Intelligent monitoring method, system and storage medium for CNC machine tools based on the Internet of Things

Abstract

The present application relates to the technical field of intelligent monitoring of machine tools, and discloses a method, system and storage medium for intelligent monitoring of CNC machine tools based on the Internet of Things. The method includes: collecting multi-source data of machine tools and transmitting it to edge nodes; processing the original data to obtain a working condition feature set and inputting it into a recursive convolutional hybrid memory network to construct a co-dependency graph, and realizing state parameter prediction through joint analysis in the time and frequency domains; converting it into a two-dimensional feature graph and extracting interactive features through SE‑ResNet; fusing two types of features to perform state classification; based on the health status judgment results combined with process parameters and historical data, predicting life and generating a diagnostic report. The present application realizes the effective fusion analysis of multi-source heterogeneous data of CNC machine tools, constructs an adaptive state recognition and fault warning mechanism, thereby realizing accurate monitoring of the machine tool operating status, intelligent diagnosis of early faults and accurate prediction of remaining life.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent monitoring of machine tools, and in particular to an intelligent monitoring method, system and storage medium for CNC machine tools based on the Internet of Things. Background Art

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, CNC machine tools, as core equipment in modern manufacturing, are receiving increasing attention for their operating status monitoring and fault diagnosis technologies. Traditional CNC machine tool monitoring methods rely primarily on periodic manual inspections and simple threshold alarm mechanisms. These methods typically use a single type of sensor to monitor specific parameters, such as vibration sensors for spindle vibration and temperature sensors for bearing temperature. These methods have significant limitations in practical applications, capturing only a single dimension of status information and failing to fully reflect the complex operating conditions of machine tools. In recent years, with the advancement of the Internet of Things (IoT) technology, some leading companies have begun experimenting with using multi-source sensors to collect machine tool operating data. Using simple data fusion methods, they implement basic status monitoring functions, such as setting fixed thresholds for signals such as vibration, temperature, and current to generate alarms.

[0003] However, existing CNC machine tool monitoring technologies still have many shortcomings. First, traditional methods have difficulty processing the complex correlations between multi-source heterogeneous data and are unable to extract implicit fault characteristics and status information from large amounts of monitoring data. Second, conventional threshold alarm methods can only detect faults that have already appeared and lack the ability to early identify and predict potential faults. Third, existing monitoring systems lack the ability to adaptively identify machine tool status under different operating conditions, making it difficult to accurately distinguish between parameter fluctuations caused by changes in normal operating conditions and abnormal signals caused by actual faults. Fourth, most systems only provide simple fault alarms, lack the precise location of the cause of the fault and the scientific prediction of the remaining service life, and cannot provide a sufficient basis for maintenance decisions. In addition, existing technologies lack consideration of edge computing and network transmission, resulting in low data processing efficiency and poor real-time performance, making it difficult to meet the needs of industrial sites for rapid response monitoring systems. Summary of the Invention

[0004] The present application provides an intelligent monitoring method, system and storage medium for CNC machine tools based on the Internet of Things, which are used to realize effective fusion analysis of multi-source heterogeneous data of CNC machine tools, build an adaptive state recognition and fault warning mechanism, thereby realizing accurate monitoring of the machine tool operating status, intelligent diagnosis of early faults and accurate prediction of remaining life.

[0005] In the first aspect, the present application provides an intelligent monitoring method for CNC machine tools based on the Internet of Things, which includes: collecting CNC machine tool spindle vibration, motor temperature, processing noise, drive current and worktable displacement data, and synchronously transmitting them to edge nodes through an industrial bus to obtain raw machine tool operation data; performing signal processing on the raw machine tool operation data to obtain a machine tool working condition feature set; inputting the machine tool working condition feature set into a recursive convolutional hybrid memory network to construct a co-dependency relationship graph of spindle speed change, tool wear, and bearing status, and realizing state parameter prediction through joint analysis in the time and frequency domains to obtain machine tool component state characteristics; converting the machine tool component state characteristics into a two-dimensional feature graph, extracting the interaction characteristics between the spindle-feed-control system through SE-ResNet, and obtaining a machine tool system state representation; fusing the machine tool component state characteristics with the machine tool system state representation, performing state classification, and obtaining a machine tool health status judgment result; based on the machine tool health status judgment result, combined with processing process parameters and historical fault data, performing remaining life prediction and fault warning to obtain a machine tool diagnostic report containing maintenance recommendations.

[0006] In a second aspect, the present application provides an intelligent monitoring system for CNC machine tools based on the Internet of Things, the intelligent monitoring system for CNC machine tools based on the Internet of Things comprising:

[0007] The transmission module is used to collect CNC machine tool spindle vibration, motor temperature, processing noise, drive current and worktable displacement data, and transmit them synchronously to the edge node through the industrial bus to obtain the original machine tool operation data;

[0008] a processing module, configured to perform signal processing on the raw data of the machine tool operation to obtain a machine tool operating condition feature set;

[0009] An analysis module is used to input the machine tool working condition feature set into a recursive convolutional hybrid memory network, construct a co-dependency relationship diagram of spindle speed change, tool wear, and bearing status, realize state parameter prediction through joint analysis in the time and frequency domains, and obtain the state characteristics of the machine tool components;

[0010] A conversion module is used to convert the state features of the machine tool components into a two-dimensional feature map, extract the interaction features between the spindle, feed, and control systems through SE-ResNet, and obtain a machine tool system state representation;

[0011] An execution module, configured to integrate the machine tool component state characteristics with the machine tool system state representation, perform state classification, and obtain a machine tool health state determination result;

[0012] The prediction module is used to perform remaining life prediction and fault warning based on the machine tool health status judgment result, combined with processing parameters and historical fault data, to obtain a machine tool diagnosis report including maintenance suggestions.

[0013] In a third aspect, an intelligent monitoring device for CNC machine tools based on the Internet of Things is provided, comprising: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the intelligent monitoring device for CNC machine tools based on the Internet of Things executes the above-mentioned intelligent monitoring method for CNC machine tools based on the Internet of Things.

[0014] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned intelligent monitoring method for CNC machine tools based on the Internet of Things.

[0015] In the technical solution provided in this application, by collecting multi-source sensor data, all-round and multi-dimensional state perception of the machine tool is achieved, the defect of limited information of a single sensor is overcome, and a data foundation is laid for a comprehensive understanding of the operating status of the machine tool; professional signal processing is performed on the original data of the machine tool operation, and specific processing methods are adopted for different types of signals, such as applying wavelet packet transform to the vibration signal to extract frequency domain features, performing dynamic window sliding average filtering on the temperature signal, calculating Mel-frequency cepstral coefficients on the acoustic signal, performing Fourier analysis on the current signal, and performing differential operations on the displacement signal, thereby extracting a rich set of working condition features and fully mining the state information contained in various signals; the machine tool working condition feature set is input into the recursive convolution hybrid memory network, and through the combination of multi-scale convolution and long-short-term memory structure, the time series correlation and co-dependency analysis of spindle speed changes, tool wear, and bearing status are realized, and a state co-dependency diagram is constructed and analyzed through joint time-frequency domain analysis. , overcoming the shortcomings of traditional methods in handling complex component interactions; by converting machine tool component state features into two-dimensional feature maps and inputting them into the SE-ResNet network, its channel attention mechanism is used to adaptively adjust the importance of different feature channels, thereby enhancing the ability to extract interaction features between the spindle, feed, and control systems and improving the accuracy of state recognition; fusing temporal features and spatial features, and combining them with the feature gating mechanism for adaptive working conditions, the feature weights are dynamically adjusted for different processing conditions, solving the problem of insufficient recognition accuracy of traditional methods under complex and changing working conditions; based on the health status judgment results and historical fault data, a component health index scoring system is established, and the remaining life is predicted by combining processing parameters and working condition loads, realizing the transition from passive response to active prediction, and providing a scientific basis for predictive maintenance; generating targeted maintenance suggestions and formulating maintenance planning plans according to the warning level, reducing maintenance costs and reducing downtime. The innovative algorithm architecture of the present invention combines the recursive convolutional hybrid memory network with the SE-ResNet network, giving full play to the advantages of the recursive convolutional hybrid memory network in time-frequency feature extraction and co-dependency modeling, as well as the strengths of SE-ResNet in spatial feature extraction. The synergistic effect of the two algorithms significantly improves the ability to recognize the status of machine tools under complex working conditions, and solves the technical problem that traditional single algorithms are difficult to simultaneously take into account temporal correlation and spatial interaction features; in particular, the recursive convolutional hybrid memory network can effectively process multi-source heterogeneous data and construct a co-dependency graph between components, while the channel attention mechanism of SE-ResNet can adaptively highlight important feature channels. The combined application of these two mechanisms in machine tool status monitoring enables the system to accurately identify subtle anomalies under complex working conditions, greatly improving the accuracy and timeliness of fault warnings, and realizing accurate assessment of the health status of machine tools and accurate prediction of fault development trends. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 This is a schematic diagram of an embodiment of an intelligent monitoring method for CNC machine tools based on the Internet of Things in an embodiment of the present application;

[0018] Figure 2 This is a schematic diagram of an embodiment of an intelligent monitoring system for CNC machine tools based on the Internet of Things in an embodiment of the present application;

[0019] Figure 3 It is a schematic block diagram of the structure of an intelligent monitoring device for CNC machine tools based on the Internet of Things in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The embodiments of the present application provide a method, system and storage medium for intelligent monitoring of CNC machine tools based on the Internet of Things. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0021] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the intelligent monitoring method of CNC machine tools based on the Internet of Things includes:

[0022] Step S101: Collect the spindle vibration, motor temperature, machining noise, drive current and worktable displacement data of the CNC machine tool, and transmit them synchronously to the edge node via the industrial bus to obtain the original data of the machine tool operation;

[0023] Step S102: performing signal processing on the original data of the machine tool operation to obtain a machine tool working condition feature set;

[0024] Step S103: Input the machine tool working condition feature set into the recursive convolutional hybrid memory network to construct a co-dependency diagram of spindle speed change, tool wear, and bearing status, and predict state parameters through joint analysis in the time and frequency domains to obtain machine tool component state features;

[0025] Step S104: converting the state features of the machine tool components into a two-dimensional feature map, extracting the interaction features between the spindle, feed, and control systems through SE-ResNet, and obtaining a representation of the machine tool system state;

[0026] Step S105: integrating the machine tool component status characteristics with the machine tool system status representation, performing status classification, and obtaining the machine tool health status determination result;

[0027] Step S106: Based on the machine tool health status determination result, combined with the processing parameters and historical fault data, perform remaining life prediction and fault warning to obtain a machine tool diagnosis report including maintenance suggestions.

