Machine tool vibration monitoring method and system based on time-frequency deep network

Abstract

The application discloses a machine tool vibration monitoring method and system based on a time-frequency deep network, relates to the technical field of machine tool state monitoring, and comprises the following steps: collecting vibration signals, spindle speed signals and machining working condition parameters, processing the vibration signals, and constructing continuous time-frequency sequences; inputting the continuous time-frequency sequences and the machining working condition parameters, extracting vibration time-frequency evolution characteristics, and constructing a future time-frequency bifurcation trajectory set; inputting the future time-frequency candidate trajectory into an improved GANomaly model, calculating the trajectory credibility, obtaining a target future time-frequency trajectory set, and performing generation path continuity verification on the target future time-frequency trajectory, calculating the generation path credibility, and obtaining an effective future time-frequency trajectory; constructing a normal evolution reference chain, calculating the time-frequency state deviation degree, determining the abnormal starting time and the abnormal starting frequency band, and outputting the machine tool vibration monitoring result according to the vibration evolution deviation result. The application realizes machine tool vibration anomaly early warning and starting point positioning through future time-frequency trajectory analysis.

Description

Technical Field

[0001] This invention relates to the field of machine tool condition monitoring technology, and in particular to a machine tool vibration monitoring method and system based on time-frequency depth networks. Background Technology

[0002] During machine tool processing, vibration signals are crucial information carriers reflecting the operating status of the equipment and the quality of machining. Existing machine tool vibration monitoring methods typically involve placing sensors in key components such as the spindle, bearings, or bed to collect vibration signals. These signals are then analyzed in time and frequency using methods such as short-time Fourier transform, wavelet transform, or order analysis to extract characteristic information reflecting the equipment's operating status. Finally, models such as support vector machines, random forests, and convolutional neural networks are used to classify and identify vibration states, thereby enabling the detection and diagnosis of machine tool faults.

[0003] With the development of deep learning technology, some existing technologies have begun to incorporate vibration analysis methods based on deep neural networks. For example, convolutional neural networks or Transformer structures are used to extract features from time-frequency graphs and identify conditions such as bearing failure, tool wear, and chatter. These methods have improved vibration feature extraction capabilities and classification accuracy to some extent, but their core remains based on the analysis of current or historical vibration data, mainly addressing the question of "what kind of anomaly has occurred." However, they lack effective modeling methods for the future evolution trend of machine tool vibration states, making it difficult to predict potential instability processes in advance.

[0004] Therefore, existing technologies have the following shortcomings: First, they rely solely on current or historical time-frequency characteristics for anomaly classification, making it impossible to predict in advance whether vibration will deviate from normal operating conditions, resulting in delayed warnings; second, they lack continuous modeling of the vibration evolution process, making it difficult to distinguish the rationality of different future evolution paths, thus affecting the accuracy of the identification results; third, they lack the ability to locate the starting point of anomalies before they occur, making it impossible to determine when and in which frequency band the anomalies begin, leading to unclear identification of complex working conditions such as chatter, tool anomalies, and early bearing damage, thus limiting the application effect of machine tool vibration monitoring technology in predictive maintenance.

[0005] Therefore, how to provide a machine tool vibration monitoring method and system based on time-frequency deep networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a machine tool vibration monitoring method and system based on time-frequency deep networks. This invention combines time-frequency analysis technology with deep learning methods. By constructing a future time-frequency bifurcation trajectory generation model, introducing an improved GANomaly model for trajectory reliability assessment, and based on the continuity verification of the generated path and vibration evolution deviation analysis, it realizes the transformation of machine tool vibration state from current identification to future evolution prediction. It determines the vibration deviation starting point and key frequency band before the anomaly occurs, and has the advantages of strong prediction ability, high identification accuracy, and clear anomaly location.

[0007] A machine tool vibration monitoring method based on a time-frequency deep network according to an embodiment of the present invention includes: The vibration signals, spindle speed signals, and machining parameters during machine tool operation are collected, and the vibration signals are processed to construct a continuous time-frequency sequence. The continuous time-frequency sequence and processing condition parameters are input into the future time-frequency trajectory generation network to extract vibration time-frequency evolution features, generate multiple future time-frequency candidate trajectories, calculate the evolution probability corresponding to each future time-frequency candidate trajectory, and construct a set of future time-frequency bifurcation trajectories based on the evolution probability. Each future time-frequency candidate trajectory in the future time-frequency bifurcation trajectory set is input into the improved GANomaly model to obtain the reconstructed trajectory and latent space features. The trajectory credibility index between each future time-frequency candidate trajectory and the reconstructed trajectory is calculated. The future time-frequency candidate trajectories are then screened to obtain the target future time-frequency trajectory set. The continuity of the generated path is verified for each target future time-frequency trajectory in the target future time-frequency trajectory set. Based on the time-frequency energy transfer relationship, frequency coupling relationship and modal evolution relationship of adjacent time steps, the credibility of the generated path is calculated. The target future time-frequency trajectory whose generated path credibility meets the preset conditions is determined as the valid future time-frequency trajectory. For each time-frequency state in the effective future time-frequency trajectory, a corresponding normal evolution reference chain is constructed. The degree of deviation between each time-frequency state and the corresponding normal evolution reference state is calculated. The earliest deviation position in the future time-frequency trajectory is determined. Based on the earliest deviation position, the abnormal start time and abnormal start frequency band are determined, and the vibration evolution deviation result is generated. The machine tool vibration monitoring results are output based on the vibration evolution deviation results.

[0008] Optionally, the process of processing the vibration signal to construct a continuous time-frequency sequence includes: The vibration signal is bandpass filtered to obtain a filtered vibration signal; The filtered vibration signal is segmented according to a preset window length and a preset sliding step size to obtain multiple vibration signal segments; Each vibration signal segment is standardized to obtain a standardized vibration signal segment. Time-frequency transformation is performed on each standardized vibration signal segment to obtain the corresponding time-frequency diagram; The continuous time-frequency sequence is constructed by arranging the time-frequency diagrams according to the time sequence of each vibration signal segment.

[0009] Optionally, the step of calculating the evolution probability corresponding to each future time-frequency candidate trajectory and constructing a set of future time-frequency bifurcation trajectories based on the evolution probability includes: A future time-frequency trajectory generation network is constructed, which includes a time-frequency sequence encoding unit, a working condition parameter encoding unit, a time-frequency working condition fusion unit, a bifurcation trajectory generation unit, and a trajectory probability calculation unit. The continuous time-frequency sequence is input into the time-frequency sequence encoding unit in chronological order to encode the vibration energy changes, frequency band position changes and time-frequency texture changes between adjacent time-frequency maps, thereby obtaining time-frequency sequence features that characterize the historical evolution of machine tool vibration. The machining condition parameters are input into the condition parameter encoding unit, and feature encoding is performed on the spindle speed, feed rate, axial depth of cut, radial depth of cut, number of cutter teeth and workpiece material to obtain the condition parameter features that characterize the current machining state. The time-frequency sequence features and condition parameter features are input into the time-frequency condition fusion unit for fusion to obtain the vibration time-frequency evolution features. The vibration time-frequency evolution characteristics are input into the bifurcation trajectory generation unit, and multiple future time-frequency candidate trajectories are generated according to the preset future prediction step size. Each future time-frequency candidate trajectory consists of multiple predicted time-frequency maps arranged in future time order. Different future time-frequency candidate trajectories correspond to different evolution paths of machine tool vibration state in the future time range. Multiple future time-frequency candidate trajectories are input into the trajectory probability calculation unit. The overall trajectory features are extracted for each future time-frequency candidate trajectory. The corresponding trajectory score is generated based on the matching degree between the overall trajectory features and the vibration time-frequency evolution features. All trajectory scores are normalized so that the sum of the evolution probabilities corresponding to each future time-frequency candidate trajectory is one. Each future time-frequency candidate trajectory is bound to its corresponding evolution probability to construct a set of future time-frequency bifurcation trajectories.

[0010] Optionally, the step of filtering future time-frequency candidate trajectories to obtain a set of target future time-frequency trajectories includes: An improved GANomaly model is constructed, which includes a first encoder, an attention-enhanced decoder, a second encoder, a frequency domain discriminator, and a time domain discriminator. The first encoder consists of three layers of residual convolutional blocks. A channel attention module is inserted after the second residual convolutional block. The attention module is used to highlight the time-frequency energy distribution related to the spindle order, the cutter tooth order, and the bearing characteristic frequency. Each future time-frequency candidate trajectory in the future time-frequency bifurcation trajectory set is used as an input sample. The input sample consists of several consecutive time-frequency images. It is input to the first encoder to obtain the first latent space features. The first latent space features are input to the attention enhancement decoder. The reconstructed trajectory is generated using the jump structure corresponding to the first encoder. The reconstructed trajectory is input to the second encoder to obtain the second latent space features. The reconstructed trajectory of the future time-frequency candidate trajectory is input into the frequency domain discriminator to obtain the frequency domain discrimination result, and simultaneously input into the time domain discriminator to obtain the time domain discrimination result. The frequency domain discriminator adopts a spectrum block convolution structure, which focuses on detecting the energy difference of the main frequency band, and the time domain discriminator adopts a one-dimensional convolution fusion self-attention structure, which focuses on detecting the difference of impact vibration. The trajectory reconstruction difference, latent space difference, frequency domain discrimination difference, and time domain discrimination difference are calculated for each future time-frequency candidate trajectory. The trajectory credibility index is obtained by weighting the four differences. The difference weighting coefficients are automatically searched in advance through the normal machine tool vibration sample training process. During the training process, a combination loss function of cross-entropy loss, reconstruction loss, latent space consistency loss, frequency band preservation loss, and energy balance loss is used. The optimizer is Adam with weighted decay. All future time-frequency candidate trajectories are sorted from high to low according to the trajectory confidence index, and the future time-frequency candidate trajectories with a trajectory confidence index higher than the set threshold are selected as the target future time-frequency trajectory.

