A fault monitoring method and system for a numerical control machine tool

Abstract

The application provides a fault monitoring method and system for a numerical control machine tool, and particularly relates to the field of numerical control machine tool fault monitoring; by uniformly collecting and time window slicing processing of multi-channel running state time sequence data, a machine tool fault monitoring training set and a test set are constructed, and the dynamic correlation relationship of each monitoring channel under the same time scale is ensured; a spectral response value is constructed based on the ratio of the change amplitude of adjacent moments in the sliding window to the channel mean amplitude value, and a disturbance response value is formed by normalization compression; combined with the disturbance deviation intensity of the driving channel and the change direction sign of the response channel, a directional dependence structure between channels is analyzed and formed; further, a reverse interference reconstruction value and a strong residual quantity are constructed to form a channel-level fusion indicator, and a path tension indicator is constructed by sample difference extraction; on this basis, a path abnormality indicator is generated and weighted, and a machine tool fault detection value is output, so that stable detection and time sequence modeling of the fault state in the running process of the numerical control machine tool are realized.

Description

Technical Field

[0001] This invention belongs to the field of CNC machine tool fault monitoring, and specifically relates to a fault monitoring method and system for CNC machine tools. Background Technology

[0002] During long-term, high-load, and multi-condition operation, CNC machine tools are prone to vibration, temperature rise, and servo malfunctions in key components such as the spindle and feed system. If monitoring is not timely, this can easily lead to a decrease in machining accuracy and unplanned downtime. Existing fault monitoring methods mostly rely on manual inspection or judgment based on single signal thresholds and empirical characteristics, which are difficult to adapt to frequent changes in operating conditions and the effects of multi-source signal coupling. They are prone to false alarms and missed alarms, and their ability to identify early, weak faults and gradual degradation is insufficient. Therefore, it is necessary to propose a fault monitoring method for CNC machine tools suitable for complex operating conditions to improve the reliability of monitoring and the effectiveness of engineering applications.

[0003] Publication No. CN117193164A uses temperature data during the operation of a CNC machine tool as the core monitoring object. It continuously collects the current operating temperature sequence and simultaneously calls multiple historical temperature sequences generated under normal and historical stable operating conditions. By performing time alignment, outlier removal, and feature normalization on the current temperature sequence and each historical temperature sequence, it constructs an operating state similarity index and a time correlation index between the two. Based on this, a weighted evaluation mechanism is introduced to comprehensively assess the reference value of different historical temperature sequences, forming a comprehensive fault degree index to characterize the current operating state. When the comprehensive fault degree index exceeds a preset threshold, it is determined that the CNC machine tool has abnormal operation and a fault. The method detects and outputs corresponding alarms or monitoring results to continuously monitor and identify the operating status of CNC machine tools. Publication No. CN116009480A introduces a fault monitoring approach that combines signal processing and intelligent discrimination. The method takes the vibration, current, and acoustic state signals collected during the operation of the CNC machine tool as input. By preprocessing the original signals and extracting time-domain, frequency-domain, or time-frequency-domain features, a set of feature parameters that can characterize the health status of the machine tool is constructed. On this basis, a fault discrimination model is introduced to identify and classify the operating status of the machine tool. The discrimination model is a pattern recognition-based analysis model that can realize fault monitoring and anomaly identification by comparing the characteristics of normal and fault states.

[0004] However, compared with these existing technologies, there are still many shortcomings: First, many methods are sensitive to changes in operating conditions, and the model or threshold is prone to mismatch when crossing operating conditions, leading to false alarms / false negatives; therefore, some studies have specifically discussed the diagnosis problem across speed / operating conditions, which reflects the long-standing existence of this pain point; Second, signals are often affected by strong noise and coupling interference, and fault characteristics are easily submerged, especially early weak faults, which require more complex denoising, time-frequency transformation or model structure to be stably identified, thus raising the threshold for engineering implementation; Third, at the engineering implementation level, although multi-source sensing solutions can increase the amount of information, they bring sensor deployment / acquisition synchronization / calibration and maintenance costs, and multi-source fusion often relies on a large amount of experimental and labeled data to select features, set thresholds or train models, and recalibration is still required when migrating to different machine tools / different production lines; at the same time, some standard-based vibration assessments provide more limits and guidance for vibration level evaluation, but do not directly solve the location and diagnosis of component-level fault mechanisms, resulting in situations where abnormalities can be detected but are difficult to interpret / locate. Summary of the Invention

[0005] This invention proposes a fault monitoring method for CNC machine tools. It involves uniformly collecting and processing time-series data of the multi-channel operation status of the CNC machine tool through time window slicing to construct a machine tool fault monitoring training set and a machine tool fault monitoring test set, ensuring the dynamic correlation of each monitoring channel at the same time scale. Based on this, a spectral response value is constructed based on the ratio of the change amplitude of adjacent moments within the sliding window to the channel mean amplitude. This is then normalized and compressed to form a disturbance response value. Further, by combining the disturbance deviation intensity of the driving channel with the change direction indicator of the response channel, a driving relationship matrix characterizing the directional dependence structure between channels is formed. Subsequently, anisotropic interference reconstruction values ​​are constructed based on the disturbance differences and driving intensity between channels. A strong residual is generated from the offset and range between the disturbance response value and the anisotropic interference reconstruction value. Path weights are constructed by unifying the amplitude of the anisotropic interference reconstruction values. The system first combines multiple samples to form a channel-level fusion indicator. Then, it extracts differences between adjacent samples to construct fusion indicator transition values. This transition value is then combined with channel index distance and propagation constraints to form anisotropic response tension, which is adjusted by folded load to obtain a path tension indicator. Based on this, the system constructs temporal transition tension values ​​using the relationship between the path tension indicator and historical samples. Furthermore, it introduces the temporal transition tension of other channels within the same sample for cross-channel difference aggregation, which is then mapped using anisotropic adjustment coefficients to form a path anomaly indicator. Finally, the path anomaly indicators of each channel are weighted and aggregated according to channel anomaly sensitivity weights, and machine tool fault detection values ​​are obtained through nonlinear compression mapping. The system then uses a logarithmic loss function to train the machine tool fault monitoring model, achieving stable detection and temporal modeling of fault states during CNC machine tool operation.

[0006] A fault monitoring method for CNC machine tools, the specific method is as follows:

[0007] S1. Collect CNC machine tool operation data and preprocess it to obtain machine tool fault monitoring training set and machine tool fault monitoring test set;

[0008] S2. Construct spectral response values ​​by comparing the change amplitude of adjacent times within the sliding window of each channel with the channel mean amplitude. Then, normalize and compress the spectral response values ​​by mean and standard deviation in the sample dimension to form perturbation response values. Combine the perturbation deviation intensity of the driving channel with the change direction of the response channel for sample-level fusion and take the average to finally obtain the driving relationship matrix that characterizes the directional dependence structure between channels.

[0009] S3. Divide the disturbance response value into channels, construct the anisotropic interference reconstruction value based on the difference amplitude between channels and the driving relationship matrix, and then form a strong residual by the offset and range between the disturbance response value and the anisotropic interference reconstruction value. Construct the path weight by unifying the amplitude of the anisotropic interference reconstruction value, and combine the path weight with the strong residual to obtain the channel-level fusion indicator.

[0010] S4. Extract the difference between adjacent samples from the fusion indicator, construct the difference amplitude comparison structure, and use the exponential form to form the fusion indicator transition value. Calculate the transition difference between the current channel and other channels, construct the propagation constraint in combination with the channel index distance, and form the anisotropic response tension after convergence. Introduce the folded load adjustment coefficient to compress and adjust the anisotropic response tension, and construct the path tension indicator.

[0011] S5. Construct the temporal transition tension value of each channel by the relationship between the path tension indicator changes of adjacent samples and historical samples, and introduce the temporal transition tension values ​​of other channels in the same sample to construct cross-channel transition differences. Combine the channel index distance to form a structural discrete weight, and accumulate the difference results to obtain the anisotropic transition tension. Introduce the anisotropic adjustment coefficient to constrain the anisotropic transition tension and form the path anomaly indicator.

[0012] S6. Construct a machine tool fault monitoring model, input the machine tool fault monitoring training set, combine the logarithmic loss function, and sequentially go through steps S2 to S5. Iteratively train the machine tool fault monitoring model until convergence, and use the machine tool fault monitoring test set to test the machine tool fault monitoring model to realize fault detection of CNC machine tools.

