Dirac delta function; is the imaginary unit; * is the convolution operator; is the time domain signal of the i-th modal component; is the frequency modulation term; is the time-varying scaling factor; is the square of the Frobenius norm; When performing modal component decomposition and screening, initialize the center frequency , the modal component ; Update by the alternating direction multiplier method: ; wherein is the (n+1)th iteration result of the i-th modal component; The minimum search operator of ; is the time-domain signal of the i-th modal component; is the Lagrange multiplier; x is the original input signal; is the superposition of all modal components; Iteration termination condition: ; Decomposition to obtain K modal components IMF(i); i = 1, 2,..., K; K is the total number of decomposition layers; Filtering fault-related IMF based on correlation coefficient threshold: ; wherein, is the Pearson correlation coefficient; is the covariance; is the standard deviation of the original signal x(t); is the standard deviation of the i-th IMF; Retain components with >0.3, and eliminate noise-dominant IMFs; usually high-frequency components i>K / 2; When performing wavelet packet energy entropy calculation, perform 3-layer wavelet packet decomposition on each filtered IMF component: Wavelet basis function: db4; wherein, supporting the transient feature extraction of non-stationary signals; Decomposition layers: 3 layers; wherein, 8 sub-bands are obtained, and the frequency resolution Δf = 625 Hz; Reconstruction sub-band range: for the typical fault frequency of a numerical control machine tool, 200-5000 Hz; When constructing energy entropy and sensitive features, calculate the energy of each sub-band ; wherein, is the wavelet packet sub-band index; is the wavelet packet coefficient; Normalize the energy distribution ; Calculate the wavelet packet energy entropy: ; Construct the fault sensitive feature vector: Fsen = [H1, H2,..., Hm], m is the number of effective IMF components; When performing feature saliency evaluation and dimensionality reduction, the relevance of the features and the fault labels is evaluated by mutual information MI: ; wherein, Y is a set of fault labels; y is a single label in Y; h is a single feature value in ; is the joint probability distribution; is the marginal probability distribution of ; is the marginal probability distribution of the label Y; is a log-likelihood ratio; Reserve the top 80% of the features before the MI value, and eliminate redundant components, that is, the features with MI <0.1 are eliminated; The screened feature vector is subjected to fractional order differential processing: ; wherein, is a ν-order fractional derivative; ν is the order of the fractional order, the default value is 0.5, and Γ() is a Gamma function; the dynamic change rate of the feature is enhanced to improve the fault recognition degree; The first-order integer-order derivative operator of time t; is an integral variable; is a fractional integral kernel function; is the value of the i-th feature set at the historical time ; is an integral infinitesimal.

[0025] The empirical mode decomposition is commonly used for non-stationary vibration signal decomposition, and modal aliasing is prone to occur, so the fault sensitive feature extraction accuracy is insufficient; in the embodiment of the application, the improved variational mode decomposition algorithm can solve the problem of modal aliasing of the traditional VMD, and by combining the wavelet packet energy entropy to quantify the fault information of each mode, the sensitive feature extraction capability for early weak faults can be enhanced, and the timeliness of fault recognition can be effectively improved.

[0026] S3: Align the current, temperature, main shaft speed features in the multi-source sensor data with the modal component features in time and space, and dynamically weight and fuse them into a fusion feature vector F(t) through an attention mechanism; the specific steps include: The current, temperature, and main shaft speed features in the multi-source sensor data are subjected to Z-score standardization: ; wherein, is a sensor feature matrix, 3 sensors, and N is a sampling point; , is the mean and standard deviation of the sensor data; The modal component features are subjected to the same standardization processing to obtain a modal feature matrix , and K is the number of IMF components; The original timestamps of , , The original timestamps of ; because there is a time delay in VMD decomposition, ≠ ; The main shaft encoder pulse signal is taken as a reference clock to convert and into angle domain timestamps θ=2πn(t)t; n(t) is the instantaneous main shaft speed; using linear interpolation method to resample to the time grid of , so that ; where the resampling interval , the synchronization error <1 ms; is the sampling frequency; output the time-aligned sensor feature matrix , modal feature matrix ; Splice the sensor feature matrix , modal feature matrix by feature dimension: ; Wherein, the row is the feature type; Column, for time samples; Maximum-minimum normalization processing is performed on the spliced ; Obtain the fault label set Y, and calculate the feature priority based on mutual information ; For each time sample t, calculate the local response by the softmax function ; Wherein, W is a learnable parameter; d is the feature dimension; fuse the weights by the formula , and output the time-varying attention weight matrix ; When performing weighted fusion, ; Sum along the feature dimension , compressed into a single-channel fusion feature, and output the fusion feature vector .

