A machine tool running state prediction method based on edge computing

Abstract

The application discloses a kind of based on edge computing's machine tool running state prediction method, comprising: step one: the multi-source state parameter data of machine tool operation is collected;Step two: the multi-source state parameter data is preprocessed in edge node;Step three: machine tool running state is modeled and predicted by improved TSMixer network, and machine tool running state prediction value and time series response score chart are output;Step four: machine tool running state prediction value is compared with running state threshold, and state evaluation label and early warning grade label are generated, and disturbance condition is analyzed based on time series response score chart;Step five: according to early warning grade label, response control strategy is executed, and running state feedback data is collected;Step six: the prediction result and early warning information are visually displayed.The application improves machine tool anomaly identification accuracy and the real-time performance of running state prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and industrial predictive analysis, and particularly relates to a machine tool running state prediction method based on edge computing. BACKGROUND

[0002] With the wide deployment of intelligent manufacturing systems and the continuous improvement of the intelligent level of industrial equipment operation, real-time perception, accurate modeling and predictive control of the running state of core processing equipment such as machine tools have become one of the key links in intelligent factories. The existing machine tool state monitoring system mainly relies on large neural network models deployed on the cloud or local logical rules to make state judgments on the collected running parameters, but there are generally the following problems in the actual industrial environment:

[0003] During the running of the machine tool, a large amount of high-frequency mutation and complex disturbance will be generated, such as tool impact, thermal expansion vibration, current mutation, etc. These non-stationary dynamic characteristics have burstiness and time-varying nature, which are difficult to be accurately identified by traditional static modeling or low-frequency filtering methods, resulting in a decrease in prediction accuracy and insufficient sensitivity of the model when dealing with high dynamic conditions; the data collected by the multi-source sensors deployed in the industrial field have problems such as inconsistent sampling frequencies, channel redundancy and severe signal fluctuations, and traditional time alignment and normalization methods cannot balance accuracy and stability; the existing prediction model based on Transformer or convolutional neural network has a complex structure and large computational overhead, and running on an edge computing node has a resource bottleneck, which is difficult to meet the dual requirements of prediction real-time and deployment flexibility; most model outputs are only single numerical results, lacking explainable feedback on state change trends and mutation points, making it difficult to assist field maintenance personnel in fault tracing and risk intervention.

[0004] Therefore, how to provide a machine tool running state prediction method based on edge computing is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a machine tool running state prediction method based on edge computing, which fully integrates edge computing technology, an improved TSMixer network and an industrial state discrimination method, and describes in detail the real-time modeling of the machine tool running state on the edge node, the identification of abnormal fluctuations and the output of explainable prediction results, which has the advantages of flexible deployment, efficient computation, sensitive mutation response and high visualization of prediction results.

[0006] According to the machine tool running state prediction method based on edge computing according to an embodiment of the present application, the following steps are included:

[0007] Step 1: Collecting multi-source state parameter data during the running of the machine tool;

[0008] Step 2: Preprocess the multi-source state parameter data in the edge nodes to generate a standardized time series dataset;

[0009] Step 3: Input the standardized time series dataset into the improved TSMixer network deployed on the edge nodes to model and predict the machine tool operating status, and output the predicted machine tool operating status value and time series response score map;

[0010] Step 4: Compare the predicted machine tool operating status with the operating status threshold to determine the machine tool operating status, generate status assessment labels and warning level labels, and analyze the disturbance situation based on the time-series response scoring chart;

[0011] Step 5: Based on the warning level label, execute the response control strategy, collect operational status feedback data, and perform incremental updates and model parameter optimization on the improved TSMixer network;

[0012] Step Six: In the visualization interface of the edge node, the predicted value of machine tool operation status, status assessment label, early warning level label, time sequence response score chart and response control strategy are visualized.

[0013] Optionally, the acquisition of multi-source status parameter data during machine tool operation specifically includes:

[0014] Industrial sensors are installed on the machine tool spindle assembly, lead screw pair, tool post, coolant pump and power distribution control module to collect multi-source status parameter data during machine tool operation;

[0015] The multi-source state parameter data includes spindle axial and radial vibration signals, spindle current signals, spindle housing surface temperature signals, lead screw servo drive current signals, guide rail friction temperature rise signals, tool post impact vibration signals, cutting heat temperature signals, coolant temperature signals, hydraulic pump operating current signals, power distribution module power supply voltage signals, power grid fluctuation signals, and power factor change signals.