[0028] It is understandable that the execution subject of this application can be a CNC machine tool intelligent monitoring system based on the Internet of Things, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0029] In an embodiment of the present application, an acceleration sensor is installed on the spindle to obtain a vibration signal, a temperature sensor is installed on the surface of the motor housing to measure the temperature rise, an acoustic sensor is set around the processing area to capture noise characteristics, a current sensor is embedded in the drive system to monitor the current waveform, and a displacement sensor is arranged on the workbench to track the motion trajectory. The data collected by these sensors are transmitted to the edge computing node via the RS485 bus or industrial Ethernet according to a predetermined sampling strategy. The sampling frequency is set according to different signal characteristics. The vibration signal uses 20kHz high-frequency sampling to capture transient characteristics, and the temperature signal uses 1Hz low-frequency sampling to track slowly changing trends. The edge node performs time synchronization to ensure that multi-source data is aligned in the time dimension to form the original data of the machine tool operation. Signal processing is performed on the original data of the machine tool operation to extract features. The vibration signal is decomposed into five layers using a wavelet packet transform. Energy distribution characteristics are extracted from different frequency bands, revealing characteristic frequency information for components such as bearings and gears. A sliding average filter is applied to the temperature signal to eliminate sudden noise and reflect the thermal state of the motor. Mel-frequency cepstral coefficients are calculated for the machining noise, and 13 main coefficients and their differential values are extracted to characterize the acoustic characteristics. A Fourier transform is performed on the current signal to analyze its harmonic components and calculate the power factor to reflect the motor load condition. The displacement signal is differentiated to calculate velocity and acceleration for motion accuracy assessment. These different modal features are organized into a structured time series matrix, with rows representing time points and columns representing feature dimensions. Normalization is performed to eliminate dimensional differences, and principal component analysis is then used to reduce the dimensionality, retaining the principal components with a contribution rate of 95%, forming a machine tool operating condition feature set. This feature set is then fed into a recursive convolutional hybrid memory network to analyze temporal correlations. Positional encoding is first applied to the feature sequence to embed temporal position information into the feature representation. Local characteristic patterns within different time windows are then extracted using a multi-scale convolutional layer. The multi-scale convolutional layer consists of parallel 1×1, 3×3, and 5×5 convolutional kernels. The convolutional output features are weighted using a channel-wise attention mechanism. The convolutional features are then fed into a long-short-term memory layer, which uses a bidirectional gated recurrent unit to capture long-term dependencies. The forget gate, input gate, and output gate precisely control the flow of information. The attention fusion layer calculates the interaction strength matrix between the three-dimensional features of spindle speed variation, tool wear, and bearing status. Non-negative matrix factorization is used to extract latent correlation factors and construct a co-dependency graph. In this graph, nodes represent spindle speed, tool wear, and bearing status features, and edge weights represent the interaction strength between features. A continuous wavelet transform is applied to the co-dependency graph, mapping the time-domain signals to the time-frequency plane. The characteristic frequencies and time-frequency correlation patterns of key components, such as spindle speed fluctuations, tool wear progression, and bearing status changes, are identified. Finally, a gated update mechanism is used to fuse the time-frequency features with contextual features to generate machine tool component status features. These machine tool component status features are converted into a two-dimensional feature map, and the SE-ResNet network is applied to extract spatial feature relationships.First, the one-dimensional feature vector is reorganized into a two-dimensional feature map with 64 channels and a size of 7×7. Basic features are extracted through an initial convolutional layer. The core of SE-ResNet is a residual structure with a channel-wise attention mechanism. Residual connections alleviate the difficulties of training deep networks. The SE module compresses the feature map into channel descriptors through global average pooling. The importance weights between channels are learned through a two-layer fully connected network, and the feature map is weighted by channel. Over the four stages of processing, the number of channels is gradually increased to 512. Finally, a global pooling operation converts the feature map into a fixed-dimensional vector, capturing the state interactions between the spindle, feed, and control systems, and generating a representation of the machine tool system state.

[0030] State classification is performed by fusing machine tool component status characteristics with machine tool system state representations. The two features are concatenated and weighted using a feature gating mechanism. For example, the weight of the spindle feature is increased in high-speed cutting conditions, while the weight of the feed system feature is increased during frequent starts and stops. This results in a fused feature that is adaptive to the working conditions. A classifier network maps the fused features into a probability distribution of machine tool states. A multi-level threshold judgment mechanism identifies different fault types. For example, a peak in the spindle vibration spectrum at the characteristic frequency of a bearing fault is identified as a bearing anomaly. Sliding window smoothing ensures state judgment stability, and a state transition is confirmed when the states of five consecutive windows are consistent during machining. By matching the current state with a library of historical abnormal patterns, the machine tool health status is determined through a comprehensive assessment. Based on the health status determination results, the remaining life is predicted and a diagnostic report is generated. A health index scoring system is established for key components. The operating load is assessed based on current machining parameters, and historical failure cases are searched for reference data. This information is combined to calculate component degradation rates and trends, and predict the remaining life. Multi-level warning thresholds are set based on the predicted values, and corresponding warnings are triggered when the predicted life falls below a specific threshold. Generate specific maintenance recommendations for different warning levels, prioritize maintenance tasks, estimate required time and costs, and integrate them into diagnostic reports for presentation via edge computing nodes and cloud platforms.

[0031] Taking a certain type of vertical machining center as an example, the monitoring system collected spindle vibration data and showed abnormal energy accumulation in the 8kHz frequency band. The temperature sensor detected that the spindle temperature rise rate reached 0.8℃ / hour. The acoustic sensor captured the increase in sound pressure level at the bearing characteristic frequency and the increase in harmonic content in the current waveform. The recursive convolution hybrid memory network analyzed the time-frequency correlation of these multi-source signals and identified the typical co-dependency pattern of bearing anomalies. The SE-ResNet network further extracted the interaction characteristics of the bearing and spindle system. The fusion analysis determined that it was an early spalling fault of the bearing inner ring. The system predicted that the remaining life of the bearing was 320 hours, planned maintenance time in advance, and generated specific step-by-step instructions for bearing replacement, avoiding production interruptions caused by sudden failures and reducing maintenance costs.

[0032] In the embodiment of the present application, by collecting multi-source sensor data, all-round and multi-dimensional state perception of the machine tool is achieved, the defect of limited information of a single sensor is overcome, and a data foundation is laid for a comprehensive understanding of the operating status of the machine tool; professional signal processing is performed on the original data of the machine tool operation, and specific processing methods are adopted for different types of signals, such as applying wavelet packet transform to the vibration signal to extract frequency domain features, performing dynamic window sliding average filtering on the temperature signal, calculating Mel-frequency cepstral coefficients on the acoustic signal, performing Fourier analysis on the current signal, and performing differential operations on the displacement signal, thereby extracting a rich set of working condition features and fully mining the state information contained in various types of signals; the machine tool working condition feature set is input into the recursive convolution hybrid memory network, and through the combination of multi-scale convolution and long-short-term memory structure, the time series correlation and co-dependency analysis of spindle speed changes, tool wear, and bearing status are realized, a state co-dependency diagram is constructed and jointly analyzed in the time and frequency domains to overcome It overcomes the deficiency of traditional methods in handling the interaction relationship between complex components; by converting the state characteristics of machine tool components into a two-dimensional feature map and inputting it into the SE-ResNet network, its channel attention mechanism is used to adaptively adjust the importance of different feature channels, thereby enhancing the ability to extract the interaction characteristics between the spindle, feed and control system and improving the accuracy of state recognition; by fusing temporal features and spatial features, and combining the feature gating mechanism of working condition adaptation, the feature weights are dynamically adjusted for different processing conditions, solving the problem of insufficient recognition accuracy of traditional methods under complex and changing working conditions; based on the health status judgment results and historical fault data, a component health index scoring system is established, and the remaining life is predicted by combining the processing parameters and working condition loads, realizing the transition from passive response to active prediction, and providing a scientific basis for predictive maintenance; generating targeted maintenance suggestions and formulating maintenance planning plans according to the warning level, reducing maintenance costs and reducing downtime. The innovative algorithm architecture of the present invention combines the recursive convolutional hybrid memory network with the SE-ResNet network, giving full play to the advantages of the recursive convolutional hybrid memory network in time-frequency feature extraction and co-dependency modeling, as well as the strengths of SE-ResNet in spatial feature extraction. The synergistic effect of the two algorithms significantly improves the ability to recognize the status of machine tools under complex working conditions, and solves the technical problem that traditional single algorithms are difficult to simultaneously take into account temporal correlation and spatial interaction features; in particular, the recursive convolutional hybrid memory network can effectively process multi-source heterogeneous data and construct a co-dependency graph between components, while the channel attention mechanism of SE-ResNet can adaptively highlight important feature channels. The combined application of these two mechanisms in machine tool status monitoring enables the system to accurately identify subtle anomalies under complex working conditions, greatly improving the accuracy and timeliness of fault warnings, and realizing accurate assessment of the health status of machine tools and accurate prediction of fault development trends.

[0033] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0034] Arrange a vibration sensor array at key locations on the CNC machine tool spindle, obtain the spindle's radial and axial vibration characteristics through multi-point measurement, and obtain the original spindle vibration signal;

[0035] Deploy a temperature sensor network in the hot spots of CNC machine tool drive motors, servo motors, and control cabinets to build a temperature field distribution monitoring system and obtain the original motor temperature signal;

[0036] Acoustic sensors are set up around the machining area of the CNC machine tool, and beamforming is used to perform spatial direction collection of machining noise to obtain the original machining noise signal;

[0037] A current monitoring module is embedded in the power supply circuit of the CNC machine tool spindle motor and feed system to collect the current and harmonic components of each phase and obtain the original drive current signal;

[0038] A displacement sensing network is built on the worktable and guide rail system of the CNC machine tool to monitor the motion trajectory and precision deviation of the worktable and obtain the original displacement signal of the worktable;

[0039] Time synchronization and alignment are performed on the original signals of spindle vibration, motor temperature, machining noise, drive current, and worktable displacement to establish a unified time base and obtain time-aligned multi-source signals.