[0013] Preferably, in step S1, the CNC machine tool fault monitoring dataset is jointly obtained by online operation monitoring data during actual production and processing, equipment no-load and load operation test data, and historical long-term operation record data. A complete time-series sample system is constructed by utilizing the complementarity of different operating condition data in the time dimension. Operating status monitoring points are set up at the spindle system, feed axis servo system, key positions of the guide rail, tool magazine mechanism, and machine tool electronic control unit. Data such as spindle speed, spindle drive current, X-axis servo motor current, Y-axis servo motor current, Z-axis servo motor current, X-axis feed speed, Y-axis feed speed, Z-axis feed speed, X-axis position deviation, Y-axis position deviation, Z-axis position deviation, spindle bearing temperature, spindle motor temperature, X-axis servo motor temperature, Y-axis servo motor temperature, Z-axis servo motor temperature, and guide rail temperature are collected. The temperature of the rail section and the temperature and humidity parameters of the machine tool's operating environment are collected and connected to a unified data acquisition platform via the CNC system interface and sensor acquisition unit. The sampling frequency is set to 1 time per second to continuously collect various operating parameters. At the same time, the machine tool number, shaft identification, machining condition identification and corresponding timestamp are recorded synchronously to form multi-source operating status time series raw data. After all the collected multi-source operating status time series raw data are transmitted to the central server, time alignment, invalid segment removal, outlier cleaning based on physical constraints and statistical criteria, missing data imputation, dimension unification and numerical normalization are performed in sequence. A fixed-length time window is used to slice the continuous time series data, and each time series segment containing 200 continuous sampling points is used as an operating status sample for subsequent fault monitoring and status discrimination method modeling and verification.

[0014] Furthermore, in step S2, the change amplitude of each channel at two adjacent moments within the sliding window is calculated to form a disturbance intensity term; the amplitude ratio of each channel value to the current channel mean is extracted, and after introducing a stabilization factor, nonlinear compression is performed to form an amplitude suppression factor; the disturbance intensity term and the amplitude suppression factor are calculated point by point, and the average of all calculation results within the window is calculated to form the spectral response value.

[0015] Based on the spectral response value of each channel, the mean and standard deviation of all samples under the current channel are calculated as an overall reference. The spectral response value of a single sample is compared with the mean of the current channel to construct the deviation. The standard deviation is proportionally processed and nonlinearly transformed to form a reciprocal suppression structure, thus forming a perturbation response value.

[0016] For the perturbation response value of each channel, the deviation of the driving channel from its own mean in the current sample is calculated, and the driving deviation intensity is constructed by combining the standard deviation of the whole sample. At the same time, the direction of change of the response channel relative to its mean in the current sample is extracted to generate the response direction label. After fusing the driving deviation intensity and the response direction label, the fusion result under all samples is averaged to obtain the driving relationship matrix that includes the directional influence between all channels.

[0017] Preferably, by fusing the perturbation intensity term and the amplitude suppression factor to construct the spectral response value, the statistical offset caused by the amplitude range of the signal can be suppressed while preserving the temporal variation characteristics of the monitored signal. This allows the spectral response value to stably reflect the dynamic variation characteristics of different channels within the sliding window. Based on this, the mean and standard deviation of the entire sample are introduced to normalize and nonlinearly compress the spectral response value, constructing a perturbation response value to achieve a unified expression of the amplitude of different channels under different operating conditions. Furthermore, the perturbation response value is used to calculate the driving deviation intensity, and combined with the response direction indicator to form the single-sample driving contribution between the driving channel and the response channel. Finally, a driving relationship matrix is ​​constructed to describe the directional dependency structure between all channel pairs, ensuring that the multi-channel time series modeling process has good stability, comparability, and scalability, and providing clear data support for anomaly propagation modeling and fault chain identification.

[0018] Furthermore, in step S3, taking the target channel in the current sample as the benchmark, the difference amplitude between its perturbation response value and that of other candidate channels is calculated sequentially. The difference amplitude is then exponentially modulated using the driving relationship matrix to obtain the interferometric modulation weight. The perturbation response value of the candidate channel is then weighted and corrected using the interferometric modulation weight. The value with the largest correction result among all candidates is selected as the anisotropic interferometric reconstruction value of the target channel in the current sample.

[0019] The maximum and minimum values ​​of the anisotropic interferometric reconstruction values ​​are extracted from the candidate channel set to form the range as the amplitude adjustment term. The difference between the disturbance response value and the anisotropic interferometric reconstruction value is calculated on the current sample channel. A sign term is constructed based on the offset direction sign. The difference value is mapped and used in the calculation of the sign term. Finally, it is combined with the amplitude adjustment term to obtain the strong residual.

[0020] Taking a single channel as the calculation object, the anisotropic interferometric reconstruction values ​​corresponding to the other channels are extracted in sequence, and the amplitude of the anisotropic interferometric reconstruction values ​​is uniformized to form path weights. The path weights are multiplied with the strong residuals of the corresponding channels and all channels are fused to form a total variable indicator. Finally, the strong residuals of the current channel are jointly calculated to obtain the channel-level fused indicator.

[0021] Preferably, during the operation of a CNC machine tool, the disturbance responses of each channel often exhibit inconsistent directions and asynchronous amplitudes. Relying solely on a single disturbance response value is insufficient to reflect the true structural anomaly state. By constructing anisotropic interference reconstruction values ​​based on the target channel, the disturbance influence of candidate channels on the target channel can be unified under the same comparison scale, thus clearly characterizing the asymmetric disturbance relationship between channels. Furthermore, by combining the range of the anisotropic interference reconstruction values ​​with the offset direction of the disturbance response values ​​to construct a strong residual, it helps to highlight the structurally significant amplitude deviation in the disturbance response. On this basis, anisotropic interference reconstruction values ​​are introduced to form path weights, and the path weights are fused with the strong residual, and then jointly calculated with the strong residual of the current channel itself. While retaining the total variable disturbance correlation information, a channel-level fusion indicator is output, thereby providing a reliable basis for the stable characterization of the anomaly state of key channels of the CNC machine tool.

[0022] Furthermore, in step S4, based on the channel-level fusion indicator, the change relationship of the same channel in adjacent samples is extracted, and a comparison structure of the current difference amplitude and the historical difference amplitude is constructed. On this basis, the current difference amplitude is embedded into an exponential interference modulation structure for modulation processing, and the modulation result is combined with the comparison structure to form a fusion indicator transition value that simultaneously contains the relative change intensity and modulation constraint characteristics.

[0023] Using the current channel as a benchmark, the differences in the fusion indicator transition values ​​of other channels are compared one by one, and cross-channel propagation attenuation constraints are introduced in combination with the channel index distance. On this basis, the directional driving modulation of the difference relationship between each channel is carried out, and the modulated difference results are converged to form the anisotropic response tension of the current channel.

[0024] Using the anisotropic response tension of the channel as the input, a nonlinear compression mapping relationship is constructed by introducing a folding load adjustment coefficient. The anisotropic response tension is proportionally adjusted to the anisotropic response tension compressed by the folding load adjustment coefficient to form a path tension indicator.

[0025] Preferably, by constructing the fusion indicator transition value and generating the anisotropic response tension, a unified modeling of the internal change trend and the difference pattern between channels is achieved. This not only retains the local change information of the current channel among continuous samples, but also introduces the difference propagation law between channels, enhancing the perception of the coordinated change between channels. Furthermore, by combining the folding load adjustment coefficient to form the path tension indicator, the relative change trend is retained while compressing the difference response within the channel. This enables the overall modeling structure to have nonlinear adjustment capability and directional discrimination capability, effectively improving the response sensitivity and expression stability of the subsequent structure in the multi-channel feature fusion process.

[0026] Furthermore, in step S5, the path tension indicator is constructed based on the path tension indicator. The change in the path tension indicator between the current sample channel and the previous sample channel is extracted, and the change in the path tension indicator of adjacent historical samples is calculated. The current trend and the historical trend are compared differentially. On the basis of eliminating the influence of the change direction, a normalization adjustment factor is introduced to perform dynamic scale constraint, thereby forming the temporal transition tension value of the channel in the current sample.

[0027] Using the temporal transition tension value of the current channel as input, the temporal transition tension values ​​of other channels in the same sample are introduced in sequence to construct the cross-channel difference in transition intensity of different channels. Under the constraint of the temporal transition tension value of the channel itself, the cross-channel difference is proportionalized to construct the relative transition deviation ratio. The channel index distance forms the structural discrete weight. The relative transition deviation ratio and structural discrete weight of all channels are weighted and accumulated to obtain the anisotropic transition tension.

[0028] An anisotropic adjustment coefficient corresponding to the channel is introduced to adjust the anisotropic transition tension of the channel under the current sample. The adjustment result and the anisotropic transition tension are combined to form a ratio that includes the original strength information and structural adaptability, thus obtaining the path anomaly indicator.