[0011] Optionally, the calculation of the generated path credibility, and the determination of the target future time-frequency trajectory whose generated path credibility meets preset conditions as a valid future time-frequency trajectory, includes: Obtain the set of future time-frequency trajectories of the target, divide each future time-frequency trajectory of the target into multiple consecutive time-frequency states according to the future time sequence, and determine the transformation relationship between two adjacent consecutive time-frequency states as a vibration generation path segment; For each vibration generation path segment, the total energy, spindle order energy, cutter tooth passage order energy, bearing characteristic frequency band energy, and structural modal frequency band energy are extracted from the previous and subsequent time-frequency states. Based on the energy change of the subsequent time-frequency state relative to the previous time-frequency state, the corresponding time-frequency energy transfer characteristics are obtained. For each vibration generation path segment, the spindle order frequency band, the cutter tooth passage order frequency band, the bearing characteristic frequency band, and the structural mode frequency band are extracted from the previous and next time-frequency states, and the corresponding frequency coupling characteristics are obtained based on the position offset, bandwidth change, and energy ratio change of each frequency band in adjacent time-frequency states. For each vibration generation path segment, the spindle modal components, tool modal components, bearing modal components, and structural modal components are extracted from the previous and next time-frequency states. Based on the enhancement, attenuation, and transfer of each modal component in adjacent time-frequency states, the corresponding modal evolution characteristics are obtained. The time-frequency energy migration features, frequency coupling features, and modal evolution features corresponding to each vibration generation path segment in the future time-frequency trajectory of the same target are input into the generation path credibility calculation unit. The generation path credibility calculation unit calculates the energy migration continuity score, frequency coupling continuity score, and modal evolution continuity score respectively. The energy migration continuity score, frequency coupling continuity score, and modal evolution continuity score are weighted and fused to obtain the generation path credibility corresponding to the future time-frequency trajectory of the target. The credibility of the generated path corresponding to the future time-frequency trajectory of each target is compared with the preset path credibility threshold. The future time-frequency trajectory of the target with a generated path credibility not lower than the preset path credibility threshold is determined as the valid future time-frequency trajectory.

[0012] Optionally, the generation of vibration evolution deviation results includes: Obtain valid future time-frequency trajectories, split each valid future time-frequency trajectory into multiple future time-frequency states according to future time sequence, and record the future time steps corresponding to each future time-frequency state; Based on the machining condition parameters, select normal future time-frequency trajectories that match the current spindle speed, feed rate, axial depth of cut, radial depth of cut, number of cutter teeth, and workpiece material code from the normal vibration evolution sample library. Align the normal future time-frequency trajectories with the effective future time-frequency trajectories step by step in chronological order to form a normal evolution reference chain. For each future time step, the total energy features, main shaft order frequency band features, cutter tooth passage order frequency band features, bearing feature frequency band features, and structural mode frequency band features of the effective future time-frequency trajectory and the normal evolution reference chain in the corresponding time-frequency state are extracted respectively. The deviation degree corresponding to the future time step is obtained according to the differences between the features. The deviation degree of each future time step is compared with the preset deviation threshold in the future time sequence. The future time step that first reaches the preset deviation threshold is determined as the earliest deviation time step. In the time-frequency state corresponding to the earliest deviation time step, the frequency band with the largest deviation degree is determined as the abnormal starting frequency band. The vibration evolution deviation results are generated based on the earliest deviation time step, the abnormal starting frequency band, the degree of deviation corresponding to each future time step, and the corresponding effective future time-frequency trajectory identifier.

[0013] Optionally, the step of outputting machine tool vibration monitoring results based on the vibration evolution deviation results includes: The earliest deviation time step in the vibration evolution deviation result is converted into the corresponding actual time to obtain the anomaly initiation time. Based on the abnormal starting frequency band in the vibration evolution deviation results, the key frequency range to which the abnormal vibration belongs is determined, and the key frequency range is matched with the spindle order frequency band, the cutter tooth passing order frequency band, the bearing characteristic frequency band and the structural mode frequency band to obtain the abnormality type identifier; Based on the deviation change information corresponding to each future time step, the vibration deviation growth trend is calculated, and the vibration instability development trend is determined based on the vibration deviation growth trend. The risk level of machine tool vibration is determined based on the onset time of the anomaly, the anomaly type identifier, and the vibration instability development trend; The system integrates the abnormal start time, abnormal start frequency band, abnormal type identifier, vibration instability development trend, and risk level to generate and output machine tool vibration monitoring results.

[0014] A machine tool vibration monitoring system based on a time-frequency deep network according to an embodiment of the present invention includes: The data acquisition and time-frequency construction module is used to acquire vibration signals, spindle speed signals and machining parameters, and process the vibration signals to construct a continuous time-frequency sequence. The future time-frequency trajectory generation module is used to extract vibration time-frequency evolution features based on continuous time-frequency sequences and processing condition parameters, generate multiple future time-frequency candidate trajectories, and construct a set of future time-frequency bifurcation trajectories; The trajectory credibility assessment module is used to input future time-frequency candidate trajectories into the improved GANomaly model to calculate trajectory credibility and filter to obtain the target future time-frequency trajectory set; The path generation verification module is used to verify the continuity of the generated path for the target's future time-frequency trajectory, calculate the credibility of the generated path, and filter out the valid future time-frequency trajectories. The evolutionary deviation analysis module is used to calculate the degree of deviation based on the effective future time-frequency trajectory and the normal evolution reference chain, determine the earliest deviation position, and generate vibration evolution deviation results. The monitoring result output module is used to output machine tool vibration monitoring results based on the vibration evolution deviation results.

[0015] The beneficial effects of this invention are: This invention expands the monitoring target of machine tool vibration from the traditional current or historical time-frequency state to the future multi-path evolution process by introducing a future time-frequency bifurcation trajectory generation mechanism. This fundamentally changes the limitation of existing technologies, which can only identify anomalies that have already occurred. By modeling and probabilistically characterizing multiple future time-frequency candidate trajectories, it can reflect the possible development trend of machine tool vibration in advance, achieving predictive monitoring before obvious vibration anomalies occur. This effectively improves the early warning capability for problems such as chatter, tool abnormalities, and early bearing damage, and reduces the processing risks caused by delayed early warning.

[0016] Furthermore, this invention employs an improved GANomaly model to assess the credibility of future time-frequency candidate trajectories and combines this with a path continuity verification mechanism to perform multi-layered constraint screening on the rationality of the future vibration evolution process. This not only avoids the problem of excessive reliance on anomalous modes in traditional methods but also effectively eliminates false predicted trajectories that do not conform to the physical evolution laws of machine tools, thereby improving the stability and accuracy of monitoring results. By introducing multi-dimensional features such as time-frequency energy transfer, frequency coupling, and modal evolution, the vibration analysis is made closer to the actual operating mechanism of machine tools, enhancing the model's adaptability to complex working conditions.

[0017] This invention, by constructing a vibration evolution deviation analysis mechanism, achieves refined localization of the anomaly formation process. It can determine the earliest deviation from the normal evolution path and its corresponding frequency band in future time-frequency trajectories, clarifying the anomaly's starting location and development trend. Compared to traditional methods that only output fault types, this invention can not only determine whether an anomaly exists but also provide the time starting point and frequency source of the anomaly, improving the interpretability and engineering application value of monitoring results and providing more reliable technical support for machine tool condition assessment and predictive maintenance. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a machine tool vibration monitoring method based on time-frequency deep networks proposed in this invention; Figure 2 This is a schematic diagram of the structure of a machine tool vibration monitoring system based on a time-frequency depth network proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1A machine tool vibration monitoring method based on time-frequency deep networks, comprising: The vibration signals, spindle speed signals, and machining parameters during machine tool operation are collected, and the vibration signals are processed to construct a continuous time-frequency sequence. The continuous time-frequency sequence and processing condition parameters are input into the future time-frequency trajectory generation network to extract vibration time-frequency evolution features, generate multiple future time-frequency candidate trajectories, calculate the evolution probability corresponding to each future time-frequency candidate trajectory, and construct a set of future time-frequency bifurcation trajectories based on the evolution probability. Each future time-frequency candidate trajectory in the future time-frequency bifurcation trajectory set is input into the improved GANomaly model to obtain the reconstructed trajectory and latent space features. The trajectory credibility index between each future time-frequency candidate trajectory and the reconstructed trajectory is calculated. The future time-frequency candidate trajectories are then screened to obtain the target future time-frequency trajectory set. The continuity of the generated path is verified for each target future time-frequency trajectory in the target future time-frequency trajectory set. Based on the time-frequency energy transfer relationship, frequency coupling relationship and modal evolution relationship of adjacent time steps, the credibility of the generated path is calculated. The target future time-frequency trajectory whose generated path credibility meets the preset conditions is determined as the valid future time-frequency trajectory. For each time-frequency state in the effective future time-frequency trajectory, a corresponding normal evolution reference chain is constructed. The degree of deviation between each time-frequency state and the corresponding normal evolution reference state is calculated. The earliest deviation position in the future time-frequency trajectory is determined. Based on the earliest deviation position, the abnormal start time and abnormal start frequency band are determined, and the vibration evolution deviation result is generated. The machine tool vibration monitoring results are output based on the vibration evolution deviation results.