[0027] The traditional method adopts simple splicing or weighted average to fuse multi-source data, which cannot dynamically adapt the importance of features under different working conditions, resulting in poor discriminability of the fusion features; in the embodiment of the application, the time offset of the multi-sensor data is eliminated by spatio-temporal alignment, and the weights of different features are adaptively allocated by using the attention mechanism, which can reduce the interference of redundant information and effectively enhance the discriminability of the fusion feature vector.

[0028] S4: Take the fusion feature vector F(t) as input and the fault type label Y(t) as output, construct a hybrid diagnostic model based on bidirectional gate recurrent unit and improved convolutional neural network, use bidirectional gate recurrent unit to capture time sequence dependence, extract local features through improved convolutional neural network, and output preliminary fault prediction probability; The specific steps include: When constructing the hybrid diagnostic model, the training set , the validation set ​Input into the pre-built hybrid model structure; The hybrid model structure is a series structure of a bidirectional gated recurrent unit (Bi-GRU) and an improved convolutional neural network (CNN), and includes a Bi-GRU time sequence feature extraction module: Number of layers: 2 layers (forward GRU + reverse GRU), number of hidden units H = 64, dropout rate = 0.3; Output: bidirectional feature splicing ; 2H = 128, due to bidirectional splicing; Improved CNN local feature extraction module: Convolutional layer 1: 3x3 convolutional kernel, number 32; step 1; padding "same"; activation function ReLU; Attention layer: channel attention module (CAM), generating weights for 32 channels Weighted output ; Convolutional layer 2: 3x3 convolutional kernel, number 64; step 1, padding "same"; activation function LeakyReLU; Global pooling: for Perform global average pooling, output ; When performing feature fusion and classification head, splice the bidirectional feature through global maximum pooling to ; Splice and , to get the fusion feature ; Fully connected layer: the output layer uses Softmax activation to get the prediction probability ; Configure the loss function: ; where L is the batch average loss value; B is the batch size; is the one-hot encoding matrix of the true label; is the model prediction probability, indicating the probability of the b-th sample being predicted as the m-th class; And configure the optimizer: AdamW, weight decay λ = 1e-4; initial learning rate η = 1e-3; use cosine annealing scheduling, decay to 0.5 times the current value every 5 epochs; Use early stopping method for training, training rounds = 50, batch size = 32, regularization coefficient = 1e-4; Output the trained hybrid diagnostic model; Input the test set into the trained hybrid diagnostic model, and output the preliminary fault prediction probability.

[0029] It should be noted that the bidirectional gated recurrent unit can capture the long-term temporal dependencies of fused features, such as fault development trends, while the improved convolutional neural network enhances key local features, such as the peak position of the shock pulse, through the attention mechanism. The two complement each other and can effectively improve the model's ability to recognize complex fault patterns. The choice of loss function and optimizer can effectively mitigate the risk of overfitting for small sample fault data. In industrial scenarios, the proportion of fault samples is usually less than 5%.

[0030] The existing single deep learning model is difficult to take into account both timing dependency and local impact characteristics, resulting in poor accuracy of the model under variable working conditions. In the embodiment of the present invention, by capturing the long-term dependency of feature sequences based on a bidirectional gated recurrent unit and improving the convolutional neural network to enhance local detail features, the two work together to effectively improve the model's ability to fit complex fault modes and improve the accuracy of preliminary predictions.

[0031] S5: Real-time collection and preprocessing of multi-source sensor data during the operation of the CNC machine tool, input of the preprocessed data into the constructed fault diagnosis model, and output of the fault diagnosis results.

[0032] Among them, the preprocessing of multi-source sensor data collected in real time during the operation of the CNC machine tool is processed through steps S2-S3, and the target fusion feature vector is output; The target fusion feature vector is input into the constructed fault diagnosis model, and the fault probability vector and fault type label are output.

[0033] In the embodiment of the present invention, an end-to-end real-time processing link is used to meet the real-time requirements of the industrial site, automatically output the fault probability vector and fault type label, and realize the transition from passive maintenance to active early warning, which can effectively reduce the loss of machine tool downtime.