[0016] Based on industrial communication protocols, multi-source state parameter data is transmitted to edge computing nodes, and cached and timestamped.

[0017] The communication protocol includes Modbus, CAN, EtherCAT, or OPC-UA.

[0018] Optionally, the step of preprocessing the multi-source state parameter data at the edge nodes to generate a standardized time-series dataset specifically involves:

[0019] Multi-source state parameter data are classified according to industrial sensor channels, and pulse anomalies in each channel are removed by using the sliding window mid-value filtering method. Duplicate sampled values ​​in each channel are removed by hash matching method. Signal interference data in each channel is identified and removed based on frequency domain energy threshold method. The data of each channel are time aligned according to the sampling frequency set by each industrial sensor to form an alignment parameter sequence.

[0020] The data of each channel of the alignment parameter sequence is mapped to the [0,1] interval by the maximum and minimum value normalization method to form a normalized parameter sequence;

[0021] The Z-Score standardization method is used to unify the dimensions of the normalized parameter sequence, forming a standardized parameter sequence.

[0022] By setting the time step and sliding window length, the standardized parameter sequence is reorganized into a three-dimensional tensor structure according to the sample batch size, the number of industrial sensor channels, and the time step, forming a standardized time series dataset.

[0023] Optionally, the improved TSMixer network includes a fluctuation-aware gating module, a channel-time bidirectional cross-modeling module, a local anomaly attention module, and a residual decomposition prediction module.

[0024] Optionally, the fluctuation sensing gating module performs fluctuation processing on the standardized time-series dataset by calculating and normalizing the fluctuation intensity score, and outputs a fluctuation time-series tensor, specifically as follows:

[0025] At each time step, a sliding window segment of the standardized time series dataset is extracted based on the sliding window length;

[0026] For each industrial sensor channel, the standard deviation of the sliding window segment is calculated, and the standard deviation of each industrial sensor channel is averaged to generate a fluctuation intensity score.

[0027] The fluctuation intensity scores of all time steps are normalized using the Softmax function to generate a sequence of gated weighted coefficients for the corresponding time steps;

[0028] Based on the gated weighted coefficient sequence, the standardized time series dataset is weighted according to the time step dimension to generate a fluctuating time series tensor.

[0029] Optionally, the channel-time bidirectional cross-modeling module performs bidirectional cross-attention modeling on the fluctuating time series tensor in both the time step dimension and the industrial sensor channel dimension, outputting a joint feature tensor, specifically:

[0030] In the time step dimension, the fluctuation time series tensor is used to extract the time series evolution features of each industrial sensor channel through one-dimensional convolution, and the consistency of the time step dimension is maintained by symmetric zero-padding to generate a time feature tensor.

[0031] The channel dimension and time step dimension of the fluctuating time series tensor of industrial sensors are permuted to form a permuted time series tensor;

[0032] At each time step, the channel features of the permutation time series tensor are extracted, and the channel features are input into the channel hybrid MLP to perform nonlinear mapping and feature compression of the channel dimension, generating the channel feature tensor.

[0033] The channel feature tensor of an industrial sensor is permuted by its channel dimension and time step dimension to generate a permuted channel feature tensor.

[0034] Based on the temporal feature tensor and the permuted channel feature tensor, the temporal-channel attention weight matrix and the channel-temporal attention weight matrix are calculated respectively.

[0035] The time-channel attention weight matrix and the time feature tensor are weighted and modulated to generate a cross-time feature tensor;

[0036] The channel-time attention weight matrix and the permutation channel feature tensor are weighted and modulated to generate the cross-channel feature tensor;

[0037] The cross-temporal feature tensor and the cross-channel feature tensor are fused element-wise with weights to form a cross-fused feature tensor.

[0038] The time feature tensor and the permutation channel feature tensor are fused element-wise with weights to form the residual feature tensor;

[0039] The cross-fusion feature tensor and the residual feature tensor are weighted and combined based on the residual fusion coefficients, and a joint feature tensor is generated through linear transformation and activation function.

[0040] Optionally, the local anomaly attention module performs perturbation enhancement processing on the joint feature tensor through a wavelet transform layer, and outputs an enhanced feature tensor, specifically:

[0041] At each time step, a sliding window segment of the joint feature tensor is extracted based on the sliding window length;

[0042] On each industrial sensor channel, high-frequency detail coefficients are obtained by performing a discrete wavelet transform on each time feature in the sliding window segment, and the mean square value of the high-frequency detail coefficients is calculated as a perturbation score.