[0021] In this embodiment, the process of processing the vibration signal to construct a continuous time-frequency sequence includes: The vibration signal is bandpass filtered to obtain a filtered vibration signal; The filtered vibration signal is segmented according to a preset window length and a preset sliding step size to obtain multiple vibration signal segments; Each vibration signal segment is standardized to obtain a standardized vibration signal segment. Time-frequency transformation is performed on each standardized vibration signal segment to obtain the corresponding time-frequency diagram; The continuous time-frequency sequence is constructed by arranging the time-frequency diagrams according to the time sequence of each vibration signal segment.

[0022] In this embodiment, the step of calculating the evolution probability corresponding to each future time-frequency candidate trajectory and constructing a set of future time-frequency bifurcation trajectories based on the evolution probability includes: A future time-frequency trajectory generation network is constructed, which includes a time-frequency sequence encoding unit, a working condition parameter encoding unit, a time-frequency working condition fusion unit, a bifurcation trajectory generation unit, and a trajectory probability calculation unit. The continuous time-frequency sequence is input into the time-frequency sequence encoding unit in chronological order to encode the vibration energy changes, frequency band position changes, and time-frequency texture changes between adjacent time-frequency maps, thereby obtaining time-frequency sequence features characterizing the historical evolution of machine tool vibration. Specifically, obtaining time-frequency sequence features characterizing the historical evolution of machine tool vibration involves: Read adjacent time-frequency maps arranged in chronological order. For the i-th time-frequency map, using a 128×128 time-frequency grid as coordinates, calculate the total energy, energy centroid frequency, first-order spindle energy, cutting tooth passage order energy, bearing inner ring characteristic energy, bearing outer ring characteristic energy, structural mode main peak energy, and texture direction entropy. Arrange these eight scalars in a predetermined order to generate an 8-dimensional single-frame energy-frequency vector. Read all time-frequency maps sequentially and concatenate all 8-dimensional vectors in chronological order to obtain an energy-frequency feature matrix of size L×8, where L is the number of historical time-frequency map frames. Calculate the absolute value of the element-wise difference between the energy-frequency vectors corresponding to the i-th and (i+1)-th time-frequency maps, and use the criterion that the absolute value of the difference is no greater than twice the standard deviation of its component mean to filter out the energy-frequency features. The frequency band smoothing sequence is generated by marking the positions where the criteria are met as 1 and generating an energy smoothing mask. Further calculations are made on two adjacent time-frequency images to determine the changes in the spindle order band center frequency, the cutter tooth passage order band center frequency, and the bearing characteristic band energy ratio. These three changes are then concatenated column-wise to form an L-1 row, 3-column frequency band change matrix. The energy-frequency band feature matrix and the frequency band change matrix are aligned by time steps and stacked vertically to form a time-frequency evolution feature matrix of size L×11. Based on the energy smoothing mask, rows satisfying the smoothing condition for three or more consecutive frames are selected in the time-frequency evolution feature matrix to establish a time-frequency evolution adjacency relationship. This adjacency relationship is written into a 0-1 matrix and output along with the time-frequency evolution feature matrix to obtain the time-frequency sequence features characterizing the historical evolution of machine tool vibration. The machining condition parameters are input into the condition parameter encoding unit to perform feature encoding on the spindle speed, feed rate, axial depth of cut, radial depth of cut, number of teeth, and workpiece material, thereby obtaining the condition parameter features characterizing the current machining state. The time-frequency sequence features and the condition parameter features are then input into the time-frequency condition fusion unit for fusion to obtain the vibration time-frequency evolution features. Specifically, the condition parameter features characterizing the current machining state are as follows: Read the spindle speed, feed rate, axial depth of cut, radial depth of cut, number of teeth, and workpiece material code from the current machining condition parameters. Read the spindle speed in revolutions per minute (RPM) as the first value, the feed rate in millimeters per minute (mm) as the second value, the axial and radial depths of cut in millimeters as the third and fourth values ​​respectively, the number of teeth as the fifth value, and the workpiece material code as the sixth value. Divide the first to fifth values ​​by their corresponding preset upper limits to obtain normalized values ​​for the spindle speed, feed rate, axial depth of cut, radial depth of cut, and number of teeth. Convert the sixth value into a unique thermal encoding vector for the material. Then, concatenate the normalized values ​​for the spindle speed, feed rate, axial depth of cut, radial depth of cut, number of teeth, and material thermal encoding vector in a fixed order to generate the first machining condition vector. Read adjacent two... The spindle speed difference, feed rate difference, axial depth of cut difference, and radial depth of cut difference within each sampling period are divided by their respective preset upper limits to obtain the speed change, feed rate change, axial depth of cut change, and radial depth of cut change. These speed change, feed rate change, axial depth of cut change, and radial depth of cut change are then concatenated sequentially to generate a second working condition vector. The first and second working condition vectors are then concatenated column-wise to obtain the original working condition feature vector. This original working condition feature vector is input into a two-layer fully connected encoding structure. The first layer maps the original working condition feature vector to a 32-dimensional intermediate vector and performs normalization and ReLU activation on the 32-dimensional intermediate vector. The second layer maps the 32-dimensional intermediate vector to a 16-dimensional working condition parameter feature vector. This 16-dimensional working condition parameter feature vector is then output as the working condition parameter feature representing the current machining state. The vibration time-frequency evolution characteristics are input into the bifurcation trajectory generation unit, and multiple future time-frequency candidate trajectories are generated according to a preset future prediction step size. Each future time-frequency candidate trajectory consists of multiple predicted time-frequency maps arranged in future time sequence. Different future time-frequency candidate trajectories correspond to different evolution paths of the machine tool vibration state within the future time range. Specifically, the multiple future time-frequency candidate trajectories are generated according to the preset future prediction step size as follows: First, the vibration time-frequency evolution feature tensor is read and input into the trajectory seed generation layer. The trajectory seed generation layer has five branches in parallel, with each branch adding a branch identifier vector at its beginning as a branch conditional embedding. Then, the branch conditional embedding is concatenated with the vibration time-frequency evolution feature tensor and fed into the corresponding branch's two-layer gated recurrent unit. The first-layer gated recurrent unit outputs and caches the hidden state for each future time step, while the second-layer gated recurrent unit recursively runs using the first-layer output and the prediction result from the previous time step as input. During the recursive phase, each branch maps its hidden state to the first future time step using two layers of one-dimensional deconvolution. A predicted time-frequency map is generated, and minimum truncation, maximum limit, and overall energy normalization are applied to it. Then, the predicted time-frequency map at the current time step is concatenated with the initial hidden state, and the next predicted time-frequency map is generated recursively through the same deconvolution structure. This process is repeated until the cumulative number of generated maps matches the preset future prediction step size. After all predicted time-frequency maps have been generated, the predicted time-frequency maps within the same branch are timestamped in the order of generation and assembled into a future time-frequency candidate trajectory. The same operation is performed on all five branches, resulting in five future time-frequency candidate trajectories with different evolution paths within the future time range. Multiple future time-frequency candidate trajectories are input into the trajectory probability calculation unit. For each future time-frequency candidate trajectory, overall trajectory features are extracted. A corresponding trajectory score is generated based on the matching degree between the overall trajectory features and the vibration time-frequency evolution features. All trajectory scores are normalized so that the sum of the evolution probabilities corresponding to each future time-frequency candidate trajectory is one. Each future time-frequency candidate trajectory is bound to its corresponding evolution probability to construct a set of future time-frequency bifurcation trajectories. Specifically, for each future time-frequency candidate trajectory, overall trajectory features are extracted, and a corresponding trajectory score is generated based on the matching degree between the overall trajectory features and the vibration time-frequency evolution features. Read all predicted time-frequency maps arranged in future time sequence from each future time-frequency candidate trajectory. For each predicted time-frequency map, calculate the total energy, energy centroid frequency, principal axis order band energy, cutter tooth passage order band energy, bearing characteristic band energy, structural mode band energy, highest energy band position, and band energy dispersion. Arrange these features in a fixed order to form a single-step trajectory feature vector. Concatenate all single-step trajectory feature vectors from the same future time-frequency candidate trajectory in future time sequence to form a trajectory feature matrix. The mean, maximum, minimum, and terminal values, as well as the change in the terminal value relative to the initial value, are calculated column-wise for the trajectory feature matrix. The calculation results are expanded in column order to generate the overall trajectory feature vector. The historical time-frequency features, frequency band migration features, and operating condition fusion features of the terminal values ​​in the vibration time-frequency evolution features are read and concatenated to generate the historical evolution reference vector. The overall trajectory feature vector and the historical evolution reference vector are adjusted to the same dimension, and the absolute difference between them is calculated bit by bit. The average of all absolute differences is used to obtain the first matching difference value. Read the first predicted time-frequency map of the future time-frequency candidate trajectory and the last historical time-frequency map in the continuous time-frequency sequence, and calculate the total energy difference, spindle order band energy difference, cutter tooth passage order band energy difference, bearing characteristic frequency band energy difference, and structural mode frequency band energy difference between the two respectively. Average the differences to obtain the initial continuous difference value; read the energy change and frequency band position change between adjacent predicted time-frequency maps in the future time-frequency candidate trajectory, and count the number of times the change direction repeatedly switches to obtain the trajectory oscillation number. The first matching difference value, the initial continuous difference value, and the number of trajectory oscillations are input into the trajectory scoring layer. The trajectory scoring layer first normalizes the three inputs, and then adds them together according to preset weights to obtain the trajectory mismatch value. The trajectory mismatch value is mapped to the trajectory score in the opposite direction, so that the smaller the trajectory mismatch value, the higher the corresponding trajectory score. The above processing is performed on all future time-frequency candidate trajectories to obtain the trajectory score corresponding to each future time-frequency candidate trajectory.