[0029] Preferably, by introducing temporal transition modeling of path tension indicators, anomaly determination no longer relies on the state at a single moment, but can reflect the evolutionary and abrupt characteristics of the channel in continuous samples, which is beneficial for early identification of the gradually accumulating anomaly trend during the operation of CNC machine tools. Based on the temporal transition tension value, anisotropic transition tension is constructed to uniformly express the transition imbalance state of different channels in the same sample, making structural anomalies under multi-channel collaborative operation conditions clearer and more identifiable. Furthermore, by combining the anisotropic adjustment coefficient to form a path anomaly indicator, different channels can obtain a consistent anomaly measurement benchmark under different load levels, different processing stages, and different structural conditions, thereby improving the stability, comparability, and continuity of anomaly representation under complex working conditions, and enhancing the ability to characterize the degree of anomaly evolution and the reliability of judgment during CNC machine tool fault monitoring.

[0030] Furthermore, in step S6, the path anomaly indications of each channel are weighted and aggregated according to the channel anomaly sensitivity weight, and the machine tool fault detection probability is obtained through nonlinear compression. The consistency between the detection result and the real fault label is constrained by the logarithmic loss function.

[0031] Input the machine tool fault monitoring training set, combine the logarithmic loss function, set the hyperparameters, construct the machine tool fault monitoring model, train the machine tool fault monitoring model until convergence, and use the machine tool fault monitoring test set to test the machine tool fault monitoring model to realize fault detection of CNC machine tools.

[0032] Preferably, by weighting and aggregating the path anomaly indicators of each channel under the constraint of channel anomaly sensitivity weights, the anomaly contributions of different channels in the overall fault evolution process of the machine tool are expressed differently, avoiding the structural information weakening and anomaly masking problems caused by the simple superposition of multiple channel anomalies. By using nonlinear compression mapping, the cumulative effect of multi-channel path anomalies is uniformly constrained within a stable numerical range, effectively suppressing the risk of numerical inflation during the anomaly accumulation process, and ensuring the continuity and comparability of machine tool fault monitoring output values ​​at different operating stages. Combined with the logarithmic loss function, the consistency between the machine tool fault detection value and the real fault label is constrained, and the channel anomaly sensitivity weights are adaptively adjusted during training, strengthening the response capability to key anomaly channels and reducing the interference of redundant channels on the monitoring results. This achieves stable fusion and reliable modeling of multi-channel path anomaly information during CNC machine tool operation, and improves the discrimination stability and detection consistency of the machine tool fault monitoring model under complex working conditions.

[0033] In the above technical solution, the technical effects and advantages provided by the present invention are as follows: The present invention constructs a multi-channel structured anomaly representation system that evolves stepwise from disturbance response value, driving relationship matrix, fusion indicator, path tension indicator, temporal transition tension and path anomaly indicator. This system enables continuous modeling of the multi-channel disturbance correlation and anomaly propagation path during the operation of CNC machine tools. It allows the anomaly coupling characteristics between different components and different monitoring channels inside the machine tool to be stably described under a unified structural framework, thereby significantly improving the overall perception capability of the machine tool fault evolution process under complex working conditions.

[0034] This invention introduces a fusion indicator transition value, anisotropic response tension, and folded load adjustment mechanism to jointly constrain the temporal variation characteristics within a channel and the anisotropic difference information between channels. This ensures that the path tension indicator and the path anomaly indicator maintain continuity and comparability in terms of numerical scale and structural expression, effectively suppressing the structural information weakening problem caused by the simple superposition of multiple channel anomalies, and enhancing the stability of identifying early weak faults and progressive anomalies.

[0035] This invention introduces channel anomaly sensitive weights into the path anomaly indicator for weighted aggregation, and combines nonlinear compression mapping and logarithmic loss function to complete model training. This enables machine tool fault detection values ​​to maintain a uniform probability expression form under different operating stages and load conditions. While ensuring consistency in discrimination, it improves the model's adaptive attention to key anomaly channels, thereby achieving reliable monitoring and robust modeling of fault states during CNC machine tool operation. Attached Figure Description

[0036] Figure 1 This is a flowchart of a fault monitoring method for CNC machine tools provided by the present invention;

[0037] Figure 2 This is a structural diagram of the driving relationship matrix provided by the present invention;

[0038] Figure 3 This is a structural diagram of the fusion indicator provided by the present invention;

[0039] Figure 4 This is a structural diagram of the path tension indicator provided by the present invention;

[0040] Figure 5 This is a structural diagram of the path anomaly indicator provided by the present invention;

[0041] Figure 6 This is a loss graph of the logarithmic loss function provided by this invention during the training process;

[0042] Figure 7 This is a distribution diagram of path anomaly indicators provided by the present invention;

[0043] Figure 8 This is a comparison chart of the detected values ​​and the actual values ​​of the machine tool fault monitoring model provided by this invention. Detailed Implementation

[0044] This invention proposes a fault monitoring method for CNC machine tools. It involves uniformly collecting and processing time-series data of the multi-channel operation status of the CNC machine tool through time window slicing to construct a machine tool fault monitoring training set and a machine tool fault monitoring test set, ensuring the dynamic correlation of each monitoring channel at the same time scale. Based on this, a spectral response value is constructed based on the ratio of the change amplitude of adjacent moments within the sliding window to the channel mean amplitude. This is then normalized and compressed to form a disturbance response value. Further, by combining the disturbance deviation intensity of the driving channel with the change direction indicator of the response channel, a driving relationship matrix characterizing the directional dependence structure between channels is formed. Subsequently, anisotropic interference reconstruction values ​​are constructed based on the disturbance differences and driving intensity between channels. A strong residual is generated from the offset and range between the disturbance response value and the anisotropic interference reconstruction value. Path weights are constructed by unifying the amplitude of the anisotropic interference reconstruction values. The system first combines multiple samples to form a channel-level fusion indicator. Then, it extracts differences between adjacent samples to construct fusion indicator transition values. This transition value is then combined with channel index distance and propagation constraints to form anisotropic response tension, which is adjusted by folded load to obtain a path tension indicator. Based on this, the system constructs temporal transition tension values ​​using the relationship between the path tension indicator and historical samples. Furthermore, it introduces the temporal transition tension of other channels within the same sample for cross-channel difference aggregation, which is then mapped using anisotropic adjustment coefficients to form a path anomaly indicator. Finally, the path anomaly indicators of each channel are weighted and aggregated according to channel anomaly sensitivity weights, and machine tool fault detection values ​​are obtained through nonlinear compression mapping. The system then uses a logarithmic loss function to train the machine tool fault monitoring model, achieving stable detection and temporal modeling of fault states during CNC machine tool operation.

[0045] Please see Figure 1 As shown in the figure, a fault monitoring method for a CNC machine tool in this application embodiment has the following specific steps.

[0046] S1. Collect CNC machine tool operation data and preprocess it to obtain machine tool fault monitoring training set and machine tool fault monitoring test set.