[0043] The perturbation scores of each industrial sensor channel are averaged to obtain the average perturbation score at the current time step.

[0044] The average perturbation scores of all time steps are normalized using the Softmax function and then weighted step-by-step with the joint feature tensor to generate an enhanced feature tensor.

[0045] Optionally, the residual decomposition prediction module performs trend modeling and perturbation modeling on the enhanced feature tensor, and outputs the predicted value of the machine tool operating status and the time series response score map, specifically:

[0046] The enhanced feature tensor is used to extract time trend features through three layers of causal convolution to generate a trend feature tensor;

[0047] Perform global average pooling on the trend feature tensor along the time step dimension to obtain the trend prediction sequence of the machine tool's operating status;

[0048] The trend prediction sequence is averaged through channels to output the average trend prediction value of the machine tool's operating status.

[0049] Calculate the residual between the enhancement feature tensor and the trend feature tensor to obtain the perturbation residual tensor;

[0050] The perturbation residual tensor is averaged along the channel dimension of the industrial sensor to obtain the time-series response score map.

[0051] The timing response score map is averaged along the time step dimension to output the average disturbance response value of the machine tool operating status.

[0052] By setting trainable fusion coefficients, the average trend prediction value and the average disturbance response value are fused to generate the machine tool operating status prediction value.

[0053] Optionally, the step of comparing the predicted machine tool operating status with the operating status threshold to determine the machine tool operating status, generating a status assessment label and a warning level label, and analyzing the disturbance based on the time-series response scoring chart specifically involves:

[0054] Set operating status thresholds and abnormal disturbance thresholds, wherein the operating status thresholds include a lower threshold and an upper threshold;

[0055] The predicted machine tool operating status is compared with the operating status threshold, specifically:

[0056] If the predicted value of the machine tool's operating status is lower than the lower limit of the threshold, it is determined to be in an underload state, and a status assessment label of L0 is generated, the warning level label is normal, and the status color is marked as green.

[0057] If the predicted value of the machine tool's operating status is not lower than the lower threshold and not higher than the upper threshold, it is determined to be in a stable state, a status assessment label L1 is generated, the warning level label is "attention", and the status color is marked as blue.

[0058] If the predicted value of the machine tool's operating status is higher than the upper limit of the threshold, it is determined to be in an overload state, and a status assessment label of L2 is generated, a warning level label of high risk is generated, and the status color is marked as red.

[0059] The time-series response scoring graph is mapped to a two-dimensional line graph, where the horizontal axis of the two-dimensional line graph is the time step index and the vertical axis is the disturbance response value.

[0060] If there are two or more time steps where the disturbance response value is higher than the abnormal disturbance threshold, the time steps that meet the condition will be highlighted in the two-dimensional line graph to form a local disturbance abnormality interval.

[0061] Optionally, the step of executing a response control strategy based on the warning level label, collecting operational status feedback data, and performing incremental updates and model parameter optimization on the improved TSMixer network specifically involves:

[0062] If the warning level label is normal, the machine tool operating parameters are kept unchanged, and the current status is recorded as a low-load operating condition benchmark sample.

[0063] If the warning level label is "Attention", the cutting speed, feed rate or coolant flow rate will be adjusted according to the preset parameter adjustment rules, and the operating status feedback data of each industrial sensor channel after adjustment will be collected.

[0064] If the warning level label is high risk, immediately issue a pause command or reduce load control, execute emergency stop protection or cooling enhancement operations, and use the adjustment process and post-execution operation status feedback data as abnormal samples of high-risk operating conditions.

[0065] The collected operational status feedback data is cached and tagged with time, and then input to the edge nodes to perform incremental updates and model parameter optimization on the improved TSMixer network;

[0066] The model parameters include fluctuation-aware gating coefficients, channel-time cross-modeling weight matrix, residual fusion coefficients, trend perturbation fusion coefficients, and activation function bias terms.

[0067] The beneficial effects of this invention are:

[0068] First, this invention deploys industrial sensors at key locations on the machine tool to collect multi-source state parameter data, which is then uploaded to edge computing nodes via an industrial communication protocol. To address the issues of temporal consistency and signal stability of the collected data, this invention performs preprocessing operations on the multi-source state parameter data at the edge nodes, including sliding window midpoint filtering, hash matching for deduplication, and frequency domain energy threshold removal. This completes time alignment, normalization, and standardization steps, constructing a structured, standardized time-series dataset, thereby improving data quality and modeling usability.