[0023] In this embodiment, the step of filtering future time-frequency candidate trajectories to obtain a set of target future time-frequency trajectories includes: An improved GANomaly model is constructed, which includes a first encoder, an attention-enhanced decoder, a second encoder, a frequency domain discriminator, and a time domain discriminator. The first encoder consists of three layers of residual convolutional blocks. A channel attention module is inserted after the second residual convolutional block. The attention module is used to highlight the time-frequency energy distribution related to the spindle order, the cutter tooth order, and the bearing characteristic frequency. Each future time-frequency candidate trajectory in the future time-frequency bifurcation trajectory set is used as an input sample. The input sample consists of several consecutive time-frequency images and is input to the first encoder to obtain the first latent space features. The first latent space features are then input to the attention-enhanced decoder, which uses the jump structure corresponding to the first encoder to generate a reconstructed trajectory. The reconstructed trajectory is then input to the second encoder to obtain the second latent space features, where: The first latent space features are obtained as follows: The predicted time-frequency maps of consecutively arranged candidates for the same future time-frequency trajectory are stacked into a four-dimensional tensor in chronological order. A two-dimensional convolutional layer of size 3×3 with a stride of 1 is used to extract local time-frequency texture features. After the output of the convolutional layer, it is passed through a batch normalization layer and a ReLU activation layer in sequence. The convolutional output is downsampled simultaneously on the time axis and frequency axis using a 2×2 max pooling layer. The pooling result is fed into a second convolutional block. The second convolutional block has the same structure as the first block but doubles the number of channels. After repeating the convolutional block and pooling structure three times, global average pooling is used after the output of the third pooling layer to compress the three-dimensional feature map into a one-dimensional vector. Then, it is mapped to a fixed-length latent space vector through a fully connected layer to obtain the first latent space feature representing the overall vibration evolution mode. The second latent space features are obtained as follows: The reconstructed trajectory output by the attention-enhanced decoder is re-stacked into a four-dimensional tensor according to the original input order. The reconstructed trajectory is then encoded a second time using a convolution-batch normalization-ReLU-pooling structure that is exactly the same as the first encoder but does not share parameters. Similarly, after repeating the convolutional blocks and pooling three times, global average pooling is used to compress the feature map into a one-dimensional vector. This vector is then mapped to the same dimension as the first latent space feature through a fully connected layer, and the second latent space feature corresponding to the reconstructed trajectory is output. The reconstructed trajectory from the future time-frequency candidate trajectory is input into the frequency domain discriminator to obtain the frequency domain discrimination result, and simultaneously input into the time domain discriminator to obtain the time domain discrimination result. The frequency domain discriminator adopts a spectrum block convolution structure, focusing on detecting energy differences in the main frequency bands, while the time domain discriminator adopts a one-dimensional convolution fusion self-attention structure, focusing on detecting differences in impact vibrations. The frequency domain discrimination result is obtained as follows: For any future time-frequency candidate trajectory and its corresponding reconstructed trajectory, firstly, each predicted time-frequency image is read in future time sequence. A two-dimensional fast Fourier transform is performed on each predicted time-frequency image across the entire range to extract the amplitude spectrum. Then, the amplitude spectrum is sliced ​​at equal widths along the frequency axis, grouped into sets of 20 adjacent frequency points, resulting in several spectral blocks. All spectral blocks at the same time step are sequentially input into a frequency domain discriminator using a three-layer two-dimensional convolutional structure. The first convolutional layer has a kernel size of 3x3 and a stride of 1, outputting eight channels. The second and third convolutional layers both have kernel sizes of 2x2. The second layer outputs 16 channels, and the third layer outputs 32 channels. After each convolutional layer, batch normalization, ReLU activation, and max pooling are performed sequentially. Global average pooling is used on the output of the third convolutional layer to generate block-level feature vectors. All block-level feature vectors within the same time step are concatenated to form a time-level feature vector, and then concatenated in chronological order to form a trajectory-level frequency domain feature vector. The trajectory-level frequency domain feature vector is input into a fully connected scorer, and the output is mapped to a percentage score using a Sigmoid function. The larger the output value, the greater the difference in frequency band energy distribution between the reconstructed trajectory and the original candidate trajectory. The time-domain discrimination result is obtained as follows: For the same pair of future time-frequency candidate trajectories and reconstructed trajectories, the energy summation along the frequency axis of each time-frequency image is obtained to obtain the time-domain envelope, transforming the entire trajectory into a vibration envelope sequence sampled at equal intervals. The vibration envelope sequence is input into a one-dimensional convolutional fusion self-attention discriminator. The first layer of the discriminator uses a convolution kernel length of 25 and a stride of 1 to output 64 channels and performs batch normalization and ReLU activation. The second layer uses a convolution kernel length of 15 and a stride of 1 to output 128 channels and performs layer normalization. The output of the second convolution layer is fed into a multi-head self-attention layer. The attention layer has four heads and uses residual connections and a feedforward network to obtain weighted time-series features. Global average pooling is performed on the time-series features to obtain a time-domain feature vector, which is input into a one-layer two-node fully connected scorer to output a time-domain difference score. The higher the score, the more significant the abrupt change in envelope energy and the greater the difference in impact amplitude. Finally, the frequency domain difference score and the time-domain difference score are returned together. The trajectory reconstruction difference, latent space difference, frequency domain discrimination difference, and temporal domain discrimination difference are calculated for each future time-frequency candidate trajectory. A trajectory reliability index is obtained by weighting these four differences. The weighting coefficients are automatically searched beforehand through a normal machine tool vibration sample training process. During training, a combined loss function of cross-entropy loss, reconstruction loss, latent space consistency loss, bandgap preservation loss, and energy balance loss is used. The optimizer is Adam with weighted decay, where the initial learning rate is set to 1 / 10,000, decaying by 10% every fifteen rounds. Specifically, the trajectory reconstruction difference, latent space difference, frequency domain discrimination difference, and temporal domain discrimination difference for each future time-frequency candidate trajectory are calculated as follows: The time-frequency energy values ​​of the same future time-frequency candidate trajectory and reconstructed trajectory at the same row and column coordinates are read frame by frame. The absolute value of the energy difference at corresponding positions in the two time-frequency images is calculated and averaged. The average difference obtained at each future time step is accumulated in chronological order and then divided by the number of future time steps to obtain the trajectory reconstruction difference. The latent space feature vectors output by the first encoder and the latent space feature vectors output by the second encoder are L2 normalized respectively. Then, the absolute values ​​are subtracted element by element in chronological order and averaged to obtain the latent space difference. The frequency domain difference score returned by the frequency domain discriminator and the time domain difference score returned by the time domain discriminator are called. The difference between the two is calculated with the preset frequency domain reference value and time domain reference value to obtain the frequency domain discrimination difference and time domain discrimination difference. The four difference weighting coefficients automatically searched during the training phase of normal machine tool vibration samples are read. The four differences are weighted and summed according to the weighting coefficients. The summation result is mapped to a percentage value in reciprocal form to obtain the final trajectory reliability index of this trajectory. All future time-frequency candidate trajectories are sorted from high to low according to the trajectory confidence index, and the future time-frequency candidate trajectories with a trajectory confidence index higher than the set threshold are selected as the target future time-frequency trajectory.