[0047] In this embodiment, the fault monitoring data of CNC machine tools is not usually derived directly from experimental results under a single ideal working condition. Instead, it is obtained by combining online operation monitoring data during actual production and processing, equipment no-load and load operation test data, historical long-term operation record data, and comparative test data under manually set working conditions. A complete time-series sample covering different processing states and operating stages is constructed through the complementary fusion of multi-source operation data, thereby meeting the research and verification needs of CNC machine tool fault monitoring methods. Operational status monitoring points are deployed in the spindle box, feed axis servo motors, key positions of the guide rails, tool magazine mechanism, and inside the machine tool electrical control cabinet. Spindle speed, spindle drive current, X-axis servo motor current, and Y-axis servo motor current are collected at each monitoring point. The system collects multi-dimensional status information during machine tool operation, including servo motor current, Z-axis servo motor current, X-axis feed speed, Y-axis feed speed, Z-axis feed speed, X-axis position deviation, Y-axis position deviation, Z-axis position deviation, spindle bearing temperature, spindle motor temperature, X-axis servo motor temperature, Y-axis servo motor temperature, Z-axis servo motor temperature, guide rail temperature, and the temperature and humidity of the machine tool's operating environment. This data is then connected to a unified data acquisition module via the CNC system interface and an external sensor acquisition unit. The sampling frequency is set to once per second, continuously recording multi-dimensional status information during machine tool operation. Simultaneously, it records the machine tool number, axis identification, machining program number, machining condition identification, load status identification, and timestamp, forming multi-source operating status time-series raw data. All collected multi-source data... After the raw data of the operational status timing is transmitted to the central server, it is first filtered and time-aligned based on timestamps from different sources, eliminating duplicate records, equipment downtime for maintenance, and obvious communication interruptions. Then, based on the machine tool's rated speed range, current safety limit, temperature allowable range, and the statistical rule of three standard deviations, abnormal sampling points are jointly judged. Records that do not meet physical constraints or statistical distribution patterns are marked as abnormal and deleted from the dataset. Missing segments in the remaining data are processed. For segments with consecutive missing durations not exceeding a preset threshold, linear interpolation is used to fill in the missing data. Specifically, based on the time ratio between the missing time and two adjacent valid sampling times, adjacent valid sampling values ​​are interpolated. Weighted estimation is used to obtain the operating parameter values ​​at the missing time points. For segments with consecutive missing durations exceeding a threshold, they are directly removed from the dataset. After missing data processing, the units and dimensions of various operating parameters are converted and standardized: rotation speed is standardized to revolutions per minute, current to amperes, temperature to degrees Celsius, position deviation to millimeters, and feed rate to millimeters per minute. Based on the standardization of dimensions, the maximum and minimum reference values ​​of each operating parameter in the historical stable operating range are statistically analyzed. The operating parameters are then linearly normalized by subtracting the corresponding minimum reference value from the current parameter value and dividing by the difference between the maximum and minimum reference values, so that each operating parameter falls within a unified numerical range of 0 to 1.After normalization, the continuous time-series data is sliced ​​into fixed-length time windows. Each time window contains 200 consecutive sampling moments of operational data, used to characterize the local operational status of the CNC machine tool within the corresponding time interval. The machine tool fault monitoring model learns the mapping relationship between machine tool fault patterns in the past 200 seconds and machine tool faults in the next 30 seconds to achieve fault monitoring. Multiple time-series data segments are generated on the time axis using a fixed-step sliding window, ultimately constructing a total of 10,000 machine tool monitoring data points. 8,000 of these data points are used as the machine tool fault monitoring training set, and the remaining 2,000 samples are used as the machine tool fault monitoring test set, providing a structurally unified, temporally continuous, and scale-consistent data foundation for subsequent fault monitoring model training and performance verification.

[0048] S2. Construct spectral response values ​​by comparing the change amplitude of adjacent times within the sliding window of each channel with the channel mean amplitude. Then, normalize and compress the spectral response values ​​according to the mean and standard deviation in the sample dimension to form perturbation response values. Combine the perturbation deviation intensity of the driving channel with the change direction indicator of the response channel to perform sample-level fusion and take the average. Finally, obtain the driving relationship matrix that represents the directional dependence structure between channels.

[0049] Furthermore, in step S2, a driving relationship matrix is ​​established, the process of which is as follows: Figure 2 As shown, the specific steps for establishing the driving relationship matrix are as follows.

[0050] S21. Calculate the change amplitude of each channel at two adjacent moments within the sliding window to form the disturbance intensity term; extract the amplitude ratio of each channel value to the current channel mean, introduce a stabilization factor and perform nonlinear compression to form the amplitude suppression factor; perform point-by-point calculations on the disturbance intensity term and the amplitude suppression factor and average all calculation results within the window to form the spectral response value.

[0051] In this embodiment, considering the characteristics of multi-source monitoring signals exhibiting insignificant amplitude changes but frequent local disturbances in the early fault stage during CNC machine tool operation, a spectral response construction method based on a combination of disturbance intensity and amplitude suppression at adjacent time points is proposed. By accumulating the amplitude changes of continuous sampling points within a sliding time window and introducing an amplitude suppression mechanism relative to the window mean, the dominant interference of high-amplitude channels on the overall characteristics is avoided, enabling the output results to more realistically reflect the activity level of potential abnormal disturbances in the machine tool's operating state, providing stable input for subsequent abnormal normalization mapping and channel-driven relationship modeling. The specific steps are as follows: First, define the disturbance intensity of channel k between adjacent time points t and t+1. Dynamic intensity reflects the magnitude of channel characteristics changes over time; the more frequent the changes, the larger the value. A window mean is introduced. Amplitude ratio term for reference The amplitude suppression factor is constructed by squaring the sample and eliminating the influence of the sign. When the signal amplitude deviates significantly from the window mean at a certain moment, the amplitude suppression factor will decrease, thereby reducing the weight on the overall spectral response. The disturbance intensity term and the amplitude suppression factor term are multiplied point by point to obtain the coupled response value at time step t. For all signals within the window... The coupling response values ​​are accumulated and averaged to obtain the spectral response value of channel k under the sample; the mathematical model of the spectral response value is:

[0052] ;

[0053] in, For the first The value of the k-th channel at the t-th time step within a time window; For the first The value of the k-th channel at the (t+1)-th time step within a time window; T represents all time steps within window k, with a value of 200; It is a very small positive number used for amplitude normalization stabilization, and its value is set to... ; For the first The sliding window mean of the k-th channel within a time window, its mathematical model is as follows: ; For the first The spectral response value of the k-th channel within a time window.

[0054] S22. Based on the spectral response value of each channel, calculate the mean and standard deviation of all samples in the current channel. As an overall reference, compare the spectral response value of a single sample with the mean of the current channel to construct the deviation. The standard deviation is proportionally processed and nonlinearly transformed to form a reciprocal suppression structure, thus forming a perturbation response value.

[0055] In this embodiment, to address the issue of amplitude scale differences in spectral response values ​​under different operating conditions, a normalized perturbation construction method based on inverse amplitude mapping is used. By performing centering and standardization operations on the spectral response values ​​of each channel, and constructing a compression mapping function based on the square of the deviation, a normalized measure of fluctuation intensity is achieved, effectively avoiding calculation biases caused by inconsistencies in mean and scale between channels. Specifically, to eliminate differences in the mean response values ​​of different channels, the deviation of the response from the mean of the current channel is calculated. Considering the fluctuations in channel scale, a standard deviation is introduced as a standardization factor to obtain the proportional deviation. To eliminate the influence of positive and negative signs and enhance the influence on outliers, the proportional deviation is squared to obtain the squared deviation term. Based on the squared deviation term, a reciprocal suppression structure is introduced to construct the perturbation response value; the mathematical model of the perturbation response value is as follows:

[0056] ;

[0057] in, Let K be the mean spectral response of channel k across all samples, and its mathematical model is: N is the total number of samples, with a value of 8000; Let K be the standard deviation of the spectral response of channel k across all samples, and its mathematical model is: ; To ensure the value is a very small positive number and avoid a denominator of zero, the value is set to... ; For the first The disturbance response value of the kth channel within a time window.

[0058] S23. For the perturbation response value of each channel, calculate the degree of deviation of the driving channel from its own mean in the current sample, and construct the driving deviation intensity by combining the standard deviation of the whole sample; at the same time, extract the direction of change of the response channel from its mean in the current sample and generate the response direction indicator; after fusing the driving deviation intensity and the response direction indicator, average the fusion result under all samples to obtain the driving relationship matrix containing the directional influence between all channels.

[0059] In this embodiment, based on the asymmetric influence characteristics between different monitoring channels of a CNC machine tool under abnormal conditions, a channel-to-channel directional driving relationship construction method is used. By constructing a sign change flag for the response channel and multiplying it sample-by-sample with the standardized deviation value of the driving channel, an asymmetric influence matrix reflecting the directional dependency is obtained. The specific steps are: calculating the directional dependency of driving channel p in the sample... The degree of deviation is taken as the driving strength of channel p, and the driving deviation strength is obtained. Its calculation method is as follows: By judging channel q in the sample The sign term is constructed relative to the direction of change of its mean to obtain the response direction indicator. Subsequently, the driving deviation intensity is multiplied by the response direction flag to obtain the directed driving contribution. The average of the sample-level directed driving contribution terms is then calculated to obtain the driving relationship matrix from channel p to channel q. The mathematical model of the driving relationship matrix is ​​as follows:

[0060] ;

[0061] in, For the first The disturbance response value of the qth channel within a time window; For the first The disturbance response value of the p-th channel within a time window; Let be the mean spectral response of channel q across all samples; This is a sign function, taking values ​​of -1, 0, and +1. For the first The driving relationship matrix of channel p to channel q within a time window.

[0062] S3. Divide the disturbance response value into channels, construct the anisotropic interference reconstruction value based on the difference amplitude between channels and the driving relationship matrix, and then form a strong residual by the offset and range between the disturbance response value and the anisotropic interference reconstruction value. Construct the path weight by unifying the amplitude of the anisotropic interference reconstruction value, and combine the path weight with the strong residual to obtain the channel-level fusion indicator.