[0040] Transmit time-aligned multi-source signals to the edge computing layer through industrial-grade communication protocols to build a data transmission channel;

[0041] Data quality assessment is performed at the edge computing layer, and validity verification and noise elimination are performed on multi-source signals to obtain the original data of machine tool operation.

[0042] Specifically, a vibration sensor array is deployed on the CNC machine tool spindle. Three-axis accelerometers are installed at the front and rear bearings, as well as in the middle of the spindle. Each sensor simultaneously measures vibration acceleration in the X, Y, and Z directions. These piezoelectric accelerometers have a measurement range of ±50g, a sensitivity of 100mV / g, and a frequency response range of 0.5Hz-10kHz. Multi-point measurements capture vibration characteristics at different spindle locations. This not only captures radial vibration data reflecting imbalance and eccentricity, but also axial vibration data reflecting axial load variations. The raw spindle vibration signal includes a timestamp, measurement location, and triaxial acceleration values. A temperature sensor network is deployed at key hotspots in the CNC machine tool's drive system. These sensors include PT100 platinum resistance temperature sensors installed on the spindle motor windings, servo motor housing, bearing housing, control cabinet heat sink, and power components. These sensors have an accuracy of ±0.5°C and a measurement range of -40°C to 125°C. Multi-point temperature monitoring creates a comprehensive temperature distribution map, providing real-time insights into the thermal state of each component. For high-precision machining centers, temperature changes directly affect the accuracy and stability of the machine tool, so temperature field monitoring is extremely critical. The raw motor temperature signal contains the timestamp, the location of the monitoring point, and the temperature value.

[0043] An acoustic sensor array consisting of at least four microphones is deployed around the machining area, using beamforming technology to spatially and directional collect machining noise. Beamforming is a signal processing technique that adjusts the phase and amplitude of the signals received by each element in the microphone array to enhance sound signals in a specific direction and suppress interfering noise from other directions. The acoustic sensor has a sampling frequency of 44.1kHz and a dynamic range of 30-120dB. These sensors can capture abnormal machining sounds, such as tool collisions, breakage, or bearing noise. The raw machining noise signal contains timestamps, direction information, and sound pressure level data. A current monitoring module is embedded in the CNC machine tool's electrical system. Hall effect current sensors are installed directly in the power supply lines of the spindle motor and each axis' servo motor. The measurement range is 0-100A with an accuracy of ±1%. For three-phase motors, each phase current is monitored separately, and the current RMS value, crest factor, and harmonic content are calculated through real-time sampling. Current characteristics directly reflect the motor's load condition and health status. Increased harmonic content often indicates a motor or control system anomaly. The raw drive current signal contains a timestamp, motor ID, phase ID, and instantaneous current value.

[0044] A displacement sensing network is constructed on the worktable and guideway system of a CNC machine tool. This involves installing linear or magnetic scales with an accuracy of ±0.01mm along the worktable's X, Y, and Z axes. Proximity sensors are also installed at key guideway locations to monitor the worktable's motion. This displacement sensing network accurately monitors the worktable's actual motion trajectory, velocity, and acceleration characteristics. By comparing the actual motion trajectory with the commanded trajectory, the position error and trajectory deviation are calculated, allowing the dynamic accuracy of the machine tool's feed system to be evaluated. The raw worktable displacement signal includes a timestamp, axis identifiers, and position coordinates.

[0045] The raw signals collected by all sensors require time synchronization and alignment. Initial alignment is performed based on the timestamps in the data packets. Then, using a time synchronization protocol (such as the IEEE 1588 Precision Time Protocol), the clocks of all sensors are synchronized to a common standard, with an error of less than 1 millisecond. For signals with different sampling frequencies, resampling is performed using an interpolation algorithm to bring all signals to the same time step. Time-aligned multi-source signals form a unified data structure, with each record containing the timestamp and valid data from all sensors at that moment.

[0046] Time-aligned multi-source signals are transmitted to the edge computing layer via industrial-grade communication protocols. A star-topology sensor network is constructed using industrial Ethernet, RS485 bus, or IEEE 802.15.4 wireless communication. Appropriate communication methods are selected based on data volume and real-time requirements: high-sampling-rate vibration and sound data are transmitted via Ethernet, while low-frequency temperature data is transmitted wirelessly. A checksum mechanism is used during data transmission to ensure data integrity, and encryption is implemented for critical data to ensure security. After receiving the multi-source signals, the edge computing layer first performs data quality assessment, including integrity checks (detecting packet loss), validity verification (determining whether the data is within a reasonable range), and noise detection (identifying sudden changes and outliers). Detected outliers are processed using different strategies based on data type: vibration data is processed using a median filter to remove impulse noise, temperature data is smoothed using a sliding average, and displacement data is processed using a Kalman filter to estimate the true state. The data after quality assessment and preliminary processing constitutes the raw machine tool operation data, which is stored in the edge node's time series database for further analysis.

[0047] For example, a vibration sensor installed on a certain vertical machining center (VMC) detected periodic variations in radial vibration amplitude. Spectral analysis revealed a significant increase in the 1X component, which correlates to spindle speed. Simultaneously, a temperature sensor detected a bearing temperature 15°C above normal, and an acoustic sensor detected abnormal noise at a specific frequency. The current monitoring module detected increased spindle motor current fluctuations and harmonic content. Displacement sensor data revealed slight periodic errors during worktable motion. The edge computing layer, through time synchronization, aligned these signals to a unified timestamp and discovered that vibration, temperature rise, noise, and current anomalies all occurred in the same time period and varied with spindle speed. Initially, the issue was suspected to be a spindle bearing problem. The data quality assessment eliminated occasional impact interference during the cutting process, retaining valid signals reflecting the bearing condition.

[0048] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0049] The original spindle vibration signal is decomposed by wavelet packet transform, and the decomposition layer is selected as 5 layers. The energy distribution of each frequency band is extracted to obtain the frequency domain characteristics of the spindle vibration.

[0050] Perform sliding average filtering on the original motor temperature signal, and dynamically adjust the window size according to the temperature change rate to eliminate instantaneous interference and obtain a smoothed temperature sequence;

[0051] Calculate the Mel-frequency cepstral coefficients of the original signal of the processed noise, extract the first 13 coefficients and their first-order differences and second-order differences to form the acoustic feature vector;

[0052] Perform Fourier analysis on the original driving current signal to calculate the fundamental component and each order harmonic content, and extract the current effective value, peak factor and power factor to obtain the current characteristics;

[0053] Perform differential operation on the original signal of worktable displacement to calculate velocity and acceleration respectively, and extract position error, trajectory deviation and jerk characteristics to obtain motion characteristics;

[0054] The spindle vibration frequency domain features, smoothed temperature series, acoustic eigenvectors, current features, and motion features are combined to construct a multidimensional feature set, which is organized into a structured time series data matrix, with each row corresponding to a time point and each column corresponding to a feature dimension.

[0055] The multidimensional feature set is standardized, and each feature is normalized to a distribution with a mean of 0 and a standard deviation of 1 using the Z-score to eliminate dimensional differences and obtain a standardized feature matrix;

[0056] Principal component analysis is performed on the standardized feature matrix, and the principal components with a cumulative contribution rate of 95% are retained to reduce feature redundancy and obtain the machine tool working condition feature set.

[0057] Specifically, in the IoT-based intelligent monitoring method for CNC machine tools, the wavelet packet transform is applied to the original spindle vibration signal for multi-scale analysis. The db4 wavelet function is first selected as the basis function, and the number of decomposition layers is set to five. The wavelet packet transform is an extension of the wavelet transform. It not only decomposes the low-frequency component but also performs an equal-width decomposition of the high-frequency component, forming a frequency band tree structure. With the five-layer decomposition, the original vibration signal is divided into 32 frequency bands, each corresponding to a portion of the signal energy. The energy eigenvalue is calculated for each frequency band, and then the energy distribution ratio of each frequency band is calculated to form the frequency domain eigenvector of the spindle vibration. These frequency band energy distribution characteristics effectively reflect the frequency composition of the spindle vibration and can identify abnormal patterns such as bearing faults, imbalance, and misalignment.

[0058] A sliding average filter is applied to the raw motor temperature signal, with the window size dynamically adjusted based on the temperature change rate. The temperature change rate is first calculated. If it exceeds a preset threshold, the window size is reduced to preserve the rapidly changing characteristics; if the rate is smaller, the window size is increased to smooth out small fluctuations. For example, during the motor startup phase, when the temperature rises rapidly, a smaller window size, such as three points, is used; during stable operation, a larger window size, such as 15 points, is used. This smoothed temperature series preserves the temperature trend while eliminating measurement noise and transient interference. Mel-frequency cepstral coefficients (MFCCs) are calculated on the raw signal of the processed noise. This is an acoustic feature representation method that mimics the human auditory perception. The sound signal is first framed, with each frame length being 25ms and the frame shift being 10ms. Each frame is windowed and then subjected to a fast Fourier transform to obtain the power spectrum. This is then mapped to the mel-frequency scale using a mel filter bank. The energy of the filtered signal is logarithmized, and finally, a discrete cosine transform is performed to obtain the MFCC coefficients. The first 13 MFCC coefficients are extracted as basic features, and their first-order differences (velocity features) and second-order differences (acceleration features) are calculated to form a 39-dimensional acoustic feature vector. These features can characterize changes in acoustic properties during machining, such as changes in sound caused by tool wear or abnormal noise caused by bearing failure.

[0059] Perform Fourier analysis on the original driving current signal to extract the frequency components in the current waveform. First, perform discrete Fourier transform on the current signal to calculate the amplitude and phase of each frequency component. The formula is:

[0060]

[0061] in Indicates the frequency The complex representation of the current component, Indicates the time domain The current sampling value of the sampling point, is the number of sampling points, is an imaginary unit. Extract the fundamental component and each order harmonic component from it and calculate the total harmonic distortion rate:

[0062]

[0063] in is the hth harmonic amplitude, is the fundamental amplitude, The highest harmonic order considered is calculated. At the same time, the current RMS value, crest factor, and power factor are calculated. These current characteristics reflect the working status and health of the motor.