[0034] Example 2, as Figure 2 As shown, an AI fault prediction and diagnosis system based on CNC machine tools includes: Diagnostic data processing module: This module monitors the operating status of CNC machine tools throughout their life cycle under different operating conditions, collects multi-source sensor data and fault record data from the machine tools, extracts characteristic parameters of different time series in the sensor data, constructs a high-dimensional feature matrix X(t), and labels the fault record data with fault type labels Y(t), combining them to form a labeled diagnostic dataset. Fault-sensitive processing module: Decomposes the non-stationary vibration signal in the high-dimensional feature matrix X(t) through an improved variational modal decomposition algorithm to obtain multiple modal components IMF(i), and calculates the fault-sensitive characteristics of each component in combination with wavelet packet energy entropy; Fusion feature vector module: align the current, temperature, main shaft speed features in the multi-source sensor data and the modal component features in time and space, and dynamically weight and fuse them into a fusion feature vector F(t) through an attention mechanism; Hybrid diagnostic model construction module: taking the fusion feature vector F(t) as input and the fault type label Y(t) as output, a hybrid diagnostic model based on a bidirectional gated recurrent unit and an improved convolutional neural network is constructed, the bidirectional gated recurrent unit is used to capture the time sequence dependence, and the improved convolutional neural network is used to extract local features, and a preliminary fault prediction probability is output; Fault prediction and diagnosis module: real-time acquisition and preprocessing of multi-source sensor data in the operation of the numerical control machine tool, input of the preprocessed data into the constructed fault diagnosis model, and output of a fault diagnosis result.

[0035] In several embodiments provided by the present application, it should be understood that the disclosed system can be implemented in other ways. For example, the above-described embodiments of the application are only illustrative, for example, the division of the modules is only a logical function division, and actual implementation can have another division mode.

[0036] The modules described as separate components can or can not be physically separated, and the components shown as modules can or can not be physical modules, which can be located in one place or distributed on multiple network modules. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs.

[0037] In addition, the function modules in each embodiment of the present application can be integrated in one processing module, or each module can exist physically, or two or more modules can be integrated in one module. The above integrated module can be realized in the form of hardware or hardware plus software function module.

[0038] It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the essential characteristics of the present application.

[0039] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and are not limited, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An AI fault prediction and diagnosis method based on CNC machine tools, characterized in that: include: S1: Monitor the operating status of CNC machine tools throughout their life cycle under different working conditions. Collect multi-source sensor data and fault record data from the machine tools. Extract characteristic parameters of different time series in the sensor data and construct a high-dimensional feature matrix X(t). Label the fault record data with fault type labels Y(t) to form a labeled diagnostic dataset. S2: Decompose the non-stationary vibration signal in the high-dimensional feature matrix X(t) using an improved variational modal decomposition algorithm to obtain multiple modal components IMF(i), and calculate the fault sensitivity characteristics of each component in combination with wavelet packet energy entropy; S3: Align the current, temperature, and spindle speed features in the multi-source sensor data with the modal component features in time and space, and dynamically weight them through the attention mechanism to form a fused feature vector F(t); S4: Taking the fused feature vector F(t) as input and the fault type label Y(t) as output, a hybrid diagnosis model based on a bidirectional gated recurrent unit and an improved convolutional neural network is constructed. The bidirectional gated recurrent unit is used to capture temporal dependencies, and the improved convolutional neural network is used to extract local features and output a preliminary fault prediction probability. S5: Real-time collection and preprocessing of multi-source sensor data during the operation of the CNC machine tool, input of the preprocessed data into the constructed fault diagnosis model, and output of the fault diagnosis results.

2. The AI ​​fault prediction and diagnosis method based on CNC machine tools according to claim 1 is characterized in that: Use a sliding window to intercept the feature vectors of consecutive time steps and construct a high-dimensional feature matrix: ; Where t is the time sample; is a single time step; is the feature vector; the window contains = 100 time steps, each column corresponds to a feature dimension, a total of 18 dimensions, the number of matrix rows is the number of time steps in the window, and the number of columns is the feature dimension; When calibrating the matrix dimension, if the total monitoring time is T, the total number of samples in the feature matrix is ​​N=(TW) / S+1, and the final high-dimensional feature matrix dimension is , S is the sliding window step size; W is the sliding window length.