[0063] Furthermore, in step S3, a fusion indicator is established, the process of which is as follows: Figure 3 As shown, the specific steps for establishing the fusion indicator are as follows.

[0064] S31. Using the target channel in the current sample as a benchmark, calculate the difference between its perturbation response value and that of other candidate channels in turn. Combine the driving relationship matrix to perform exponential modulation on the difference amplitude to obtain the interferometric modulation weight. Use the interferometric modulation weight to perform weighted correction on the perturbation response value of the candidate channel. Select the value with the largest correction result among all candidates as the anisotropic interferometric reconstruction value of the target channel in the current sample.

[0065] In this embodiment, during the operation of the CNC machine tool, there are complex coupling relationships between the spindle system, feed system, drive motor, and structural components. Different monitoring channels exhibit inconsistent disturbance amplitudes, propagation directions, and response timings under abnormal conditions. To avoid misjudging the structure by simply treating multi-channel disturbances as synchronous changes, this step addresses the issue of heterogeneity in disturbance directions between channels by introducing a disturbance reconstruction method based on anisotropic interference mechanisms. Based on the obtained driving relationships, a cross-channel disturbance difference term is constructed and direction-dependent exponential interference modulation is applied to achieve a structured reconstruction of the channel responses under multi-source disturbance interaction conditions, thereby forming a steady-state disturbance estimate that reflects the complex operating state of the CNC machine tool. The specific steps are as follows: For the sample... The target channel q is fixed, and candidate interference paths are constructed one by one among the remaining candidate channels p. For each channel pair p and q, the difference in their perturbation response values ​​under the current sample is calculated to characterize the degree of inconsistency in the perturbation propagation direction. Then, the driving relationship matrix of channel p to channel q is introduced. By embedding the driving strength and perturbation difference term together into an exponential decay structure, direction-dependent interferometric modulation weights are formed to suppress the influence of directionally mismatched channels on the current channel structure estimation. The modulated interferometric weights are then multiplied and fused with the perturbation response value of channel p to obtain candidate reconstructed perturbation values ​​based on anisotropic interferometry correction. Finally, in all... The result with the largest response after interferometric correction is selected from the candidate results and taken as the anisotropic interferometric reconstruction value of channel p in the current sample, thereby completing the steady-state perturbation estimation of the multi-channel perturbation structure; the mathematical model of the anisotropic interferometric reconstruction value is:

[0066] ;

[0067] in, For the first The disturbance response value of the qth channel within a time window; This is the interference modulation coefficient, used to control the degree of influence of disturbance differences on the interference intensity. It is initially set to 0.5 and can be set according to the actual situation, with a value range of 0.1 to 2.0, to limit abnormal amplification and reduction. For exponentiation; The maximum value operation is used to select the result with the largest response after interference correction; For the first The disturbance response value of the qth channel within a time window; For the sample The anisotropic interferometric reconstruction value of the lower channel p is used to characterize the structural steady-state disturbance estimation result under multi-channel directional interferometric conditions.

[0068] S32. Extract the maximum and minimum values ​​of the anisotropic interference reconstruction values ​​from the candidate channel set to form the range as the amplitude adjustment term. Calculate the difference between the disturbance response value and the anisotropic interference reconstruction value on the current sample channel. Construct a sign term based on the offset direction sign. Map the difference value and participate in the calculation of the sign term. Finally, combine it with the amplitude adjustment term to obtain the strong residual.

[0069] In this embodiment, based on the anisotropic deviation between the perturbation response value and the anisotropic interference reconstruction value, an enhancement mechanism combining channel range and direction cancellation suppression is proposed; a strong residual is formed by constructing the product response expression between the direction sign term and the range coefficient to capture structural deviations; the specific steps are as follows: firstly, in the sample The maximum and minimum values ​​of the anisotropic interference reconstruction values ​​are extracted from the candidate channel set, and their range is calculated as the sample amplitude adjustment term. For the target channel q, the difference between its disturbance response value and the anisotropic interference reconstruction value is calculated, and its direction sign is extracted to form a sign term. The difference is then squared to form a non-negative amplification term. The sign term is multiplied by the non-negative amplification term to form the directional enhancement residual. Finally, this residual is multiplied by the amplitude adjustment term to generate a strong residual, which describes the intensity of the channel's deviation from the current structure. The mathematical model of the strong residual is as follows:

[0070] ;

[0071] in, This is the sign function, which takes values ​​of −1, 0, and +1, and is used to describe the offset direction. For the sample The reconstructed value of the anisotropic interference in the lower channel q; For the sample Strong residuals on channel q.

[0072] S33. Taking a single channel as the calculation object, extract the anisotropic interferometric reconstruction values ​​corresponding to the other channels in sequence, and perform amplitude uniformization processing on the anisotropic interferometric reconstruction values ​​to form path weights. Multiply the path weights with the strong residuals of the corresponding channels and fuse all channels to form a total variable indicator. Finally, perform joint calculation with the strong residuals of the current channel itself to obtain the channel-level fused indicator.

[0073] In this embodiment, based on the anisotropic interference reconstruction value and strong residual enhancement value of candidate channel p and target channel q, the anisotropic interference relationship between channels is regarded as a total variable interaction structure, and the strong residual is regarded as the channel-level structural deviation intensity. For each channel k across the entire channel range, a fusion indicator is constructed; the specific steps are: read the anisotropic interference reconstruction value of channel p one by one. Strong residual with channel q ,use Will Transform the path weights into exponential type and introduce absolute values ​​to ensure the consistency of the interference intensity amplitude. Then, combine the path weights with... Perform product coupling to form an indicator for channel k, for all The indicators are accumulated to obtain the total variable indicator for channel k, and finally compared with the strong residual of channel q. Perform product union to output channel-level fused indicator. This ensures that each channel k obtains a fusion result jointly determined by the anisotropic interferometric reconstruction value and the strong residual; the mathematical model of the fusion indicator is:

[0074] ;

[0075] in, The modulation coefficient has an initial value of 2 and is adjusted according to the actual value required for the fusion indicator. The value range is from -5 to 5 to limit abnormal amplification and reduction. For exponential path weights; To perform absolute value operations on the reconstructed values ​​of the anisotropic interference; For the sample The channel-level fusion indicator for the middle channel k.

[0076] S4. Extract the differences between adjacent samples from the fusion indicator, construct a difference magnitude comparison structure, and use an exponential form to form the fusion indicator transition value. Calculate the transition difference between the current channel and other channels, construct propagation constraints in combination with channel index distance, and form an anisotropic response tension after convergence. Introduce a folded load adjustment coefficient to compress and adjust the anisotropic response tension, and construct the path tension indicator.

[0077] Furthermore, in step S4, the path tension indicator is established, and the process is as follows: Figure 4 As shown, the specific steps for establishing the path tension indicator are as follows.

[0078] S41. Based on the channel-level fusion indicator, extract the change relationship of the same channel in adjacent samples, construct a comparison structure between the current difference amplitude and the historical difference amplitude, and on this basis, embed the current difference amplitude into the exponential interference modulation structure for modulation processing, and combine the modulation result with the comparison structure to form a fusion indicator transition value that simultaneously contains the relative change intensity and modulation constraint characteristics.

[0079] In this embodiment, during the fault evolution process of the CNC machine tool, the channel-level fusion indication quantity corresponding to each monitoring channel is... Will follow the sample The advancement produces changes of varying magnitudes. Among them, the short-term fluctuations caused by transient load disturbances and the continuous changes caused by structural degradation have significant differences in transition amplitude. To distinguish between these two types of change characteristics, the specific steps are as follows: Under the condition that channel k is fixed, firstly, the sample index is used... , , Corresponding channel-level fusion indicator , , Calculate the current adjacent difference magnitude Difference range with historical adjacent areas The ratio of the former to the latter describes the amplification of the current channel state change relative to historical changes; a stability factor is added to the denominator of the ratio. This is used to suppress the numerical amplification effect caused by excessively small historical difference amplitudes; subsequently, an exponential modulation term is constructed based on the current difference amplitude. Through the interference modulation coefficient The abnormal amplification magnitude is suppressed, thereby forming a fusion indicator transition value that simultaneously reflects the relative strength of the transition and its stable regulation characteristics. The mathematical model for the fusion indicator transition value is as follows:

[0080] ;

[0081] in, For the sample Channel-level fusion indicator for channel k; For the sample Channel-level fusion indicator for channel k; The stabilization factor takes a value of ; For the sample The fusion indicator transition value of the middle channel k.