[0064] Perform differential operation on the original signal of the worktable displacement to calculate the velocity and acceleration. The formula is:

[0065]

[0066]

[0067] in represents the displacement value at time t, is the sampling time interval, For speed, is acceleration. At the same time, position error, trajectory deviation and acceleration characteristics are extracted. Position error Calculated as actual position With instruction position The difference:

[0068]

[0069] These motion characteristics can reflect the dynamic accuracy of the worktable and the status of the feed system.

[0070] The aforementioned features are combined to construct a multidimensional feature set, organized as a structured time series data matrix, with each row corresponding to a time point and each column corresponding to a feature dimension. This matrix contains status information for different machine tool components, but suffers from inconsistent dimensions. Therefore, the multidimensional feature set is standardized, using the Z-score method to normalize each feature to a distribution with a mean of 0 and a standard deviation of 1:

[0071]

[0072] in is the standardized eigenvalue, is the original eigenvalue, is the mean of the j-th column feature, is the standard deviation of the feature in column j. This standardization process eliminates the dimensional differences between different features and makes all features have the same weight in subsequent analysis.

[0073] Finally, principal component analysis (PCA) is performed on the standardized feature matrix. By calculating the eigenvalues and eigenvectors of the feature covariance matrix, the original high-dimensional features are projected into the principal component space. The principal components with a cumulative contribution rate of 95% are selected, significantly reducing feature dimensionality and redundancy, resulting in a more compact set of machine tool operating condition features.

[0074] For example, a certain type of CNC milling machine revealed a significant increase in the energy content of the 7th and 8th frequency bands after wavelet packet transform decomposition of the raw spindle vibration signal during machining. Dynamic window filtering of the motor temperature signal showed a steady upward trend. The MFCC features of the machining noise revealed abnormal changes in the 3rd and 5th coefficients. Fourier analysis of the current signal revealed an increase in the 5th harmonic content, with the total harmonic distortion rising to 8%. The position error calculated from the displacement signal fluctuated significantly at specific points. After organizing these features into a time-series matrix, normalization eliminated dimensional differences. Principal component analysis reduced the original 100-plus dimensional features to 15 principal components, preserving critical information. By analyzing these combined features, the characteristic patterns of early-stage spindle bearing failures were accurately captured, laying the foundation for subsequent condition identification.

[0075] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0076] The machine tool working condition feature set is input into the temporal encoding layer of the recursive convolutional hybrid memory network, and the feature sequence is obtained through position encoding and time step embedding processing;

[0077] The feature sequence is input into the multi-scale convolution layer of the recursive convolution hybrid memory network. The local feature patterns of different time windows are extracted through parallel 1×1, 3×3, and 5×5 convolution kernels. Channel attention weighting is performed on the convolution output features to obtain multi-scale temporal convolution features.

[0078] The multi-scale temporal convolutional features are input into the long short-term memory layer, and the long-term dependencies are captured using forward and backward bidirectional gated recurrent units. The information flow is controlled by the forget gate, input gate, and output gate to obtain long-term context features.

[0079] Perform attention fusion layer processing on long-term context features to calculate the pairwise interaction strength matrix between the three-dimensional features of spindle speed change, tool wear, and bearing status. Apply non-negative matrix factorization to the interaction strength matrix to extract potential correlation factors and obtain a compressed feature interaction representation.

[0080] A weighted directed graph structure is constructed based on the compressed feature interaction representation. The nodes are set as spindle speed change features, tool wear features, and bearing status features. The edge weights represent the interaction strength between features. The node feature information is aggregated through graph convolution operations to obtain a co-dependency graph.

[0081] Continuous wavelet transform is performed on the co-dependency graph to map the time domain signal to the time-frequency plane. The energy distribution and phase relationship of each frequency band are calculated. The characteristic frequency components are extracted by the maximum energy tracking algorithm to obtain the time-frequency characteristic spectrum.

[0082] The time-frequency feature spectrum and long-term context features are fused through a gated update mechanism. Adaptive weights are assigned according to the energy distribution of different frequency bands. The current observation and historical state information are integrated to obtain the state characteristics of machine tool components.

[0083] Specifically, the machine tool operating condition feature set is input into the temporal encoding layer of a recursive convolutional hybrid memory network, which consists of a feature embedding unit and a position encoding unit. The feature embedding unit is a fully connected layer that maps feature vectors of different dimensions into a unified d-dimensional space, typically set to 512. Specifically, each feature vector in the machine tool operating condition feature set is multiplied by a weight matrix and added with a bias term, unifying the different dimensions of the original features into a uniform dimension. For example, features of different dimensions, such as original spindle vibration, temperature, and acoustic features, are mapped into vectors of consistent dimensions through a linear transformation to facilitate subsequent processing. The position encoding unit adds encoding information representing temporal position, using a combination of sine and cosine functions to generate a fixed pattern. For each position and dimension in the sequence, the corresponding encoding value is calculated. The position encoding is then added to the feature embedding to produce a feature sequence. This processed feature sequence contains both the semantic information of the original features and temporal position information, enabling the model to distinguish data at different time points. The feature sequence is fed into a multi-scale convolutional layer, which consists of multiple parallel convolutional branches with kernels of three different sizes: 1×1, 3×3, and 5×5. 1×1 convolutions primarily fuse information across the channel dimension, 3×3 convolutions capture medium-range temporal correlations, and 5×5 convolutions focus on longer-range temporal patterns. Each convolutional branch is followed by batch normalization and a ReLU activation function to enhance nonlinear expression and prevent gradient vanishing. After the convolution operation, channel-wise attention weighting is applied to the output features of all branches. This recalibrates the features by learning the importance weights for each channel, emphasizing important channels and suppressing irrelevant ones. Finally, the weighted features are combined to form a multi-scale temporal convolutional feature, which captures local pattern information at different time scales.

[0084] The multi-scale temporal convolutional features are input into the long short-term memory layer, which uses a forward and backward bidirectional gated recurrent unit (BiGRU) structure that can simultaneously consider past and future contextual information. Each GRU unit contains two gating mechanisms: a reset gate and an update gate. The reset gate determines how to combine the new input with the previous state, and the update gate determines how much previous state information to retain. The forward GRU processes from the beginning to the end of the time series, capturing forward temporal information; the backward GRU processes from the end to the beginning of the time series, capturing reverse temporal information. The results of the bidirectional processing are spliced in the feature dimension to obtain long-term contextual features containing the complete context. This structure is particularly suitable for capturing the gradual process of machine tool component status changes over time, such as long-term development trends such as bearing wear and temperature increase.

[0085] The long-term context features are processed using an attention fusion layer specifically designed to analyze the interactions between three key factors: spindle speed variation, tool wear, and bearing condition. The long-term context features are first decomposed into sub-features corresponding to the three factors. Then, the interaction strength matrix is calculated between each pair of factors, such as the correlation strength between spindle speed and bearing condition, spindle speed and tool wear, and tool wear and bearing condition. These interaction matrices are then reduced in dimensionality using non-negative matrix factorization techniques to extract potential correlation factors, eliminate redundant information, and obtain a compressed representation of feature interactions. This decomposition method preserves key interaction patterns while significantly reducing computational complexity.

[0086] A weighted directed graph structure is constructed based on a compressed feature interaction representation. The nodes of this graph represent the spindle speed change characteristics, tool wear characteristics, and bearing status characteristics. The edges represent the interactions between nodes, and the edge weights indicate the strength of the interactions. The advantage of the graph structure is that it can represent the complex relationship network between multiple components, considering not only direct influences but also simulating indirect transmission effects. For example, a change in spindle speed may directly affect the bearing status, or it may indirectly affect the bearing status by affecting tool wear. Through graph convolution operations, each node can aggregate information from adjacent nodes, enabling the transmission and fusion of information within the graph structure, ultimately obtaining a co-dependency graph that contains the complete interaction relationships.

[0087] A continuous wavelet transform is performed on the co-dependency graph to convert the time-domain signal into the time-frequency domain. The continuous wavelet transform uses wavelet functions of different scales and positions to convolve with the signal, obtaining frequency information while maintaining temporal resolution, making it particularly suitable for analyzing non-stationary signals. For machine tool monitoring, this transform can precisely locate the time points and frequency characteristics of specific fault modes. After the transformation, the energy distribution and phase relationship of each frequency band are calculated, and the main characteristic frequency components are identified using a maximum energy tracking algorithm. For example, for bearing faults, the energy changes of the characteristic frequencies of the inner race, outer race, and rolling elements can be accurately extracted; for tool wear, the evolution of cutting vibration frequencies can be tracked. These analysis results are organized into a time-frequency feature spectrum, which contains a complete representation of the fault characteristics in both time and frequency dimensions.

[0088] The time-frequency feature spectrum is fused with long-term contextual features through a gated update mechanism. This mechanism uses a set of learnable gating units to determine how to combine frequency and time domain information. The gating units calculate weight coefficients based on the energy distribution of different frequency bands and weight the frequency domain features, focusing on frequency bands with significant energy. At the same time, the gating units also control the retention ratio of historical state information, retaining more historical information when the state is stable and preferring current observations when the state suddenly changes. This adaptive fusion mechanism balances long-term trends and instantaneous changes to obtain comprehensive and accurate machine tool component state characteristics, providing a reliable basis for subsequent state classification and life prediction.

[0089] For example, a high-speed machining center is equipped with vibration sensors, temperature sensors, acoustic sensors, current sensors, and displacement sensors to monitor the operating status of various components in real time. After collecting data for a period of time, feature extraction is used to obtain a machine condition feature set, including spindle vibration frequency domain features, motor temperature series, acoustic feature vectors, current features, and motion features. This feature set is then processed in a recursive convolutional hybrid memory network. First, the temporal encoding layer maps features of different dimensions into a 512-dimensional vector and adds positional encoding. A multi-scale convolutional layer uses 1×1, 3×3, and 5×5 convolution kernels to extract feature patterns at different time scales in parallel. The long-short-term memory layer captures long-term dependencies. The attention fusion layer analyzes the interactions between spindle speed changes, tool wear, and bearing status, constructing a co-dependency graph. A continuous wavelet transform maps this relationship graph to the time-frequency plane, identifying characteristic frequency patterns of bearing anomalies. A gated update mechanism fuses the time-frequency features with temporal context features to obtain a feature representation that accurately reflects the health status of the component.