3. The AI ​​fault prediction and diagnosis method based on CNC machine tools according to claim 2 is characterized in that: When configuring the improved variational mode decomposition (VMD) parameters, first determine the number of adaptive decomposition layers, and then dynamically optimize the number of decomposition layers K based on the kurtosis maximization criterion; configure the penalty factor α: ; Among them, Energy() represents the modal energy concentration, and the optimization goal is to maximize the energy separation of each IMF component.

4. The AI ​​fault prediction and diagnosis method based on CNC machine tools according to claim 3 is characterized in that: The time-varying scaling term γ(t) is introduced into the improved variational mode decomposition VMD objective function: ;in, is the set of modal components to be decomposed; is the set of center frequencies of each modal component; i is the modal component index; is the time partial derivative operator; is the Dirac delta function; is the imaginary unit; * is the convolution operator; is the time domain signal of the i-th modal component; is the frequency modulation term; is the time-varying scaling factor; is the square of the Frobenius norm.

5. The AI ​​fault prediction and diagnosis method based on CNC machine tools according to claim 4 is characterized in that: Initialize the center frequency when performing modal component decomposition and screening , modal components ; Iterative update via alternating direction multiplier method: ;in, is the n+1th iteration result of the i-th modal component; about Minimum search operator of ; is the time domain signal of the i-th modal component; is the Lagrange multiplier; x is the original input signal; is the sum of all modal components.

6. The AI ​​fault prediction and diagnosis method based on CNC machine tools according to claim 5 is characterized in that: When constructing energy entropy and sensitive features, calculate the energy of each sub-band ;in, is the wavelet packet subband index; is the wavelet packet coefficient; Normalized energy distribution ; Calculate the wavelet packet energy entropy: ; Construct the fault-sensitive feature vector: Fsen=[H1,H2,...,Hm], where m is the number of effective IMF components.

7. The AI ​​fault prediction and diagnosis method based on CNC machine tools according to claim 6 is characterized in that: The sensor feature matrix , modal characteristic matrix Concatenate by feature dimension: ; Among them, rows are feature types; columns are time samples; Perform maximum-minimum normalization on the splicing and output the joint feature matrix .

8. The AI ​​fault prediction and diagnosis method based on CNC machine tools according to claim 7 is characterized in that: When building a hybrid diagnostic model, the training set , validation set Input to a pre-built hybrid model structure; When performing feature fusion and classification, the bidirectional features are concatenated and compressed into ; Splicing and , get the fusion features .

9. The AI ​​fault prediction and diagnosis method based on CNC machine tools according to claim 8 is characterized in that: Configure the loss function: ; Where L is the average loss value of the batch; B is the batch size; is the one-hot encoding matrix of the true label; is the model prediction probability, which indicates the probability that the bth sample is predicted to be the mth class.

10. An AI fault prediction and diagnosis system based on CNC machine tools, characterized in that: include: Diagnostic data processing module: This module monitors the operating status of CNC machine tools throughout their life cycle under different operating conditions, collects multi-source sensor data and fault record data from the machine tools, extracts characteristic parameters of different time series in the sensor data, constructs a high-dimensional feature matrix X(t), and labels the fault record data with fault type labels Y(t), combining them to form a labeled diagnostic dataset. Fault-sensitive processing module: Decomposes the non-stationary vibration signal in the high-dimensional feature matrix X(t) through an improved variational modal decomposition algorithm to obtain multiple modal components IMF(i), and calculates the fault-sensitive characteristics of each component in combination with wavelet packet energy entropy; Fusion feature vector module: aligns the current, temperature, and spindle speed features in multi-source sensor data with the modal component features in time and space, and dynamically weights and fuses them into a fused feature vector F(t) through an attention mechanism; Hybrid diagnosis model construction module: This module takes the fused feature vector F(t) as input and the fault type label Y(t) as output. It constructs a hybrid diagnosis model based on a bidirectional gated recurrent unit and an improved convolutional neural network. The bidirectional gated recurrent unit is used to capture temporal dependencies, and the improved convolutional neural network is used to extract local features and output a preliminary fault prediction probability. Fault prediction and diagnosis module: collects and preprocesses multi-source sensor data during the operation of CNC machine tools in real time, inputs the preprocessed data into the constructed fault diagnosis model, and outputs the fault diagnosis results.

Images