[0082] S42. Using the current channel as a benchmark, compare the differences in fusion indicator transition values ​​with other channels one by one, and introduce cross-channel propagation attenuation constraints in combination with channel index distance. On this basis, perform directional driving modulation on the difference relationship between each channel, and converge the modulated difference results to form the anisotropic response tension of the current channel.

[0083] In this embodiment, during the operation of the CNC machine tool, multiple channels are affected by disturbances from the same source abnormal signal, exhibiting a structurally enhanced trend. To avoid the non-identifiable propagation error caused by traditional similarity-based weighting, this step introduces a cross-channel tension-driven mechanism. It utilizes the difference between the convolution path and the index to construct propagation path weights, and combines this with the difference in fusion indicator transition values ​​to form anisotropic response values ​​in the sense of physical propagation, thereby describing the asymmetric disturbance diffusion behavior between channels. Specifically, the steps are: calculating the fusion indicator transition value between channel c and channel k. ,use Representing the physical index distance of channels, this enables long-range conduction attenuation modeling and is used to construct perturbed convolutional paths across channels. The farther the channel, the more its weights are affected. The inhibition, and then through Strengthening the response driving force in the dominant propagation direction, the final squared result converges to obtain the anisotropic response tension of channel k. The mathematical model for the anisotropic response tension is:

[0084] ;

[0085] in, For the first The driving relationship matrix of channel c to channel k within a time window; For the sample The fusion indicator transition value of the middle channel c; For propagation attenuation constraints, representing the relative index distance between channels, used to describe the propagation attenuation of the physical location of the channels; For the sample The anisotropic response tension of the middle channel k.

[0086] S43. Using the channel's anisotropic response tension as input, a nonlinear compression mapping relationship is constructed by introducing a folding load adjustment coefficient. The anisotropic response tension is proportionally adjusted to the anisotropic response tension compressed by the folding load adjustment coefficient to form a path tension indicator.

[0087] In this embodiment, under the complex operating conditions of a CNC machine tool, the channel anisotropic response tension undergoes nonlinear amplification during cross-path propagation due to the superposition of local anomalies, causing the fusion judgment to be overly sensitive to extreme disturbances. To address the high-amplitude instability of the channel anisotropic response tension, a folding load suppression mechanism is introduced. By constructing a controlled compression mapping relationship, a stable and comparable path tension indicator is formed. The specific steps are as follows: using the channel anisotropic response tension... As input, by introducing the folded load adjustment coefficient A proportional compression relationship is constructed so that the tension gradually enters the controlled range as the amplitude increases, thereby avoiding the disproportionate influence of high-amplitude anisotropic responses during path propagation. Folding mapping maintains response resolution in the low-amplitude range and suppresses excessive amplification in the high-amplitude range, ensuring that the tension results from different channels are comparable. The final output is the path tension indicator. While maintaining the information on anisotropic structural differences, it possesses stability and continuity, and can be directly used for multi-channel anomaly integrated analysis; the mathematical model of the path tension indicator is:

[0088] ;

[0089] in, This is the folding load adjustment coefficient, used to control the compressive strength of the anisotropic response tension in the high-amplitude range. The initial setting is 0.8, which can be adjusted according to the stability of equipment operation. The value range is from 0.5 to 1.5, limiting abnormal amplification and reduction. For the sample The path tension indicator of the middle channel k under cross-channel disturbance propagation.

[0090] S5. Construct the temporal transition tension value of each channel by the relationship between the path tension indicator changes of adjacent samples and historical samples, and introduce the temporal transition tension values ​​of other channels in the same sample to construct cross-channel transition differences. Combine the channel index distance to form a structural discrete weight, and perform weighted accumulation of the difference results to obtain the anisotropic transition tension. Introduce the anisotropic adjustment coefficient to constrain the anisotropic transition tension and form the path anomaly indicator.

[0091] Furthermore, in step S5, a path anomaly indicator is established, the process of which is as follows: Figure 5 As shown, the specific steps for establishing a path anomaly indicator are as follows.

[0092] S51. Based on the path tension indicator, the change in the path tension indicator between the current sample channel and the previous sample channel is extracted, and the change in the path tension indicator of adjacent historical samples is calculated. The current trend is compared with the historical trend by difference. On the basis of eliminating the influence of the change direction, a normalization adjustment factor is introduced to perform dynamic scale constraint, and the temporal transition tension value of the channel in the current sample is formed.

[0093] In this embodiment, during the continuous operation of the CNC machine tool, the channel path tension exhibits an evolutionary characteristic of both phased accumulation and sudden transitions in the sample sequence. Simply relying on the difference between adjacent samples is insufficient to distinguish between actual structural transitions and inertial continuation. This step uses the path tension indicator as input and introduces a differential comparison mechanism between current and historical changes, based on the tension change trajectory of the same channel in continuous samples, to construct a temporal transition tension value sensitive to abnormal path tension transitions. The specific steps are: by calculating... Extract the change in path tension of the current sample relative to the previous sample to describe the tension adjustment trend at the current stage; then calculate... The historical changes in path tension from the previous stage are constructed to describe the inertial evolution direction of the channel in recent operation. The difference between the two is the degree of deviation between the current trend and the historical inertial trend, which can be used to identify whether the machine tool channel has undergone an unexpected transition at the current moment. Taking the absolute value of the difference ensures that the machine tool fault monitoring model focuses only on the intensity of the change rather than the direction, enhancing the ability to identify sudden transitions. Furthermore, a normalized adjustment factor is introduced. The transition amplitude is dynamically standardized to impose amplitude scale constraints, forming a temporal transition tension value; the mathematical model of the temporal transition tension value is as follows:

[0094] ;

[0095] in, This is a tension stability factor used to prevent non-physical numerical amplification of path tension within the low-amplitude operating range, ensuring the stability of temporal transition tension in continuous operating scenarios. Its value is... ; For the sample The path tension indicator of the middle channel k under cross-channel disturbance propagation; For the sample The path tension indicator of the middle channel k under cross-channel disturbance propagation; For the sample The temporal transition tension value of the middle channel k.

[0096] S52. Taking the temporal transition tension value of the current channel as input, the temporal transition tension values ​​of the other channels in the same sample are introduced in sequence to construct the cross-channel difference in the transition intensity of different channels. Under the constraint of the temporal transition tension value of the channel itself, the cross-channel difference is proportionalized to construct the relative transition deviation ratio. The channel index distance forms the structural discrete weight. The relative transition deviation ratio and structural discrete weight of all channels are weighted and accumulated to obtain the anisotropic transition tension.

[0097] In this embodiment, during the multi-channel collaborative operation of a CNC machine tool, although the temporal transition tensions of each channel originate from the same processing state, under abnormal induced conditions, their transition amplitudes and evolution rhythms often exhibit obvious inconsistencies and directional deviations. To avoid symmetrical cancellation and linear weakening caused by simple channel differences, this step uses the temporal transition tension of the channel itself as a reference benchmark, introduces the relative transition deviation ratio and structural discrete weight of cross-channel transitions, and constructs anisotropic transition tensions describing the intensity of anisotropic disturbances between channels. The specific steps are: using the temporal transition tension value of channel k... As a benchmark, the temporal transition tension values ​​of other channels m in the same sample are introduced. ,pass Describes the cross-channel differences in transition intensity across different channels; to avoid the dominance of high-amplitude channels due to using only absolute differences, an additional factor is introduced into the denominator. This method maps cross-channel differences to relative transition deviation ratios, enabling anisotropic transition metrics to adaptively match the channel's own transition scale, and introduces a stability factor. To ensure the continuity of numerical calculations under low transition conditions; furthermore, considering the correlation between the CNC machine tool channels in terms of physical structure and control path, the channel index distance is used. Constructing structural discrete weights This makes the contribution of structurally closer channels to the anisotropic transition tension more significant; finally, by multiplying the relative transition deviation ratios of all remaining channels m with the structural discrete weights and then weighting and accumulating the results, the channel k in the sample is formed. The anisotropic transition tension under The mathematical model for the anisotropic transition tension is as follows:

[0098] ;

[0099] in, These are the channels other than channel k; For the sample The timing transition tension value of the middle channel m; The stabilization factor takes a value of This is used to avoid the denominator being zero and to suppress the amplification of extreme ratios; For the sample The anisotropic transition tension of the middle channel k.

[0100] S53. Introduce the anisotropic adjustment coefficient corresponding to the channel to adjust the anisotropic transition tension of the channel under the current sample. Combine the adjustment result with the anisotropic transition tension to form a ratio that includes the original strength information and structural adaptability, and obtain the path anomaly indicator.