[0090] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0091] The state features of the machine tool components are reorganized into a two-dimensional feature map of the shape C×H×W, where C is the number of channels, H and W are the height and width of the feature map, and the initial two-dimensional feature representation is obtained;

[0092] The initial two-dimensional feature representation is input into the initial convolution layer of SE-ResNet. Through the convolution operation of 7×7 convolution kernel and stride 2, the low-level spatial features are extracted to obtain the basic feature map;

[0093] The basic feature map is input into the residual block with the SE module, and the backbone path is constructed through two 3×3 convolutional layers and skip connections to obtain the residual features;

[0094] Apply global average pooling operation to the residual features to compress the feature map of each channel into a single value to obtain the channel descriptor;

[0095] The channel descriptor is input into a two-layer fully connected network. The first layer uses ReLU activation after dimensionality reduction, and the second layer uses Sigmoid activation after dimensionality recovery to obtain the importance weight of each channel;

[0096] Apply importance weights to each channel of the residual feature and perform channel-wise weighting to complete feature recalibration and obtain a feature map with attention.

[0097] The feature map with attention is passed through the four stages of SE-ResNet in sequence, with the number of channels being 64, 128, 256, and 512 respectively. Each stage contains multiple residual blocks with SE modules to obtain multi-scale feature representation.

[0098] A global average pooling operation is performed on the multi-scale feature representation to convert the feature map into a 512-dimensional feature vector, capturing the interaction pattern between the machine tool spindle, feed and control system to obtain the machine tool system state representation.

[0099] Specifically, the machine tool component state features are reorganized into a two-dimensional feature map of shape C×H×W, where the initial number of channels C is set to 64, and the feature map height H and width W are both set to 7. This reorganization process reorganizes a one-dimensional vector according to a preset arrangement, similar to converting a one-dimensional audio signal into a spectrogram. Specifically, the one-dimensional feature vector of length d (assuming d = 3136) is rearranged into a three-dimensional tensor of 64×7×7, with each channel representing a feature type, such as spindle vibration, temperature, or acoustic features, to form the initial two-dimensional feature representation. This initial two-dimensional feature representation is input into the initial convolutional layer of the SE-ResNet, which uses a 7×7 convolution kernel and a stride of 2 for feature extraction. The convolution operation slides the convolution kernel across the input feature map and calculates the dot product between the local region and the convolution kernel to extract local features. Larger convolution kernels (7×7) capture more spatial information and are suitable for extracting low-level texture and edge features. The output of the convolutional layer is processed by batch normalization and ReLU activation function to obtain the basic feature map, and the size becomes 64×4×4 (due to the stride of 2, the size is halved).

[0100] The base feature map is input into a residual block with a SE module. The residual block is a core component of the ResNet network, consisting of two 3×3 convolutional layers forming the backbone path, while retaining a skip connection. In the backbone path, the first 3×3 convolutional layer is followed by batch normalization and ReLU activation, while the second 3×3 convolutional layer is followed by only batch normalization. The skip connection directly passes the input features to the output of the backbone path for summation, forming a residual learning structure. This design addresses the difficulty of training deep networks, enabling the network to learn a residual mapping rather than a full mapping. The resulting residual features maintain the size of the feature map. Global average pooling is applied to the residual features, compressing the feature map of each channel into a single value. This is calculated by taking the average value across all spatial locations of each channel. For example, for a residual feature with a shape of 64×4×4, global average pooling results in a feature with a shape of 64×1×1, meaning that each channel is compressed into a scalar value. These values constitute channel descriptors, which describe the global response strength of each channel.

[0101] The channel descriptors are fed into a two-layer fully connected network. The first layer reduces the number of channels from 64 to 4 (a dimensionality reduction ratio of 16) and introduces nonlinearity using the ReLU activation function. The second layer restores the number of channels from 4 to 64 and uses the Sigmoid activation function to constrain the output values between 0 and 1, forming the importance weights for each channel. These two layers of fully connected networks learn the correlations between channels and generate weight coefficients that reflect the importance of each channel.

[0102] The calculated importance weights are applied to each channel of the residual features, performing a channel-wise weighting. This involves multiplying the feature map of each channel by the corresponding weight coefficient to complete the feature recalibration. This process selectively enhances the features of important channels and suppresses those of unimportant channels, implementing a channel-attention mechanism and resulting in an attended feature map. The channel-attention mechanism enables the network to adaptively adjust the importance of different feature channels. This mechanism is particularly suitable for processing multi-sensor data from CNC machine tools, as the importance of each sensor varies under different working conditions. The attended feature map is sequentially passed through four stages of SE-ResNet, with 64, 128, 256, and 512 channels, respectively. Each stage contains multiple residual blocks with SE modules. In the first residual block of each stage (except the first one), a convolution operation with a stride of 2 is used for downsampling, halving the spatial dimension of the feature map while increasing the number of channels. This design enables the network to gradually reduce the spatial dimension while increasing the number of channels, capturing more abstract and high-level feature representations. The residual blocks in each stage contain SE modules to continuously adjust the importance of each channel. The resulting multi-scale feature representation is 512×1×1.

[0103] Finally, a global average pooling operation is performed on the multi-scale feature representation, converting the feature map into a 512-dimensional feature vector. Since the previous processing has reduced the spatial dimension of the feature map to 1×1, the global pooling here primarily converts the three-dimensional tensor into a one-dimensional vector. This 512-dimensional feature vector captures the complex interactions between the machine tool spindle, feed system, and control system, forming a representation of the machine tool system state.

[0104] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0105] The machine tool component state characteristics and the machine tool system state representation are spliced in the feature dimension to construct a joint feature vector and obtain a fused feature vector;

[0106] A feature gating mechanism is introduced into the fused feature vector. When the spindle load fluctuates greatly under high-speed cutting conditions, the weight of the spindle-related features is increased. When the feed system starts and stops frequently, the weight of the feed system features is increased. This results in a weighted fusion feature that is adaptive to the working conditions.

[0107] The weighted fusion features are input into a classifier network consisting of three fully connected layers with 256, 128, and 5 nodes respectively to obtain the probability distribution of the machine tool state.

[0108] Based on the probability distribution of machine tool status, a multi-level threshold judgment standard is set. When the vibration spectrum of the machine tool spindle reaches a peak at the characteristic frequency of the bearing fault, it is judged as a bearing abnormality; when the tool wear index exceeds the threshold, it is judged as a tool abnormality, and the preliminary status judgment result is obtained;

[0109] The preliminary state determination results are smoothed by a time sliding window. When the states are consistent within 5 consecutive time windows during the machining process, the state transition is confirmed. In the case of an empty cutting state, this is extended to 8 windows to obtain a stable state determination.

[0110] The stable state judgment is matched with the historical abnormal pattern library for similarity. When abnormal vibration occurs during heavy cutting, the cutting chatter pattern is matched first. When abnormal temperature rise occurs after long-term operation, the bearing wear pattern is matched first to determine the fault type.

[0111] Based on the fault type determination, the impact of the fault on the machining quality is evaluated by combining the current cutting parameters, workpiece material hardness, and tool type to obtain a fault level assessment.

[0112] Based on the fault level assessment, the machine tool health status is divided into four levels: normal operation, minor abnormality, moderate fault and serious fault. In the case of a serious fault, an emergency stop signal is triggered. In the case of a minor abnormality, only an early warning is issued without intervening in production, thus obtaining the machine tool health status judgment result.

[0113] Specifically, the machine tool component state features extracted by the Transformer network and the machine tool system state representation extracted by the SE-ResNet network are concatenated in the feature dimension. This process directly concatenates the two feature vectors into a longer vector. Assuming the dimension of the machine tool component state features is 256 and the dimension of the machine tool system state representation is 512, the resulting fused feature vector has a dimension of 768. This concatenation operation preserves the full information of the two different feature extraction methods, including the temporal dependencies captured by the Transformer and the spatial interaction features extracted by the SE-ResNet, forming a more comprehensive feature representation. A feature gating mechanism is introduced into the fused feature vector to achieve adaptive feature weighting based on the working condition. The gating mechanism is a technique that dynamically adjusts feature weights based on the current working condition. This control is achieved by setting a set of learnable parameters and multiplying them by the corresponding feature components to control the flow of information. In the intelligent monitoring of CNC machine tools, specialized weight adjustment strategies are designed for different working conditions. For example, when high-speed cutting conditions and large spindle load fluctuations are detected, the gating mechanism automatically increases the weight of features related to the spindle. When frequent starts and stops of the feed system are detected, the weight of the feed system features is increased. This adaptive adjustment is achieved through a combination of rule-based and learning methods. Basic rules are first established through expert knowledge, and then weight parameters are gradually optimized through machine learning, enabling the model to prioritize the most relevant features for different machining conditions.

[0114] The condition-adaptive weighted fusion features are input into the classifier network, which consists of three fully connected layers with 256, 128, and 5 nodes, respectively. The first layer maps the 768-dimensional weighted fusion features to a 256-dimensional space, using the ReLU activation function to introduce a nonlinear transformation. The second layer further compresses the features to 128 dimensions, also using ReLU activation. The final layer maps the features into a 5-dimensional output corresponding to five machine states: normal, spindle abnormality, feed abnormality, tool abnormality, and control system abnormality. The output layer uses the Softmax activation function to convert the raw scores into a probability distribution, with each class assigned a probability value between 0 and 1, and the sum of all class probabilities equal to 1. The resulting machine state probability distribution intuitively reflects the model's confidence in the current machine state. Based on the machine state probability distribution, a multi-level threshold judgment criterion is set to identify abnormal states. This judgment criterion combines probability thresholds and feature indicators. For the probability threshold, a probability greater than 0.6 is typically set for the main class to be considered definitively. Specific abnormality characteristics are also considered. For example, if a bearing anomaly is suspected, the spindle vibration spectrum is further checked to see if significant peaks appear at the bearing fault characteristic frequencies (such as the inner race, outer race, and rolling element characteristic frequencies). For tool anomalies, the tool wear indicator is checked to see if it exceeds a preset threshold. This dual determination mechanism combines the advantages of data-driven and mechanism-based analysis, improving both accuracy and interpretability, and ultimately providing a preliminary state determination result.