[0024] In this embodiment, the calculation of the generated path credibility, and the determination of the target future time-frequency trajectory whose generated path credibility meets the preset conditions as the valid future time-frequency trajectory, includes: Obtain the set of future time-frequency trajectories of the target, divide each future time-frequency trajectory of the target into multiple consecutive time-frequency states according to the future time sequence, and determine the transformation relationship between two adjacent consecutive time-frequency states as a vibration generation path segment; For each vibration generation path segment, the total energy, spindle order energy, cutter tooth passage order energy, bearing characteristic frequency band energy, and structural modal frequency band energy are extracted from the previous and subsequent time-frequency states. Based on the energy change of the subsequent time-frequency state relative to the previous time-frequency state, the corresponding time-frequency energy transfer characteristics are obtained. Specifically, the corresponding time-frequency energy transfer characteristics are obtained as follows: Five energy statistical bands are pre-divided in the entire time-frequency grid, corresponding to the total energy band, the first-order band of the main shaft, the order band of the cutting teeth, the characteristic frequency band of the bearing, and the frequency band of the structural modes, respectively. For the previous time-frequency state in the vibration generation path segment, the grid is traversed row by row, and the energy values ​​belonging to each statistical band are accumulated to obtain the total energy value, main shaft energy value, cutting tooth energy value, bearing energy value, and structural mode energy value of the previous time-frequency state in sequence. The same accumulation operation is performed on the next time-frequency state in the vibration generation path segment to obtain the corresponding five energy values ​​of the next time-frequency state. Calculate the difference between the energy values ​​of the previous time-frequency state and the corresponding energy values ​​of the next time-frequency state. Encode the difference in three directions: increase, decrease, and basically remain unchanged. The difference is set to increase if it is 3 percentage points higher than the previous time-frequency state, decrease if it is less than 3 percentage points higher, and basically remain unchanged if it is in between. The energy change directions obtained from the five statistical bands are sequentially concatenated to form an energy change direction vector containing five three-valued symbols; the ratio of the absolute value of the five energy differences to the corresponding energy value of the previous time-frequency state is calculated, and the ratios are sequentially concatenated to form a five-dimensional energy change amplitude vector; the energy change direction vector and the energy change amplitude vector are sequentially merged to generate a time-frequency energy migration feature vector of length 10. For each vibration generation path segment, the spindle order frequency band, cutter tooth passage order frequency band, bearing characteristic frequency band, and structural mode frequency band are extracted from the previous and subsequent time-frequency states. Based on the positional shift, bandwidth variation, and energy proportion variation of each frequency band in adjacent time-frequency states, the corresponding frequency coupling characteristics are obtained. Specifically, the obtained frequency coupling characteristics are as follows: The center frequencies and default bandwidths of four target frequency bands are marked in the entire time-frequency grid. The center frequencies of the main shaft order, the center frequencies of the cutting teeth, the center frequencies of the bearing characteristic frequency band, and the center frequencies of the structural mode frequency band are stored respectively. For the previous time-frequency state in the vibration generation path segment, the row where the center frequency is located is found using the frequency axis index. The effective range of the frequency band is obtained by expanding it upwards and downwards by half of the default bandwidth. The energy within the range is counted and the three frequency values ​​of the upper boundary, lower boundary, and center are recorded. The total energy obtained is divided by the total energy of the previous time-frequency state to obtain the energy percentage. The four frequency band information of the previous state are recorded in the order of center, upper boundary, lower boundary, and energy percentage. Perform the same frequency band extraction process on the next time-frequency state to obtain the corresponding four frequency band information of the next state. For the same target frequency band, calculate the absolute value of the difference between the center frequency of the next state and the center frequency of the previous state as the position offset, calculate the absolute value of the difference between the bandwidth of the next state and the bandwidth of the previous state as the bandwidth change, and calculate the absolute value of the difference between the energy proportion of the next state and the energy proportion of the previous state as the proportion change. If the position offset is higher than three times the frequency resolution, it is marked as obvious drift; if it is lower than the frequency resolution, it is marked as stable; if it is between the two, it is marked as slight drift. If the bandwidth change exceeds one-tenth of the default bandwidth, it is marked as expansion or contraction; if it does not exceed, it is marked as stable. If the energy proportion change exceeds 5%, it is marked as energy mutation; if it does not exceed, it is marked as energy stability. After converting the position offset marker, bandwidth change marker, and energy percentage change marker into integer encoded values ​​in a predetermined order, they are concatenated to generate a frequency coupling feature vector of length 12. For each vibration generation path segment, the spindle modal components, tool modal components, bearing modal components, and structural modal components are extracted from the previous and subsequent time-frequency states. Based on the enhancement, attenuation, and transition of each modal component in adjacent time-frequency states, the corresponding modal evolution features are obtained. Specifically, the obtained modal evolution features are as follows: Four modal identification curves are pre-set in the entire time-frequency grid, corresponding to the first-order natural mode of the spindle, the tool overhang mode, the bearing ring resonance mode, and the bed structure mode, respectively. For the previous time-frequency state in the vibration generation path segment, the time-frequency diagram is decomposed into multiple intrinsic modal component curves using two-dimensional empirical modal decomposition. The curves are sorted from high to low energy, and an intrinsic modal component is matched for each identification curve based on the principle of minimizing the distance between the center frequency of the curve and the four modal identification curves. Record the energy values, center frequencies, and bandwidths of the matched intrinsic modal components, which are then used as the spindle modal energy, tool modal energy, bearing modal energy, and structural modal energy, along with their corresponding center frequencies and bandwidths. Perform the same decomposition and matching process on the subsequent time-frequency state in the vibration generation path segment to obtain the energy, center frequency, and bandwidth of the four modal components corresponding to the subsequent state. For the same modal component, compare the energy of the subsequent state with that of the previous state. If the energy increase exceeds 15% of the previous state's energy, it is marked as enhancement; if the decrease exceeds 15%, it is marked as attenuation; otherwise, it is marked as... Record as basically maintained; compare the center frequency of the next state with the center frequency of the previous state. If the offset exceeds twice the frequency resolution, it is marked as frequency drift; otherwise, it is marked as stable; compare the bandwidth of the next state with the bandwidth of the previous state. If the bandwidth increases by more than 10% of the bandwidth of the previous state, it is marked as widening; if it shrinks by more than 10%, it is marked as narrowing; otherwise, it is marked as bandwidth stable; encode the judgment results of the four modal components in the three dimensions of energy change, frequency change and bandwidth change in the order of spindle, tool, bearing and structure. Each judgment result occupies one bit, and a total of 12 bits of three-value code is obtained; After encoding, three real-valued features are added: the ratio of the absolute value of the modal energy difference between the previous and next states to the total energy of the previous state, the ratio of the modal center frequency offset to the total bandwidth of the frequency axis, and the ratio of the absolute value of the modal bandwidth difference to the bandwidth of the previous state. These three real-valued features are then concatenated in modal order. Finally, the 12-bit code and the 12 real-valued features are merged sequentially to generate a modal evolution feature vector of length 24. The time-frequency energy transfer characteristics, frequency coupling characteristics, and modal evolution characteristics corresponding to each vibration generation path segment in the future time-frequency trajectory of the same target are input into the generation path credibility calculation unit. The generation path credibility calculation unit calculates the energy transfer continuity score, frequency coupling continuity score, and modal evolution continuity score respectively. The energy transfer continuity score, frequency coupling continuity score, and modal evolution continuity score are weighted and fused to obtain the generation path credibility corresponding to the future time-frequency trajectory of the target. The calculation of the energy transfer continuity score, frequency coupling continuity score, and modal evolution continuity score is as follows: Within the same target future time-frequency trajectory, the ten-dimensional time-frequency energy migration features are read sequentially according to the path segments. For the energy change direction of the same statistical band, if the direction markings of three consecutive segments remain consistent, it is counted as an energy migration stable event; if a sudden change in direction occurs, it is counted as an energy migration break event. The number of all energy migration stable events and energy migration break events is summarized, and the ratio of the number of stable events to the total number of segments is used as the basic score for energy migration continuity. Then, the basic score is smoothed and penalized according to the mean square deviation of the energy change amplitude vector of each segment to obtain the energy migration continuity score. The 12-dimensional frequency coupling features are read sequentially, and the time series of three marker bits in the same target frequency band are compared. If the position offset, bandwidth change and energy ratio change markers maintain the same judgment results in adjacent segments, it is counted as a frequency coupling stable event. The ratio of the number of frequency coupling stable events to the total number of segments is used as the basic score of frequency coupling continuity. Then, the amplitude penalty is applied according to the absolute mean of the frequency band offset and bandwidth change of each segment to obtain the frequency coupling continuity score. The 24-dimensional modal evolution features are read, and the energy enhancement marker, frequency drift marker, and bandwidth change marker of the same mode are detected in segment order. If the three markers of three consecutive segments are consistent, it is counted as a modal evolution stabilization event. The ratio of the total number of the four types of modal stabilization events to the total number of segments is used as the basic score of modal evolution continuity. The basic score is then adjusted by combining the comprehensive mean square value of the ratio of energy difference of the four types of modes, the ratio of center frequency offset, and the ratio of bandwidth difference, to obtain the modal evolution continuity score. The energy migration continuity score, the frequency coupling continuity score, and the modal evolution continuity score are weighted and added according to a preset ratio to output the credibility of the generation path of the target's future time-frequency trajectory. The credibility of the generated path corresponding to the future time-frequency trajectory of each target is compared with the preset path credibility threshold. The future time-frequency trajectory of the target with a generated path credibility not lower than the preset path credibility threshold is determined as the valid future time-frequency trajectory.