[0101] In this embodiment, during the multi-channel collaborative operation of a CNC machine tool, the anisotropic transition tension exhibits significant amplitude imbalance characteristics under different channel structures and load conditions. Directly using the anisotropic transition tension for channel-level expression easily introduces scale drift and structural bias. This step takes the anisotropic transition tension as input and introduces a self-constrained anisotropic mapping mechanism to incorporate the transition strength of the channel itself into the constraint construction process, forming a channel-level path anomaly indicator. The specific steps are: taking channel k in the sample... The anisotropic transition tension under As a fundamental input, anisotropic transition tension is introduced into the molecule to maintain the original response scale of the channel transition strength, and then an anisotropic adjustment coefficient is simultaneously introduced into the denominator. homogeneous transition tension term This allows the transition strength of the channel itself to participate in the constraint construction process in reverse. The anisotropic transition tension is gradually constrained by the adaptive anisotropic limitations of the channel structure load characteristics during numerical growth, enabling different channels to form a consistent path anomaly indicator scale under their respective load conditions, thus completing the construction of the path anomaly indicator quantity. The mathematical model of the path anomaly indicator quantity is as follows:

[0102] ;

[0103] in, is the anisotropic adjustment coefficient for channel k, used to describe the structural bearing characteristics of the channel under the accumulation of transition tension during long-term operation. The initial value is set to 0.7 based on the historical stable state of the channel. Different parameters can be set according to the actual scenario, with a range of (0,2], to limit abnormal amplification and reduction. For the sample The path anomaly indicator for channel k.

[0104] S6. Construct a machine tool fault monitoring model, input the machine tool fault monitoring training set, combine the logarithmic loss function, and sequentially go through steps S2 to S5. Iteratively train the machine tool fault monitoring model until convergence, and use the machine tool fault monitoring test set to test the machine tool fault monitoring model to realize fault detection of CNC machine tools.

[0105] S61. The path anomaly indicators of each channel are weighted and aggregated according to the channel anomaly sensitivity weight. The machine tool fault detection probability is obtained through nonlinear compression, and the consistency between the detection results and the actual fault labels is constrained by the logarithmic loss function.

[0106] In this embodiment, the path anomaly indicators of each channel in the sample are used as the basic input. By introducing channel anomaly sensitivity weights, the anomaly contributions of different channels in the overall machine tool fault evolution process are weighted. At the sample level, multi-channel path anomaly information is uniformly aggregated to form a single machine tool fault monitoring output value. This machine tool fault monitoring output value comprehensively reflects the overall cumulative level of multi-channel path anomalies in the current sample. Before obtaining the machine tool fault monitoring output value, to prevent numerical expansion of the linear aggregation result during anomaly accumulation, the machine tool fault monitoring output value is further subjected to a nonlinear compression process using a Sigmoid mapping function. This ensures that the aggregated fault response is uniformly mapped to a stable range of 0 to 1, thus forming the final machine tool fault detection value. Its mathematical model is as follows:

[0107] ;

[0108] in, The channel anomaly sensitivity weights are initially set to 1, representing the weight of each channel. This is a learnable parameter that gradually adjusts the values ​​of the anomaly sensitivity weights for each channel, so that the weights of channels that contribute more to the monitoring error in the fault samples automatically increase, while the weights of channels that contribute less to the monitoring error gradually decrease. K is the number of channels, with a value of 19. The Sigmoid function is a mapping function that maps the values ​​to between 0 and 1. For machine tool fault monitoring model, samples The output value of machine tool fault monitoring;

[0109] A logarithmic loss function is constructed to maximize the probabilistic consistency between the output value of the machine tool fault monitoring model and the actual fault label. This ensures that the final fault detection value approaches 1 when it is close to a faulty sample and approaches 0 when it is close to a normal sample, exhibiting good discriminability and stability. The mathematical model of the logarithmic loss function is as follows:

[0110] ;

[0111] in, For the sample The actual fault value; batch is the number of samples used for training the machine tool fault monitoring model, with a value of 128; This represents the value of the logarithmic loss function.

[0112] S62. Input the machine tool fault monitoring training set, combine the logarithmic loss function, set the hyperparameters, construct the machine tool fault monitoring model, train the machine tool fault monitoring model until convergence, and use the machine tool fault monitoring test set to test the machine tool fault monitoring model to realize fault detection of CNC machine tools.

[0113] In this embodiment, a machine tool fault monitoring model is constructed. The model takes a machine tool fault monitoring training set as input and sequentially performs disturbance response value construction and driving relationship matrix analysis, anisotropic interference reconstruction, and fusion indicator generation to form channel-level fusion indicators. Based on this, the differences in the changes of the fusion indicators between adjacent samples are extracted to construct fusion indicator transition values. Anisotropic response tension is formed by combining channel index distance and propagation constraints, and path tension indicators are obtained after folded load adjustment. Furthermore, a temporal transition tension value is constructed based on the changes of the path tension indicators between the current sample and historical samples. The temporal transition tension of other channels is introduced into the same sample for cross-channel difference aggregation to form anisotropic transition tension. After anisotropic adjustment coefficient mapping, path anomaly indicators are generated. The machine tool fault monitoring model further weights and aggregates the path anomaly indicators of each channel according to channel anomaly sensitivity weights, and obtains machine tool fault detection values ​​through nonlinear compression mapping. Training optimization is performed using a logarithmic loss function to achieve fault state detection during CNC machine tool operation.

[0114] Furthermore, the machine tool fault monitoring model proposed in this invention is implemented using the Python programming language, and model construction and training are completed based on the PyTorch computing framework. During the training phase, the machine tool fault monitoring model uses the Adam optimizer to update the channel anomaly sensitive weights and related learnable parameters. The initial learning rate is set to 0.001, and an exponential decay strategy is introduced to dynamically adjust the learning rate. The learning rate decay coefficient is set to 0.95 to enhance the stability and convergence consistency of parameter updates under complex working conditions. In terms of hyperparameter configuration, the training batch size is set to 128, and the total number of training epochs is set to 2000. To avoid gradient instability caused by multi-channel path anomaly indicators during backpropagation, a gradient pruning mechanism is introduced, and the pruning threshold is set to 2.0, thereby ensuring the numerical stability and training reliability of the machine tool fault monitoring model during long-term training.

[0115] Furthermore, the machine tool fault monitoring training set is input into the constructed machine tool fault monitoring model for training. The trend of the logarithmic loss function used during the training process is shown in the figure. Figure 6It can be seen that during the training process, the log loss function initially represents a gradual establishment of the mapping relationship between path tension indicators, temporal transition tension, anisotropic transition tension, and path anomaly indicators. The log loss function value is at a relatively high level with some fluctuations. As the training rounds continue to increase, the channel anomaly sensitive weights are continuously adjusted under the inverse optimization constraint. The machine tool fault monitoring model gradually forms a stable channel-level anomaly fusion structure. The loss function value shows a continuous downward trend and eventually converges to a stable range, i.e., 0.1 to 0.15. This indicates that the constructed machine tool fault monitoring model has good numerical stability and convergence characteristics under complex multi-channel disturbance conditions and can reliably monitor the machine tool fault status.

[0116] Statistical analysis was performed on the path anomaly indicators corresponding to each sample during continuous machine tool operation, and the time-series distribution results of the path anomaly indicators were obtained as follows: Figure 7 As shown in the figure, during the normal operation of the machine tool, the path anomaly indicator is generally in a low-amplitude stable range, showing only slight fluctuations. This indicates that the multi-channel path tension structure remains stable at this time, and no obvious abnormal accumulation characteristics have appeared. As the machine tool operation time progresses, in the initial stage of the anomaly, the path anomaly indicator gradually increases, reflecting that the anisotropic transition tension between channels and the path tension indicator begin to deviate continuously, and the anomaly structure evolves from local disturbance to path-level accumulation. In the stable stage of the anomaly, the path anomaly indicator enters a higher amplitude range and maintains a continuous distribution, indicating that the abnormal path related to the machine tool fault has formed a stable structural feature. The above distribution and change process of the path anomaly indicator clearly describes the entire process of the machine tool evolving from normal operation to fault state.

[0117] To verify the effectiveness of the machine tool fault monitoring model in the fault identification stage, the fault monitoring results of the machine tool fault monitoring test set were compared and analyzed. The results are as follows: Figure 8 As shown in the figure, the horizontal axis represents the evolution process of the test sample within a continuous operating time window, and the vertical axis represents the machine tool fault monitoring value at the corresponding moment. As can be seen from the figure, the fault monitoring value output by the machine tool fault monitoring model is highly consistent with the actual fault state on the time axis. The response change process is continuous and stable without obvious deviation, indicating that the constructed machine tool fault monitoring model can accurately capture the evolution characteristics of the fault state in the time dimension, realize stable tracking of the intensity and trend of machine tool fault occurrence, and thus verify that the machine tool fault monitoring model has good discrimination consistency and time series modeling reliability under complex operating disturbance conditions.