[0115] A time sliding window smoothing process is performed on the preliminary state judgment results to eliminate the influence of instantaneous fluctuations. Sliding window smoothing is a time series filtering method that considers multiple consecutive judgment results in the time dimension and obtains more stable judgments through majority voting or weighted averaging. In CNC machine tool monitoring, the stability of state judgment is crucial to avoid misjudgments caused by instantaneous interference. In specific implementation, during normal processing, the state judgment must be consistent in 5 consecutive time windows (each window is usually 1-3 seconds) before the state transition is confirmed; in the air-cut state, due to the smaller signal fluctuations and the more obvious interference effects, the confirmation is extended to 8 consecutive windows of consistency. This differentiated strategy adapts to the signal characteristics under different working conditions and obtains stable and reliable state judgment.

[0116] The stable state judgment is matched against the historical abnormal pattern library for similarity to further refine the fault type. The historical abnormal pattern library is a knowledge base built from long-term accumulation of typical fault cases. Each fault mode contains key feature descriptions and corresponding diagnostic information. The similarity matching process uses a distance-based comparison method to calculate the Euclidean distance or cosine similarity between the current abnormal feature and each pattern in the library, and the pattern with the highest similarity is selected as the matching result. The matching process also takes into account the current working condition information. For example, when abnormal vibration occurs during heavy cutting, the cutting chatter mode is prioritized; when the machine tool experiences abnormal temperature rise after long-term operation, the bearing wear mode is prioritized. This matching strategy that takes working conditions into consideration improves the accuracy and specificity of fault diagnosis.

[0117] Based on the fault type determination, a fault level assessment is performed in conjunction with the current machining parameters. This assessment considers the fault characteristics along with machining parameters (such as spindle speed, feed rate, and depth of cut), workpiece material properties (such as hardness and strength), and tool type. For example, the impact of the same bearing vibration anomaly when machining carbide is much greater than when machining aluminum alloy; tool wear develops more rapidly under high-speed cutting than under low-speed cutting. By comprehensively analyzing these factors, the fault's severity, development trends, and potential impact on machining quality are calculated, resulting in a structured fault level assessment.

[0118] Based on fault level assessment, the machine tool health status is divided into four levels: normal operation, minor anomaly, moderate failure, and severe failure, and corresponding measures are taken according to the level. During normal operation, continuous monitoring is performed without intervention; during minor anomalies, a warning message is issued, but the machine tool continues production and an inspection is recommended at an appropriate time; during moderate failures, a warning is issued and a timely shutdown and maintenance is recommended; and during severe failures, an emergency shutdown signal is triggered to prevent the failure from escalating and causing safety accidents. This graded response strategy balances production continuity and equipment safety, avoiding unnecessary downtime and preventing major equipment accidents.

[0119] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0120] Based on the results of machine tool health status determination, a health index scoring system for key machine tool components is constructed. Health scores are assigned to the spindle, feed, bearings, and tools respectively to obtain the health index of machine tool components.

[0121] Extract the current machining process parameters from the machine tool control system, including spindle speed, feed rate, cutting depth, and workpiece material information, establish the working condition-wear relationship, and obtain the working condition load assessment results;

[0122] Retrieve historical cases similar to the current fault type from the historical fault database, extract the fault development process, maintenance records and component life data, and obtain a reference life data set;

[0123] The machine tool component health index, working load assessment results, and reference life data set are fused to calculate the component degradation rate and trend to obtain the component remaining life prediction value;

[0124] Set multi-level warning thresholds based on the predicted remaining life of the component. When the predicted life is lower than the threshold, the corresponding level of warning signal is triggered to obtain the fault warning level;

[0125] Based on the fault warning level, combined with the machine tool maintenance manual and historical maintenance experience, specific suggestions for component replacement, maintenance or parameter adjustment are generated, and a maintenance suggestion list is obtained;

[0126] Based on the maintenance recommendation list, the maintenance tasks are sorted by urgency, and the time and cost required for maintenance are estimated to obtain a maintenance plan;

[0127] The component remaining life prediction value, fault warning level, maintenance recommendation list and maintenance planning scheme are integrated into a structured report, which is presented in graphic and text form through edge nodes and cloud platforms to obtain a machine tool diagnostic report including maintenance recommendations.

[0128] Specifically, a health index scoring system for key machine tool components is constructed based on the machine tool health status assessment results. This system is a numerical method for quantifying component health. By mapping various characteristic indicators to scores between 0 and 100, it intuitively reflects component health. In implementation, independent scoring models are developed for key components such as the spindle, feed system, bearings, and tool. Each model comprehensively considers the component's characteristic indicators. For example, the spindle score considers vibration amplitude, spectral characteristics, and temperature rise rate; the bearing score considers characteristic frequency energy ratio, temperature, and noise level; the feed system score considers positioning accuracy, repeatability, and tracking error; and the tool score considers cutting force variation, wear width estimation, and machined surface roughness. The health score for each component is calculated by weighted average, forming a structured machine tool component health index matrix. Current machining process parameters are extracted from the machine tool control system. These parameters directly affect the load conditions and wear rates of machine tool components. Spindle speed determines the motion speed and cyclic load of bearings; feed speed affects the dynamic load of the feed system; depth of cut is related to spindle torque and cutting force; and workpiece material information determines machining difficulty and tool wear rate. These parameters are read directly through the CNC system's communication interface, eliminating the need for additional sensors. Once these parameters are acquired, a working condition-wear correlation model is established. This model, based on a combination of empirical formulas and data-driven approaches, maps machining parameters to load factors and wear rates. For example, for tool wear, the wear rate index is calculated based on cutting speed, feed rate, depth of cut, and workpiece material hardness. For bearing life, the equivalent load is calculated based on spindle speed, radial force, and axial force. These calculations determine the load level and expected wear rate of each component under the current working conditions, forming a working condition load assessment result.

[0129] A content-based similarity retrieval method is used to retrieve historical cases similar to the current fault type from the historical fault database. First, a query vector is constructed based on the currently determined fault type, containing key elements such as fault location, symptom characteristics, and operating parameters. The similarity between this vector and each record in the historical case database is then calculated, and the cases with the highest similarity are selected as references. Three key pieces of information are extracted from these historical cases: the fault development process (describing the timeline of how the fault progressed from early symptoms to complete failure), maintenance records (documenting the specific repair measures and results), and component life data (recording the actual service life of similar components under similar operating conditions). This information is structured and organized to form a reference life dataset, providing empirical evidence for remaining life prediction. The machine tool component health index, operating load assessment results, and the reference life dataset are fused to calculate component degradation rates and trends. This data fusion utilizes a weighted integration approach, combining the current measured health status, expected wear due to operating loads, and the statistical patterns of historical data to comprehensively predict component degradation. The calculation begins by determining the component's current degradation level based on the health index. The expected degradation rate is then adjusted based on the operating load assessment results. Finally, the degradation model parameters are modified by referring to the statistical distribution in the historical data set. For example, for a bearing component, if the health index is 85 (mild degradation), the current load is high, and historical data indicates an average operating time of approximately 1,000 hours under similar conditions, a comprehensive calculation is used to estimate the remaining life. This multi-source data fusion prediction method considers both real-time conditions and the impact of operating conditions and historical experience, improving the comprehensiveness and accuracy of the prediction.

[0130] Multi-level warning thresholds are set based on the predicted remaining life of a component, enabling tiered early warning. Warning thresholds are typically categorized into three levels: caution (remaining life less than twice the equipment's planned maintenance cycle), warning (remaining life less than one planned maintenance cycle), and emergency (remaining life less than a safety margin, typically 48 hours). Different warning levels correspond to different response strategies: caution only records and conducts regular monitoring, warning requires planned maintenance, and emergency requires immediate downtime for maintenance. When a component's predicted remaining life falls below a threshold, a warning signal of the corresponding level is triggered, resulting in a fault warning level determination. Maintenance recommendations are generated for each fault warning level. This process is based on a combination of rule-based reasoning and case-based retrieval, incorporating standard procedures from machine tool maintenance manuals and best practices from historical maintenance experience. For each fault type and warning level, the system provides pre-defined maintenance action templates, including component replacement instructions, maintenance procedures, or recommended operating parameter adjustments. These templates are parameterized and customized based on the specific fault scenario and machine tool model to generate targeted maintenance recommendations. For example, for bearing warnings, maintenance recommendations include specific bearing models, disassembly procedures, replacement precautions, and commissioning verification methods; for tool wear warnings, recommendations include tool regrinding or replacement, cutting parameter adjustment plans, etc. These suggestions are organized into a structured maintenance recommendation list.

[0131] Based on the maintenance recommendation list, maintenance tasks are prioritized and planned according to urgency, maintenance difficulty, and resource constraints. The prioritization process considers several factors: fault warning level (urgent level takes precedence over warning level), component importance (critical components take precedence over less important components), maintenance dependencies (some maintenance must be performed in a specific order), and resource constraints (such as the availability of personnel, tools, and spare parts). After prioritization, the total time and cost required for maintenance are estimated based on the standard working hours and required resources for each maintenance task, resulting in a maintenance plan that includes a list of maintenance items, execution sequence, required time, personnel requirements, and cost budget. The component remaining life prediction, fault warning level, maintenance recommendation list, and maintenance plan are integrated into a structured report. The report uses a standardized format and includes summary information (outlining key findings and urgent matters), detailed condition assessment (detailed analysis of the health status of each component), predictive maintenance recommendations (life prediction and maintenance planning), and historical comparative analysis (comparison with historical data). The report content is converted into data format and presented through the local display interface of the edge node and the remote access interface of the cloud platform. It supports graphical display (such as health status dashboard, trend chart and early warning prompts) and detailed text descriptions to form a machine tool diagnostic report for maintenance personnel and managers to refer to and make decisions.