[0025] In this embodiment, generating the vibration evolution deviation result includes: Obtain valid future time-frequency trajectories, split each valid future time-frequency trajectory into multiple future time-frequency states according to future time sequence, and record the future time steps corresponding to each future time-frequency state; Based on the machining condition parameters, a normal future time-frequency trajectory matching the current spindle speed, feed rate, axial depth of cut, radial depth of cut, number of cutter teeth, and workpiece material code is selected from the normal vibration evolution sample library. This normal future time-frequency trajectory is then aligned step-by-step with the effective future time-frequency trajectory in chronological order to form a normal evolution reference chain. Specifically, forming the normal evolution reference chain involves: The spindle speed, feed rate, axial depth of cut, radial depth of cut, number of teeth, and workpiece material code from the current machining condition parameters are concatenated into a condition index vector in a predetermined order. An inverted index is established in the normal vibration evolution sample library, using spindle speed as the first key and workpiece material code as the second key. The inverted index is used to search for sets of normal future time-frequency trajectories with the same spindle speed and material code. If the set contains fewer than three trajectories, the search continues with a 5% increase in both the spindle speed range and the range below the spindle speed until at least three trajectories are found. For each retrieved normal future time-frequency trajectory, the corresponding feed rate, axial depth of cut, radial depth of cut, and number of teeth are extracted. The differences between these parameters and the current condition parameters are calculated item by item. The weighted sum of the absolute values ​​of the differences is used as the condition distance value. The trajectories are then sorted from smallest to largest, and the distance is selected accordingly. The two normal future time-frequency trajectories with the smallest values ​​are selected as candidate reference trajectories. If the number of future time steps of one of the candidate reference trajectories is greater than the number of time steps of the effective future time-frequency trajectory, a time period of equal length to the effective future time-frequency trajectory is extracted from the beginning of the candidate reference trajectory. If the number of time steps is less than the number of time steps of the effective future time-frequency trajectory, the time steps are supplemented by linear extrapolation at the end of the candidate reference trajectory. The two extracted or supplemented candidate reference trajectories are matched one-to-one with the effective future time-frequency trajectory according to the time steps. The total energy and principal axis order energy difference between the three time-frequency maps at the same time step are calculated. The candidate reference trajectory with the smaller difference is determined as the final normal future time-frequency trajectory. Finally, the normal future time-frequency trajectory is paired one-to-one with the effective future time-frequency trajectory according to the future time sequence to form a normal evolution reference chain with the same length as the effective future time-frequency trajectory. For each future time step, the total energy features, main shaft order frequency band features, cutter tooth passage order frequency band features, bearing characteristic frequency band features, and structural mode frequency band features of the effective future time-frequency trajectory and the normal evolution reference chain in the corresponding time-frequency state are extracted. The deviation degree corresponding to the future time step is obtained based on the differences between the features. Specifically, obtaining the deviation degree corresponding to the future time step based on the differences between the features is as follows: In the time-frequency grid, a total energy statistical region and four energy statistical bands are preset: main shaft order, cutter tooth passing order, bearing characteristic frequency band, and structural mode frequency band. For the same future time step, the effective future time-frequency trajectory and the two time-frequency maps corresponding to the normal evolution reference chain at that time step are read sequentially. The pixel energy is accumulated in the five statistical regions respectively to obtain the total energy value of the two time-frequency maps and the energy values ​​of the four target frequency bands. The difference between each energy value of the effective trajectory and the corresponding energy value of the reference chain is calculated, and the absolute value of the difference is divided by the corresponding energy value of the reference chain to obtain the percentage. The five difference ratios are weighted and summed according to preset weights, with the total energy difference having the highest weight, followed by the spindle order and cutter tooth order differences, and the bearing and structural mode differences having relatively lower weights. The weighted sum is compared with a preset deviation threshold. If the result does not exceed the threshold, the time step is determined to be in the normal evolution range. If the result exceeds the threshold for the first time, the time step is recorded as the earliest deviation position, and the target frequency band with the largest difference at that time is marked as the abnormal starting frequency band. The weighted sum is stored as the deviation value of the time step. The deviation degree of each future time step is compared with the preset deviation threshold in the future time sequence. The future time step that first reaches the preset deviation threshold is determined as the earliest deviation time step. In the time-frequency state corresponding to the earliest deviation time step, the frequency band with the largest deviation degree is determined as the abnormal starting frequency band. Based on the earliest deviation time step, the abnormal starting frequency band, the deviation degree corresponding to each future time step, and the corresponding valid future time-frequency trajectory identifier, vibration evolution deviation results are generated. Specifically, the generation of vibration evolution deviation results is as follows: The earliest deviation time step is read from the actual machining cycle time and recorded as the abnormal start time. The abnormal start frequency band is matched with the predefined spindle order band, cutter tooth order band, bearing characteristic band, and structural mode band to obtain the target frequency band name to which the abnormal start frequency band belongs and is written into the abnormal start frequency band field. The deviation degree values ​​corresponding to each time step are written into the deviation curve vector in the order of future time steps. At the same time, the maximum value, mean value, and terminal value of the deviation degree are calculated and written into the deviation statistics field. Based on the maximum value and mean value of the deviation curve vector, the vibration deviation is judged according to the preset three threshold ranges of mild, moderate, and severe, and the judgment result is written into the risk level field. Finally, the effective future time-frequency trajectory identifier, abnormal start time, abnormal start frequency band, deviation curve vector, deviation statistics field, and risk level field are encapsulated into a vibration evolution deviation result record.

[0026] In this embodiment, the step of outputting machine tool vibration monitoring results based on the vibration evolution deviation results includes: The earliest deviation time step in the vibration evolution deviation result is converted into the corresponding actual time to obtain the anomaly initiation time. Based on the abnormal starting frequency band in the vibration evolution deviation results, the key frequency range to which the abnormal vibration belongs is determined, and the key frequency range is matched with the spindle order frequency band, the cutter tooth passing order frequency band, the bearing characteristic frequency band and the structural mode frequency band to obtain the abnormality type identifier; Based on the deviation change information corresponding to each future time step, the vibration deviation growth trend is calculated, and the vibration instability development trend is determined based on the vibration deviation growth trend. The risk level of machine tool vibration is determined based on the onset time of the anomaly, the anomaly type identifier, and the vibration instability development trend; The system integrates the abnormal start time, abnormal start frequency band, abnormal type identifier, vibration instability development trend, and risk level to generate and output machine tool vibration monitoring results.

[0027] refer to Figure 2 A machine tool vibration monitoring system based on time-frequency deep networks, comprising: The data acquisition and time-frequency construction module is used to acquire vibration signals, spindle speed signals and machining parameters, and process the vibration signals to construct a continuous time-frequency sequence. The future time-frequency trajectory generation module is used to extract vibration time-frequency evolution features based on continuous time-frequency sequences and processing condition parameters, generate multiple future time-frequency candidate trajectories, and construct a set of future time-frequency bifurcation trajectories; The trajectory credibility assessment module is used to input future time-frequency candidate trajectories into the improved GANomaly model to calculate trajectory credibility and filter to obtain the target future time-frequency trajectory set; The path generation verification module is used to verify the continuity of the generated path for the target's future time-frequency trajectory, calculate the credibility of the generated path, and filter out the valid future time-frequency trajectories. The evolutionary deviation analysis module is used to calculate the degree of deviation based on the effective future time-frequency trajectory and the normal evolution reference chain, determine the earliest deviation position, and generate vibration evolution deviation results. The monitoring result output module is used to output machine tool vibration monitoring results based on the vibration evolution deviation results.

[0028] Example 1: In a continuous milling cycle, a machine tool performs sidewall finishing on the same batch of metal workpieces. Three-axis accelerometers are installed at the spindle head, bearing housing, and bed. The system synchronously reads the spindle speed signal and machining parameters. A total of 3600 vibration segments are collected during this cycle, each lasting 0.2 seconds, with a sampling frequency of 20000Hz and 4000 sampling points per segment. The spindle speed ranges from 5200–5800 rpm, the feed rate from 420–510 mm / min, the axial depth of cut is 1.5 mm, the radial depth of cut is 1.2 mm, the tool has 4 teeth, and the workpiece material is coded as M2. Under these conditions, traditional monitoring methods typically only trigger an alarm after a significant increase in the root mean square value of vibration, easily missing the weak changes before tool wear, early bearing damage, or chatter formation.

[0029] After receiving the vibration signal, the system performs a bandpass filter on the original signal from 50 to 8000 Hz. Before filtering, the average signal-to-noise ratio of the vibration signal at the spindle box measuring point was 16.4 dB, which was improved to 21.1 dB after filtering. Subsequently, the system segments the vibration data according to a 0.2-second window and a 0.05-second sliding step, and standardizes each segment of vibration data. The system performs a short-time Fourier transform on each window to generate a time-frequency image of size 128×128. Twenty consecutive time-frequency images form a continuous time-frequency sequence. In a certain test sample, the total energy of the last time-frequency image in the continuous time-frequency sequence is 0.386, the spindle order band energy is 0.074, the cutter tooth passage order band energy is 0.106, the bearing characteristic frequency band energy is 0.032, and the structural mode frequency band energy is 0.041.