[0118] Based on the same inventive concept as the above-described method embodiments, this invention provides a fault monitoring system for CNC machine tools. The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. It includes a data acquisition module for acquiring multi-channel operating status timing data of the CNC machine tool during operation, and performing time synchronization and windowing of the acquired data to form a machine tool operating status timing input for fault monitoring and analysis. A fault monitoring output module, connected to the data acquisition module, detects the real-time monitored CNC machine tool data and outputs real-time fault monitoring results of the CNC machine tool's operating status. When the computer program is executed by the processor, it implements the steps of a fault monitoring method for CNC machine tools.

[0119] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A fault monitoring method for CNC machine tools, characterized in that, Includes the following steps: Collect and preprocess CNC machine tool operation data; The variation amplitude and amplitude suppression relationship within the sliding window of each channel are fused to form the spectral response value. The spectral response value is then used to construct the offset modulation to obtain the perturbation response value. The perturbation response value is then used to characterize the relationship between the deviation intensity and direction between channels, generating the driving relationship matrix between channel pairs. The disturbance response values ​​are divided by channel. Anisotropic interferometric reconstruction values ​​are constructed based on the difference amplitude between channels and the driving relationship matrix. Strong residuals are generated by the offset and range between the disturbance response values ​​and the anisotropic interferometric reconstruction values. Path weights are constructed by unifying the amplitude of the anisotropic interferometric reconstruction values. Channel-level fusion indicators are formed by combining the strong residuals. The change amplitude of each channel between two adjacent times within the sliding window is calculated to form the disturbance intensity term. The amplitude ratio of each channel value to the current channel mean is extracted, and after introducing a stabilization factor, it is nonlinearly compressed to form an amplitude suppression factor. The perturbation intensity term and the amplitude suppression factor are calculated point by point, and the average of all calculation results within the window is calculated to form the spectral response value. Based on the spectral response value of each channel, the mean and standard deviation of all samples under the current channel are calculated as an overall reference. The spectral response value of a single sample is compared with the mean of the current channel to construct the deviation. The standard deviation is proportionally processed and nonlinearly transformed to introduce an inverse suppression structure to form the perturbation response value. For the perturbation response value of each channel, the degree of deviation of the driving channel from its own mean in the current sample is calculated, and the driving deviation intensity is constructed by combining it with the standard deviation of the whole sample. Simultaneously, the direction of change of the response channel relative to its mean in the current sample is extracted to generate a response direction indicator; After fusing the driving deviation intensity and the response direction indicator, the fusion results for all samples are averaged to obtain a driving relationship matrix that includes the directional influence between all channel pairs. The fusion indicator is extracted from the differences between adjacent samples to form the fusion indicator transition value. The transition difference between the current channel and other channels is calculated and propagation constraints are constructed by combining the channel index distance. The anisotropic response tension is obtained by convergence and then compressed and adjusted by the folded load adjustment coefficient to construct the path tension indicator. Taking the target channel in the current sample as the benchmark, the difference amplitude between its perturbation response value and other candidate channels is calculated sequentially. The difference amplitude is exponentially modulated by the driving relationship matrix to obtain the interferometric modulation weight. The perturbation response value of the candidate channel is weighted and corrected using the interferometric modulation weight. The value with the largest correction result is selected from all candidates as the anisotropic interferometric reconstruction value of the target channel in the current sample. The candidate channel set is then used to extract... The maximum and minimum values ​​of the anisotropic interferometric reconstruction values ​​are taken to form the range, which is used as the amplitude adjustment term. The difference between the disturbance response value and the anisotropic interferometric reconstruction value is calculated on the current sample channel. A sign term is constructed based on the offset direction sign. The difference value is mapped and used in the calculation of the sign term. Finally, it is combined with the amplitude adjustment term to obtain the strong residual. Taking a single channel as the calculation object, the anisotropic interferometric reconstruction values ​​corresponding to the other channels are extracted in sequence. The amplitude of the anisotropic interferometric reconstruction values ​​is uniformized to form path weights. The path weights are multiplied with the strong residual of the corresponding channel and all channels are fused to form a total variable indicator. Finally, it is combined with the strong residual of the current channel to obtain the channel-level fused indicator. Based on the relationship between the path tension indicator changes of adjacent samples and historical samples, a temporal transition tension value is constructed. The temporal transition tension values ​​of other channels within the same sample are introduced to form cross-channel transition differences. The structural discrete weights are constructed by combining the channel index distance and weighted accumulation to obtain the anisotropic transition tension. After being constrained and mapped by the anisotropic adjustment coefficient, a path anomaly indicator is formed. Train a machine tool fault monitoring model and output machine tool fault detection results based on the machine tool fault monitoring model.

2. The fault monitoring method for CNC machine tools according to claim 1, characterized in that, CNC machine tool operation data is obtained by combining online monitoring of actual production and processing, no-load and load operation tests of equipment, and long-term historical operation records. Operation status monitoring points are set up in the spindle system, feed axis servo system, guide rail position and electrical control unit to collect data and form multi-source operation status time sequence raw data. The multi-source operating status time series raw data are sequentially processed by time alignment, invalid segment removal, outlier cleaning based on physical constraints and statistical criteria, missing data imputation, dimension unification and numerical normalization. The continuous time series data is sliced ​​using a fixed-length time window to construct CNC machine tool fault monitoring data, which includes machine tool fault monitoring training set and machine tool fault monitoring test set.

3. The fault monitoring method for CNC machine tools according to claim 1, characterized in that, Based on the channel-level fusion indicator, the variation relationship of the same channel in adjacent samples is extracted, and a comparison structure between the current difference amplitude and the historical difference amplitude is constructed. On this basis, the current difference amplitude is embedded into an exponential interference modulation structure for modulation processing, and the modulation result is combined with the comparison structure to form a fusion indicator transition value that simultaneously contains the relative change intensity and modulation constraint characteristics. Using the current channel as a benchmark, the differences in the fusion indicator transition values ​​of other channels are compared one by one, and cross-channel propagation attenuation constraints are introduced in combination with the channel index distance. On this basis, the directional driving modulation of the difference relationship between each channel is carried out, and the modulated difference results are converged to form the anisotropic response tension of the current channel. Using the anisotropic response tension of the channel as the input, a nonlinear compression mapping relationship is constructed by introducing a folding load adjustment coefficient. The anisotropic response tension is proportionally adjusted to the anisotropic response tension compressed by the folding load adjustment coefficient to form a path tension indicator.

4. The fault monitoring method for CNC machine tools according to claim 1, characterized in that, Based on the path tension indicator, the change in the path tension indicator between the current sample channel and the previous sample channel is extracted, and the change in the path tension indicator of adjacent historical samples is calculated. The current trend is compared with the historical trend by difference. After eliminating the influence of the change direction, a normalization adjustment factor is introduced to perform dynamic scale constraint, forming the temporal transition tension value of the channel in the current sample. Using the temporal transition tension value of the current channel as input, the temporal transition tension values ​​of other channels in the same sample are introduced in sequence to construct the cross-channel difference in transition intensity of different channels. Under the constraint of the temporal transition tension value of the channel itself, the cross-channel difference is proportionalized to construct the relative transition deviation ratio. The channel index distance forms the structural discrete weight. The relative transition deviation ratio and structural discrete weight of all channels are weighted and accumulated to obtain the anisotropic transition tension. An anisotropic adjustment coefficient corresponding to the channel is introduced to adjust the anisotropic transition tension of the channel under the current sample. The adjustment result and the anisotropic transition tension are combined to form a ratio that includes the original strength information and structural adaptability, thus obtaining the path anomaly indicator.

5. The fault monitoring method for a CNC machine tool according to claim 1, characterized in that, The path anomaly indicators of each channel are weighted and aggregated according to the channel anomaly sensitivity weight, and the machine tool fault detection probability is obtained through nonlinear compression. The consistency between the detection results and the actual fault labels is constrained by the logarithmic loss function. Input the machine tool fault monitoring training set, combine the logarithmic loss function, set the hyperparameters, construct the machine tool fault monitoring model, train the machine tool fault monitoring model until convergence, and use the machine tool fault monitoring test set to test the machine tool fault monitoring model to realize fault detection of CNC machine tools.

6. A fault monitoring system for CNC machine tools, characterized in that, It includes a processor and a memory, the processor being used to process instructions stored in the memory to implement a fault monitoring method for a CNC machine tool according to any one of claims 1-5.

Images