[0132] The above describes the CNC machine tool intelligent monitoring method based on the Internet of Things in the embodiment of the present application. The following describes the CNC machine tool intelligent monitoring system based on the Internet of Things in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the intelligent monitoring system for CNC machine tools based on the Internet of Things includes:

[0133] The transmission module is used to collect CNC machine tool spindle vibration, motor temperature, processing noise, drive current and worktable displacement data, and transmit them synchronously to the edge node through the industrial bus to obtain the original machine tool operation data;

[0134] A processing module is used to perform signal processing on the original data of the machine tool operation to obtain a machine tool working condition feature set;

[0135] An analysis module is used to input the machine tool working condition feature set into a recursive convolutional hybrid memory network, construct a co-dependency relationship diagram of spindle speed change, tool wear, and bearing status, realize state parameter prediction through joint analysis in the time and frequency domains, and obtain the state characteristics of the machine tool components;

[0136] A conversion module is used to convert the state features of the machine tool components into a two-dimensional feature map, extract the interaction features between the spindle, feed, and control systems through SE-ResNet, and obtain a machine tool system state representation;

[0137] An execution module, configured to integrate the machine tool component state characteristics with the machine tool system state representation, perform state classification, and obtain a machine tool health state determination result;

[0138] The prediction module is used to perform remaining life prediction and fault warning based on the machine tool health status judgment result, combined with processing parameters and historical fault data, to obtain a machine tool diagnosis report including maintenance suggestions.

[0139] Through the collaborative cooperation of the above components and the collection of multi-source sensor data, all-round and multi-dimensional state perception of the machine tool is achieved, the defect of limited information of a single sensor is overcome, and the data foundation is laid for a comprehensive understanding of the operation status of the machine tool; professional signal processing is performed on the original data of the machine tool operation, and specific processing methods are adopted for different types of signals, such as applying wavelet packet transform to extract frequency domain features of vibration signals, performing dynamic window sliding average filtering on temperature signals, calculating Mel-frequency cepstral coefficients on acoustic signals, performing Fourier analysis on current signals, and performing differential operations on displacement signals, thereby extracting a rich set of working condition features and fully mining the state information contained in various signals; the machine tool working condition feature set is input into the recursive convolution hybrid memory network, and through the combination of multi-scale convolution and long-short-term memory structure, the time series correlation and co-dependency analysis of spindle speed changes, tool wear, and bearing status are realized, and a state co-dependency diagram is constructed and combined through time-frequency domain. Analysis overcomes the shortcomings of traditional methods in handling complex component interactions; by converting machine tool component state features into two-dimensional feature maps and inputting them into the SE-ResNet network, its channel attention mechanism is used to adaptively adjust the importance of different feature channels, thereby enhancing the ability to extract interaction features between the spindle, feed, and control systems and improving the accuracy of state recognition; by fusing temporal features and spatial features, and combining them with the feature gating mechanism for adaptive working conditions, the feature weights are dynamically adjusted for different processing conditions, solving the problem of insufficient recognition accuracy of traditional methods under complex and changing working conditions; based on the health status judgment results and historical fault data, a component health index scoring system is established, and the remaining life is predicted by combining processing parameters and working condition loads, realizing the transition from passive response to active prediction, and providing a scientific basis for predictive maintenance; generating targeted maintenance suggestions and formulating maintenance planning plans according to the warning level, reducing maintenance costs and reducing downtime. The innovative algorithm architecture of the present invention combines the recursive convolutional hybrid memory network with the SE-ResNet network, giving full play to the advantages of the recursive convolutional hybrid memory network in time-frequency feature extraction and co-dependency modeling, as well as the strengths of SE-ResNet in spatial feature extraction. The synergistic effect of the two algorithms significantly improves the ability to recognize the status of machine tools under complex working conditions, and solves the technical problem that traditional single algorithms are difficult to simultaneously take into account temporal correlation and spatial interaction features; in particular, the recursive convolutional hybrid memory network can effectively process multi-source heterogeneous data and construct a co-dependency graph between components, while the channel attention mechanism of SE-ResNet can adaptively highlight important feature channels. The combined application of these two mechanisms in machine tool status monitoring enables the system to accurately identify subtle anomalies under complex working conditions, greatly improving the accuracy and timeliness of fault warnings, and realizing accurate assessment of the health status of machine tools and accurate prediction of fault development trends.

[0140] above Figure 2The intelligent monitoring system for CNC machine tools based on the Internet of Things in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The intelligent monitoring device for CNC machine tools based on the Internet of Things in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0141] Figure 3 This is a schematic diagram of the structure of an IoT-based intelligent monitoring device for CNC machine tools provided by an embodiment of the present invention. This IoT-based intelligent monitoring device for CNC machine tools 300 may vary significantly depending on configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors), memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage medium 330 may be either transient or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown), each of which may include a series of instruction operations within the IoT-based intelligent monitoring device for CNC machine tools 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, executing the series of instruction operations stored in the storage medium 330 on the IoT-based intelligent monitoring device for CNC machine tools 300 to implement the steps of the IoT-based intelligent monitoring method for CNC machine tools.

[0142] The IoT-based CNC machine tool intelligent monitoring device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the IoT-based intelligent monitoring device for CNC machine tools shown does not constitute a limitation on the IoT-based intelligent monitoring device for CNC machine tools provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0143] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions. When the instructions are executed on a computer, the computer executes the steps of the intelligent monitoring method for CNC machine tools based on the Internet of Things.

[0144] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0145] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling an IoT-based CNC machine tool intelligent monitoring device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0146] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent monitoring method for CNC machine tools based on the Internet of Things, characterized in that: The method comprises: Collect CNC machine tool spindle vibration, motor temperature, processing noise, drive current and worktable displacement data, and transmit them synchronously to edge nodes via the industrial bus to obtain the original machine tool operation data; performing signal processing on the raw data of the machine tool operation to obtain a machine tool working condition feature set; The machine tool working condition feature set is input into the recursive convolution hybrid memory network to construct a co-dependency graph of spindle speed change, tool wear, and bearing status, and state parameter prediction is achieved through joint analysis in the time and frequency domains to obtain machine tool component state features, including: inputting the machine tool working condition feature set into the temporal coding layer of the recursive convolution hybrid memory network, and obtaining a feature sequence through position coding and time step embedding processing; inputting the feature sequence into the multi-scale convolution layer of the recursive convolution hybrid memory network, extracting local feature patterns of different time windows through parallel 1×1, 3×3, and 5×5 convolution kernels, performing channel attention weighting on the convolution output features, and obtaining multi-scale temporal convolution features; inputting the multi-scale temporal convolution features into the long-short-term memory layer, using forward and backward bidirectional gated recurrent units to capture long-term dependencies, and controlling the flow of information through forgetting gates, input gates, and output gates to obtain long-term context features; performing attention weighting on the long-term context features The intention and force fusion layer is processed to calculate the pairwise interaction intensity matrix between the three-dimensional features of spindle speed change, tool wear, and bearing status, apply non-negative matrix decomposition to the interaction intensity matrix, extract potential correlation factors, and obtain a compressed feature interaction representation; based on the compressed feature interaction representation, a weighted directed graph structure is constructed, and the nodes are set as spindle speed change features, tool wear features, and bearing status features. The edge weights represent the interaction intensity between features, and the node feature information is aggregated through graph convolution operations to obtain a co-dependency graph; a continuous wavelet transform is performed on the co-dependency graph to map the time domain signal to the time-frequency plane, calculate the energy distribution and phase relationship of each frequency band, and extract the characteristic frequency components through the maximum energy tracking algorithm to obtain a time-frequency feature spectrum; the time-frequency feature spectrum and the long-term context feature are fused through a gated update mechanism, and adaptive weights are assigned according to the energy distribution of different frequency bands. The current observation and historical status information are integrated to obtain the state characteristics of the machine tool components; The state features of the machine tool components are converted into a two-dimensional feature map, and the interaction features between the spindle, feed, and control systems are extracted through SE-ResNet to obtain the state representation of the machine tool system; fusing the state characteristics of the machine tool components with the state representation of the machine tool system, performing state classification, and obtaining a machine tool health state determination result; Based on the machine tool health status determination result, combined with processing parameters and historical fault data, remaining life prediction and fault warning are performed to obtain a machine tool diagnosis report including maintenance suggestions.

2. The intelligent monitoring method for CNC machine tools based on the Internet of Things according to claim 1 is characterized in that: The data of spindle vibration, motor temperature, machining noise, drive current and worktable displacement of the CNC machine tool are collected and synchronously transmitted to the edge node via the industrial bus to obtain the original data of the machine tool operation, including: Arrange a vibration sensor array at key locations on the CNC machine tool spindle, obtain the spindle's radial and axial vibration characteristics through multi-point measurement, and obtain the original spindle vibration signal; Deploy a temperature sensor network in the hot spots of CNC machine tool drive motors, servo motors, and control cabinets to build a temperature field distribution monitoring system and obtain the original motor temperature signal; Acoustic sensors are set up around the machining area of the CNC machine tool, and beamforming is used to perform spatial direction collection of machining noise to obtain the original machining noise signal; A current monitoring module is embedded in the power supply circuit of the CNC machine tool spindle motor and feed system to collect the current and harmonic components of each phase and obtain the original drive current signal; A displacement sensing network is built on the worktable and guide rail system of the CNC machine tool to monitor the motion trajectory and precision deviation of the worktable and obtain the original displacement signal of the worktable; Time synchronization and alignment are performed on the spindle vibration original signal, the motor temperature original signal, the machining noise original signal, the drive current original signal, and the worktable displacement original signal to establish a unified time reference and obtain a time-aligned multi-source signal; Transmitting the time-aligned multi-source signals to the edge computing layer through an industrial-grade communication protocol to build a data transmission channel; Data quality assessment is performed at the edge computing layer, validity verification and noise elimination are performed on multi-source signals, and the original data of machine tool operation is obtained.