[0030] After a continuous time-frequency sequence is input into the future time-frequency trajectory generation network, the time-frequency sequence encoding unit extracts the total energy, energy centroid frequency, spindle order energy, cutter tooth passage order energy, bearing characteristic frequency band energy, and structural modal frequency band energy frame by frame, and calculates the energy change and frequency band drift between adjacent time-frequency maps. In this sample, the cutter tooth passage order energy increases from 0.082 to 0.106 within 20 frames, and the structural modal frequency band energy increases from 0.028 to 0.041. The operating condition parameter encoding unit reads the spindle speed (5520 rpm), feed rate (468 mm / min), axial depth of cut (1.5 mm), radial depth of cut (1.2 mm), number of cutter teeth (4), and material code M2, generating 16-dimensional operating condition parameter features, which are then fused with the time-frequency sequence features to obtain the vibration time-frequency evolution features.

[0031] The bifurcation trajectory generation unit sets up 5 trajectory branches, each generating a predicted time-frequency map for the next 10 time steps. Each time step corresponds to 0.05 seconds, so each future time-frequency candidate trajectory covers 0.5 seconds of the future. The first candidate trajectory shows stable vibration, with a total energy of 0.402 at step 10; the second candidate trajectory shows enhanced tooth order energy, with a tooth passage order energy of 0.151 at step 10; the third candidate trajectory shows enhanced structural mode frequency band, with a structural mode frequency band energy of 0.128 at step 10; the fourth candidate trajectory shows enhanced bearing characteristic frequency band, with a bearing characteristic frequency band energy of 0.071 at step 10; and the fifth candidate trajectory shows enhanced broadband impact, with a total energy of 0.487 at step 10. The trajectory probability calculation unit calculates the evolution probabilities of the 5 trajectories as 0.18, 0.31, 0.27, 0.09, and 0.15, respectively, and binds them with the corresponding candidate trajectories to form a set of future time-frequency bifurcation trajectories.

[0032] After the set of future time-frequency bifurcation trajectories is fed into the improved GANomaly model, each candidate trajectory serves as an input sample, and each sample consists of 10 predicted time-frequency maps. The first encoder outputs 64-dimensional first latent space features, the attention-enhanced decoder generates the reconstructed trajectory, and the second encoder outputs 64-dimensional second latent space features. For the second candidate trajectory, the trajectory reconstruction difference is 0.043, the latent space difference is 0.051, the frequency domain discrimination difference is 0.038, the time domain discrimination difference is 0.046, and the trajectory confidence is 0.82. For the fifth candidate trajectory, the trajectory reconstruction difference is 0.118, the latent space difference is 0.142, the frequency domain discrimination difference is 0.133, the time domain discrimination difference is 0.156, and the trajectory confidence is 0.54. The system sets the confidence threshold to 0.70 and retains the first, second, and third trajectories as the target set of future time-frequency trajectories.

[0033] The generated path verification module divides the target future time-frequency trajectory into 9 adjacent path segments and extracts time-frequency energy transfer, frequency coupling, and modal evolution features. In the second trajectory, the cutting tooth's order energy continuously increases from 0.111 in step 1 to 0.151 in step 10. Eight of the nine segments are marked as stable enhancements, with an energy transfer continuity score of 0.89. The maximum drift of the spindle order frequency band position is one frequency grid, with a frequency coupling continuity score of 0.84. The tool modal components show an enhancing trend in seven segments, with a modal evolution continuity score of 0.86. The weighted generated path confidence score is 0.86, which is higher than the threshold of 0.75, and is therefore determined to be a valid future time-frequency trajectory.

[0034] The evolution deviation analysis module retrieves normal trajectories matching the current operating condition from the normal vibration evolution sample library. The sample library contains 7200 normal evolution samples. The system retrieved 96 normal future trajectories within a range of spindle speed (±5%) of 5520 rpm, feed rate (±8%) of 468 mm / min, consistent depth of cut, consistent number of teeth, and consistent material code. After sorting by operating condition distance, the system selects the two closest candidate reference trajectories and aligns them step-by-step with the effective future time-frequency trajectory, ultimately forming a normal evolution reference chain. The deviation of the second effective trajectory in future steps 1 to 3 is 0.11, 0.16, and 0.19, respectively, below the deviation threshold of 0.25. In step 4, the deviation reaches 0.28, exceeding the threshold for the first time. Therefore, the system identifies step 4 as the earliest deviation position, corresponding to a relative time of 186.20 seconds after machining starts, with the abnormal initial frequency band being the tooth passage order frequency band.

[0035] In the same simulated machining cycle, the traditional STFT threshold method detected that the cutting tooth energy exceeded the alarm threshold at 188.65 seconds, the current time-frequency CNN classification method identified the tool abnormality at 187.90 seconds, while this invention determined the earliest deviation position at 186.20 seconds and issued a warning at 186.25 seconds. This invention is 2.40 seconds faster than the STFT threshold method and 1.65 seconds faster than the current time-frequency CNN classification method. Subsequent verification showed that the third tooth of the tool showed slight wear in this stage, with a measured wear width of approximately 0.08 mm, and the surface roughness increased from 0.82 micrometers to 1.14 micrometers, corresponding to the tool tooth order abnormality output by this invention.

[0036] In the test comparison, the training samples consisted of 9800 groups, including 7200 normal samples, 1200 tool wear samples, 800 early bearing damage samples, and 600 chatter development samples; the test samples consisted of 2600 groups. The anomaly identification accuracy of the traditional STFT threshold method was 76.8%, the current time-frequency CNN classification method was 84.5%, and this invention was 93.1%. The false positive rate for normal samples of the traditional STFT threshold method was 11.6%, the current time-frequency CNN classification method was 7.9%, and this invention was 3.8%. Regarding the accuracy of anomaly initiation frequency band localization, the traditional STFT threshold method was 61.4%, the current time-frequency CNN classification method was 72.6%, and this invention was 90.8%. Regarding the average error of the anomaly initiation time, the traditional STFT threshold method was 1.38 seconds, the current time-frequency CNN classification method was 0.81 seconds, and this invention was 0.29 seconds.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine tool vibration monitoring method based on time-frequency deep networks, characterized in that, include: The vibration signals, spindle speed signals, and machining parameters during machine tool operation are collected, and the vibration signals are processed to construct a continuous time-frequency sequence. The continuous time-frequency sequence and processing condition parameters are input into the future time-frequency trajectory generation network to extract vibration time-frequency evolution features, generate multiple future time-frequency candidate trajectories, calculate the evolution probability corresponding to each future time-frequency candidate trajectory, and construct a set of future time-frequency bifurcation trajectories based on the evolution probability. Each future time-frequency candidate trajectory in the future time-frequency bifurcation trajectory set is input into the improved GANomaly model to obtain the reconstructed trajectory and latent space features. The trajectory credibility index between each future time-frequency candidate trajectory and the reconstructed trajectory is calculated. The future time-frequency candidate trajectories are then screened to obtain the target future time-frequency trajectory set. The continuity of the generated path is verified for each target future time-frequency trajectory in the target future time-frequency trajectory set. Based on the time-frequency energy transfer relationship, frequency coupling relationship and modal evolution relationship of adjacent time steps, the credibility of the generated path is calculated. The target future time-frequency trajectory whose generated path credibility meets the preset conditions is determined as the valid future time-frequency trajectory. For each time-frequency state in the effective future time-frequency trajectory, a corresponding normal evolution reference chain is constructed. The degree of deviation between each time-frequency state and the corresponding normal evolution reference chain is calculated. The earliest deviation position in the effective future time-frequency trajectory is determined. Based on the earliest deviation position, the abnormal start time and abnormal start frequency band are determined, and vibration evolution deviation results are generated. The machine tool vibration monitoring results are output based on the vibration evolution deviation results.

2. The machine tool vibration monitoring method based on time-frequency deep network according to claim 1, characterized in that, The process of processing the vibration signal to construct a continuous time-frequency sequence includes: The vibration signal is bandpass filtered to obtain a filtered vibration signal; The filtered vibration signal is segmented according to a preset window length and a preset sliding step size to obtain multiple vibration signal segments; Each vibration signal segment is standardized to obtain a standardized vibration signal segment. Time-frequency transformation is performed on each standardized vibration signal segment to obtain the corresponding time-frequency diagram; The continuous time-frequency sequence is constructed by arranging the time-frequency diagrams according to the time sequence of each vibration signal segment.

3. The machine tool vibration monitoring method based on time-frequency deep network according to claim 1, characterized in that, The calculation of the evolution probability corresponding to each future time-frequency candidate trajectory, and the construction of a set of future time-frequency bifurcation trajectories based on the evolution probability, includes: A future time-frequency trajectory generation network is constructed, which includes a time-frequency sequence encoding unit, a working condition parameter encoding unit, a time-frequency working condition fusion unit, a bifurcation trajectory generation unit, and a trajectory probability calculation unit. The continuous time-frequency sequence is input into the time-frequency sequence encoding unit in chronological order to encode the vibration energy changes, frequency band position changes and time-frequency texture changes between adjacent time-frequency maps, thereby obtaining time-frequency sequence features that characterize the historical evolution of machine tool vibration. The machining condition parameters are input into the condition parameter encoding unit, and the spindle speed, feed rate, axial depth of cut, radial depth of cut, number of cutter teeth and workpiece material are encoded to obtain the condition parameter features that characterize the current machining state. The time-frequency sequence features and condition parameter features are input into the time-frequency condition fusion unit for fusion to obtain the vibration time-frequency evolution features. The vibration time-frequency evolution characteristics are input into the bifurcation trajectory generation unit, and multiple future time-frequency candidate trajectories are generated according to the preset future prediction step size. Each future time-frequency candidate trajectory consists of multiple predicted time-frequency maps arranged in future time order. Different future time-frequency candidate trajectories correspond to different evolution paths of machine tool vibration state in the future time range. Multiple future time-frequency candidate trajectories are input into the trajectory probability calculation unit. The overall trajectory features are extracted for each future time-frequency candidate trajectory. The corresponding trajectory score is generated based on the matching degree between the overall trajectory features and the vibration time-frequency evolution features. All trajectory scores are normalized so that the sum of the evolution probabilities corresponding to each future time-frequency candidate trajectory is one. Each future time-frequency candidate trajectory is bound to its corresponding evolution probability to construct a set of future time-frequency bifurcation trajectories.

4. The machine tool vibration monitoring method based on time-frequency deep network according to claim 1, characterized in that, The process of filtering future time-frequency candidate trajectories to obtain a set of target future time-frequency trajectories includes: An improved GANomaly model is constructed, which sequentially includes a first encoder, an attention-enhanced decoder, a second encoder, a frequency domain discriminator, and a time domain discriminator. The first encoder consists of three layers of residual convolutional blocks. A channel attention module is inserted after the second residual convolutional block. The channel attention module is used to highlight the time-frequency energy distribution related to the spindle order, the cutter tooth order, and the bearing characteristic frequency. Each future time-frequency candidate trajectory in the future time-frequency bifurcation trajectory set is used as an input sample. The input sample consists of several consecutive time-frequency images. It is input to the first encoder to obtain the first latent space features. The first latent space features are input to the attention enhancement decoder. The reconstructed trajectory is generated using the jump structure corresponding to the first encoder. The reconstructed trajectory is input to the second encoder to obtain the second latent space features. The future time-frequency candidate trajectory and the reconstructed trajectory are respectively input into the frequency domain discriminator to obtain the frequency domain discrimination result, and simultaneously input into the time domain discriminator to obtain the time domain discrimination result. The frequency domain discriminator adopts a spectrum block convolution structure, which focuses on detecting the energy difference of the main frequency band, and the time domain discriminator adopts a one-dimensional convolution fusion self-attention structure, which focuses on detecting the difference of impact vibration. The trajectory reconstruction difference, latent space difference, frequency domain discrimination difference, and time domain discrimination difference are calculated for each future time-frequency candidate trajectory. The trajectory credibility index is obtained by weighting the four differences. The difference weighting coefficients are automatically searched in advance through the normal machine tool vibration sample training process. During the training process, a combination loss function of cross-entropy loss, reconstruction loss, latent space consistency loss, frequency band preservation loss, and energy balance loss is used. The optimizer is Adam with weighted decay. All future time-frequency candidate trajectories are sorted from high to low according to the trajectory confidence index, and the future time-frequency candidate trajectories with a trajectory confidence index higher than the set threshold are selected as the target future time-frequency trajectory.

5. The machine tool vibration monitoring method based on time-frequency deep network according to claim 1, characterized in that, The calculation of path reliability, and the determination of target future time-frequency trajectories whose path reliability meets preset conditions as valid future time-frequency trajectories, includes: Obtain the set of future time-frequency trajectories of the target, divide each future time-frequency trajectory of the target into multiple consecutive time-frequency states according to the future time sequence, and determine the transformation relationship between two adjacent consecutive time-frequency states as a vibration generation path segment; For each vibration generation path segment, the total energy, spindle order energy, cutter tooth passage order energy, bearing characteristic frequency band energy, and structural modal frequency band energy are extracted from the previous and subsequent time-frequency states. Based on the energy change of the subsequent time-frequency state relative to the previous time-frequency state, the corresponding time-frequency energy transfer characteristics are obtained. For each vibration generation path segment, the spindle order frequency band, the cutter tooth passage order frequency band, the bearing characteristic frequency band, and the structural mode frequency band are extracted from the previous and next time-frequency states, and the corresponding frequency coupling characteristics are obtained based on the position offset, bandwidth change, and energy ratio change of each frequency band in adjacent time-frequency states. For each vibration generation path segment, the spindle modal components, tool modal components, bearing modal components, and structural modal components are extracted from the previous and next time-frequency states. Based on the enhancement, attenuation, and transfer of each modal component in adjacent time-frequency states, the corresponding modal evolution characteristics are obtained. The time-frequency energy migration features, frequency coupling features, and modal evolution features corresponding to each vibration generation path segment in the future time-frequency trajectory of the same target are input into the generation path credibility calculation unit. The generation path credibility calculation unit calculates the energy migration continuity score, frequency coupling continuity score, and modal evolution continuity score respectively. The energy migration continuity score, frequency coupling continuity score, and modal evolution continuity score are weighted and fused to obtain the generation path credibility corresponding to the future time-frequency trajectory of the target. The credibility of the generated path corresponding to the future time-frequency trajectory of each target is compared with the preset path credibility threshold. The future time-frequency trajectory of the target with a generated path credibility not lower than the preset path credibility threshold is determined as the valid future time-frequency trajectory.

6. The machine tool vibration monitoring method based on time-frequency deep network according to claim 1, characterized in that, The generated vibration evolution deviation results include: Obtain valid future time-frequency trajectories, split each valid future time-frequency trajectory into multiple future time-frequency states according to future time sequence, and record the future time steps corresponding to each future time-frequency state; Based on the machining condition parameters, select normal future time-frequency trajectories that match the current spindle speed, feed rate, axial depth of cut, radial depth of cut, number of cutter teeth, and workpiece material code from the normal vibration evolution sample library. Align the normal future time-frequency trajectories with the effective future time-frequency trajectories step by step in chronological order to form a normal evolution reference chain. For each future time step, the total energy features, main shaft order frequency band features, cutter tooth passage order frequency band features, bearing feature frequency band features, and structural mode frequency band features of the effective future time-frequency trajectory and the normal evolution reference chain in the corresponding time-frequency state are extracted respectively. The deviation degree corresponding to the future time step is obtained according to the differences between the features. The deviation degree of each future time step is compared with the preset deviation threshold in the future time sequence. The future time step that first reaches the preset deviation threshold is determined as the earliest deviation time step. In the time-frequency state corresponding to the earliest deviation time step, the frequency band with the largest deviation degree is determined as the abnormal starting frequency band. The vibration evolution deviation results are generated based on the earliest deviation time step, the abnormal starting frequency band, the degree of deviation corresponding to each future time step, and the corresponding effective future time-frequency trajectory identifier.

7. The machine tool vibration monitoring method based on time-frequency deep network according to claim 1, characterized in that, The step of outputting machine tool vibration monitoring results based on the vibration evolution deviation results includes: The earliest deviation time step in the vibration evolution deviation result is converted into the corresponding actual time to obtain the anomaly initiation time. Based on the abnormal starting frequency band in the vibration evolution deviation results, the key frequency range to which the abnormal vibration belongs is determined, and the key frequency range is matched with the spindle order frequency band, the cutter tooth passing order frequency band, the bearing characteristic frequency band and the structural mode frequency band to obtain the abnormality type identifier; Based on the deviation change information corresponding to each future time step, the vibration deviation growth trend is calculated, and the vibration instability development trend is determined based on the vibration deviation growth trend. The risk level of machine tool vibration is determined based on the onset time of the anomaly, the anomaly type identifier, and the vibration instability development trend; The system integrates the abnormal start time, abnormal start frequency band, abnormal type identifier, vibration instability development trend, and risk level to generate and output machine tool vibration monitoring results.

8. A machine tool vibration monitoring system based on time-frequency deep networks, comprising executing the machine tool vibration monitoring method based on time-frequency deep networks as described in any one of claims 1 to 7, characterized in that, include: The data acquisition and time-frequency construction module is used to acquire vibration signals, spindle speed signals and machining parameters, and process the vibration signals to construct a continuous time-frequency sequence. The future time-frequency trajectory generation module is used to extract vibration time-frequency evolution features based on continuous time-frequency sequences and processing condition parameters, generate multiple future time-frequency candidate trajectories, and construct a set of future time-frequency bifurcation trajectories; The trajectory credibility assessment module is used to input future time-frequency candidate trajectories into the improved GANomaly model to calculate trajectory credibility and filter to obtain the target future time-frequency trajectory set; The path generation verification module is used to verify the continuity of the generated path for the target's future time-frequency trajectory, calculate the credibility of the generated path, and filter out the valid future time-frequency trajectories. The evolutionary deviation analysis module is used to calculate the degree of deviation based on the effective future time-frequency trajectory and the normal evolution reference chain, determine the earliest deviation position, and generate vibration evolution deviation results. The monitoring result output module is used to output machine tool vibration monitoring results based on the vibration evolution deviation results.

Images