3. The intelligent monitoring method for CNC machine tools based on the Internet of Things according to claim 2 is characterized in that: The signal processing of the raw data of the machine tool operation to obtain a machine tool working condition feature set includes: Applying wavelet packet transform to the original spindle vibration signal, selecting 5 decomposition layers, extracting the energy distribution of each frequency band, and obtaining the frequency domain characteristics of the spindle vibration; Performing a sliding average filter on the original motor temperature signal, dynamically adjusting the window size according to the temperature change rate to eliminate instantaneous interference and obtain a smoothed temperature sequence; Calculating Mel-frequency cepstral coefficients of the processed noise original signal, extracting the first 13 coefficients and their first-order differences and second-order differences to form an acoustic feature vector; Performing Fourier analysis on the original driving current signal to calculate the fundamental component and each order harmonic content, and extracting the current effective value, peak factor and power factor to obtain the current characteristics; Performing differential operation on the original displacement signal of the worktable to calculate the velocity and acceleration respectively, and extracting the position error, trajectory deviation and acceleration characteristics to obtain the motion characteristics; Combining the main shaft vibration frequency domain features, the smoothed temperature sequence, the acoustic feature vector, the current feature, and the motion feature to construct a multidimensional feature set, which is organized into a structured time series data matrix, where each row corresponds to a time point and each column corresponds to a feature dimension; The multidimensional feature set is standardized, and each feature is normalized to a distribution with a mean of 0 and a standard deviation of 1 using a Z-score to eliminate dimensional differences and obtain a standardized feature matrix; A principal component analysis is performed on the standardized feature matrix, and the principal components with a cumulative contribution rate of 95% are retained to reduce feature redundancy and obtain a machine tool working condition feature set.

4. The intelligent monitoring method for CNC machine tools based on the Internet of Things according to claim 1 is characterized in that: The state features of the machine tool components are converted into a two-dimensional feature map, and the interaction features between the spindle, feed, and control systems are extracted through SE-ResNet to obtain the state representation of the machine tool system, including: Reorganize the state features of the machine tool components into a two-dimensional feature map of shape C×H×W, where C is the number of channels, H and W are the height and width of the feature map, to obtain an initial two-dimensional feature representation; The initial two-dimensional feature representation is input into the initial convolution layer of SE-ResNet, and a convolution operation with a 7×7 convolution kernel and a stride of 2 is performed to extract low-level spatial features to obtain a basic feature map; The basic feature map is input into the residual block with the SE module, and the backbone path is constructed through two 3×3 convolutional layers and skip connections to obtain the residual feature; Applying a global average pooling operation to the residual features to compress the feature map of each channel into a single value to obtain a channel descriptor; The channel descriptor is input into a two-layer fully connected network. The first layer uses ReLU activation after dimensionality reduction, and the second layer uses Sigmoid activation after dimensionality restoration to obtain the importance weight of each channel; Applying the importance weight to each channel of the residual feature to perform channel-wise weighting, completing feature recalibration, and obtaining a feature map with attention; The attention-added feature map is sequentially passed through four stages of SE-ResNet, with the number of channels being 64, 128, 256, and 512 respectively. Each stage contains multiple residual blocks with SE modules to obtain a multi-scale feature representation. A global average pooling operation is performed on the multi-scale feature representation to convert the feature map into a 512-dimensional feature vector, capturing the interaction pattern between the machine tool spindle, feed and control system to obtain the machine tool system state representation.

5. The intelligent monitoring method for CNC machine tools based on the Internet of Things according to claim 1 is characterized in that: The fusing of the machine tool component state characteristics and the machine tool system state representation, performing state classification, and obtaining a machine tool health state determination result includes: splicing the machine tool component state feature and the machine tool system state representation in the feature dimension to construct a joint feature vector to obtain a fused feature vector; A feature gating mechanism is introduced into the fused feature vector. When the spindle load fluctuates greatly under high-speed cutting conditions, the weight of the spindle-related features is increased. When the feed system starts and stops frequently, the weight of the feed system features is increased, thereby obtaining a weighted fusion feature that is adaptive to the working conditions. The weighted fusion features are input into a classifier network consisting of three fully connected layers with 256, 128 and 5 nodes respectively to obtain the probability distribution of the machine tool state; Based on the probability distribution of the machine tool state, a multi-level threshold judgment standard is set, and when the vibration spectrum of the machine tool spindle has a peak at the bearing fault characteristic frequency, it is determined that the bearing is abnormal; When the tool wear index exceeds the threshold, it is determined that the tool is abnormal and a preliminary status determination result is obtained; Perform time sliding window smoothing on the preliminary state determination result, confirm the state transition when the state is consistent in 5 consecutive time windows during the processing process, and extend it to 8 windows in the empty cutting state to obtain a stable state determination; Perform similarity matching between the stable state judgment and the historical abnormal pattern library, and when abnormal vibration occurs during heavy cutting processing, give priority to matching the cutting chatter pattern; When an abnormal temperature rise occurs after a long period of operation, the bearing wear pattern is matched first to determine the fault type; Based on the fault type determination, the impact of the fault on the machining quality is evaluated in combination with the current cutting parameters, workpiece material hardness, and tool type to obtain a fault level assessment; Based on the fault level assessment, the machine tool health status is divided into four levels: normal operation, minor abnormality, moderate fault and serious fault. In the case of a serious fault, an emergency stop signal is triggered. In the case of a minor abnormality, only an early warning is issued without intervening in production, thereby obtaining the machine tool health status judgment result.

6. The method for intelligent monitoring of CNC machine tools based on the Internet of Things according to claim 1, characterized in that: Based on the machine tool health status determination result, combined with processing parameters and historical fault data, remaining life prediction and fault warning are performed to obtain a machine tool diagnostic report containing maintenance recommendations, including: Based on the health status determination results of the machine tool, a health index scoring system for key machine tool components is constructed, and health scores are assigned to the spindle, feed, bearing, and tool respectively to obtain the health index of the machine tool components; Extract the current machining process parameters from the machine tool control system, including spindle speed, feed rate, cutting depth, and workpiece material information, establish the working condition-wear relationship, and obtain the working condition load assessment results; Retrieve historical cases similar to the current fault type from the historical fault database, extract the fault development process, maintenance records and component life data, and obtain a reference life data set; fusing the health index of the machine tool component, the working load assessment result, and the reference life data set to calculate the degradation rate and trend of the component and obtain a predicted value of the remaining life of the component; A multi-level warning threshold is set according to the predicted value of the remaining life of the component, and when the predicted life is lower than the threshold, a warning signal of the corresponding level is triggered to obtain a fault warning level; Based on the fault warning level, combined with the machine tool maintenance manual and historical maintenance experience, specific suggestions for component replacement, maintenance or parameter adjustment are generated to obtain a maintenance suggestion list; Based on the maintenance suggestion list, the maintenance tasks are sorted according to the urgency, and the time and cost required for maintenance are estimated to obtain a maintenance plan; The predicted remaining life value of the components, fault warning level, maintenance recommendation list and maintenance planning scheme are integrated into a structured report, which is presented in graphic and text form through edge nodes and cloud platforms to obtain a machine tool diagnostic report containing maintenance recommendations.

7. An intelligent monitoring system for CNC machine tools based on the Internet of Things, characterized in that: For implementing the method for intelligent monitoring of CNC machine tools based on the Internet of Things according to any one of claims 1 to 6, the intelligent monitoring system for CNC machine tools based on the Internet of Things comprises: The transmission module is used to collect CNC machine tool spindle vibration, motor temperature, processing noise, drive current and worktable displacement data, and transmit them synchronously to the edge node through the industrial bus to obtain the original machine tool operation data; a processing module, configured to perform signal processing on the raw data of the machine tool operation to obtain a machine tool operating condition feature set; The analysis module is used to input the machine tool working condition feature set into the recursive convolution hybrid memory network, construct a co-dependency graph of spindle speed change, tool wear, and bearing status, realize state parameter prediction through joint analysis in the time and frequency domains, and obtain the state characteristics of machine tool components, including: inputting the machine tool working condition feature set into the temporal coding layer of the recursive convolution hybrid memory network, and obtaining a feature sequence through position coding and time step embedding processing; inputting the feature sequence into the multi-scale convolution layer of the recursive convolution hybrid memory network, extracting local feature patterns of different time windows through parallel 1×1, 3×3, and 5×5 convolution kernels, performing channel attention weighting on the convolution output features, and obtaining multi-scale temporal convolution features; inputting the multi-scale temporal convolution features into the long-short-term memory layer, using forward and backward bidirectional gated recurrent units to capture long-term dependencies, and controlling the flow of information through forgetting gates, input gates, and output gates to obtain long-term context features; and The feature performs attention fusion layer processing, calculates the pairwise interaction intensity matrix between the three-dimensional features of spindle speed change, tool wear, and bearing status, applies non-negative matrix decomposition to the interaction intensity matrix, extracts potential correlation factors, and obtains a compressed feature interaction representation; constructs a weighted directed graph structure based on the compressed feature interaction representation, and the nodes are set as spindle speed change features, tool wear features, and bearing status features. The edge weights represent the interaction intensity between features, and the node feature information is aggregated through graph convolution operations to obtain a co-dependency graph; performs continuous wavelet transform on the co-dependency graph, maps the time domain signal to the time-frequency plane, calculates the energy distribution and phase relationship of each frequency band, extracts the characteristic frequency components through the maximum energy tracking algorithm, and obtains the time-frequency feature spectrum; the time-frequency feature spectrum and the long time series context feature are fused through a gated update mechanism, and adaptive weights are assigned according to the energy distribution of different frequency bands. The current observation and historical status information are integrated to obtain the state characteristics of the machine tool components; A conversion module is used to convert the state features of the machine tool components into a two-dimensional feature map, extract the interaction features between the spindle, feed, and control systems through SE-ResNet, and obtain a machine tool system state representation; An execution module, configured to integrate the machine tool component state characteristics with the machine tool system state representation, perform state classification, and obtain a machine tool health state determination result; The prediction module is used to perform remaining life prediction and fault warning based on the machine tool health status judgment result, combined with processing parameters and historical fault data, to obtain a machine tool diagnosis report including maintenance suggestions.

8. An intelligent monitoring device for CNC machine tools based on the Internet of Things, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the method for intelligent monitoring of CNC machine tools based on the Internet of Things according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor executes the intelligent monitoring method for CNC machine tools based on the Internet of Things according to any one of claims 1 to 6.

Images