[0069] Secondly, the improved TSMixer network proposed in this invention is deployed in edge nodes, enabling efficient modeling and real-time prediction with a lightweight structure. The improved TSMixer network includes a fluctuation-aware gating module, a channel-time bidirectional cross-modeling module, a local anomaly attention module, and a residual decomposition prediction module. The fluctuation-aware gating module generates time-step gating coefficients by calculating the standard deviation of a sliding window, achieving dynamic enhancement of high-fluctuation segments. The channel-time bidirectional cross-modeling module constructs a cross-attention relationship between the time step and channel dimensions through one-dimensional convolution and channel hybrid MLP, improving the collaborative modeling capability between state variables. The local anomaly attention module introduces discrete wavelet transform to perform weighted processing on perturbation enhancement features, improving the model's response strength to anomaly changes. The residual decomposition prediction module divides the prediction process into two sub-tasks: trend modeling and perturbation modeling, generating predicted machine tool operating status values ​​and time-series response score maps, enhancing the interpretability and structural clarity of the prediction results.

[0070] Furthermore, this invention generates status assessment labels and early warning level labels by comparing the predicted machine tool operating status with the operating status threshold. Combined with the time-series response scoring map, it locates high-risk disturbance intervals, achieving refined and visualized status judgment. At the edge nodes, response control strategies are executed based on the early warning level labels, and operating status feedback data is collected to perform incremental updates and model parameter optimization on the improved TSMixer network, enhancing the model's adaptability and continuous iteration capability. The predicted machine tool operating status, label information, and time-series response scoring map are visualized in the edge node's visualization interface, supporting local early warning response and remote cloud synchronization, forming a complete edge intelligent prediction closed loop.

[0071] In summary, this invention significantly improves the modeling capabilities for high-frequency disturbances, non-stationary behavior, and multi-source data correlation features while ensuring the model's lightweight nature and edge deployability. It achieves high-precision, high-responsiveness, and strong interpretability prediction of machine tool operating status, effectively meeting the comprehensive needs of industrial sites for real-time performance, safety, and adaptability. Attached Figure Description

[0072] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0073] Figure 1 This is a schematic diagram of a machine tool operating status prediction method based on edge computing proposed in this invention;

[0074] Figure 2 This is a diagram of the improved TSMixer network structure in a machine tool operating status prediction method based on edge computing proposed in this invention.

[0075] Figure 3 This is a schematic diagram of the channel-time bidirectional cross-modeling module in a machine tool operation status prediction method based on edge computing proposed in this invention;

[0076] Figure 4 This is a schematic diagram of the residual decomposition prediction module in a machine tool operation status prediction method based on edge computing proposed in this invention. Detailed Implementation

[0077] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0078] refer to Figures 1-4 A machine tool operating status prediction method based on edge computing includes the following steps:

[0079] Step 1: Collect multi-source status parameter data during machine tool operation;

[0080] Step 2: Preprocess the multi-source state parameter data in the edge nodes to generate a standardized time series dataset;

[0081] Step 3: Input the standardized time series dataset into the improved TSMixer network deployed on the edge nodes to model and predict the machine tool operating status, and output the predicted machine tool operating status value and time series response score map;

[0082] Step 4: Compare the predicted machine tool operating status with the operating status threshold to determine the machine tool operating status, generate status assessment labels and warning level labels, and analyze the disturbance situation based on the time-series response scoring chart;

[0083] Step 5: Based on the warning level label, execute the response control strategy, collect operational status feedback data, and perform incremental updates and model parameter optimization on the improved TSMixer network;

[0084] Step Six: In the visualization interface of the edge node, the predicted value of machine tool operation status, status assessment label, early warning level label, time sequence response score chart and response control strategy are visualized.

[0085] In this invention, the predicted machine tool operating status is mapped into a line graph according to a time step sequence, and a status assessment label is superimposed to form an operating status trend graph. Warning level labels are displayed in the operating status trend graph using color coding: green indicates normal, blue indicates attention, and red indicates high risk. A two-dimensional line graph generated based on the time-series response scoring graph is visualized. The execution time, adjustment parameters, and operating status feedback data of each response control strategy are recorded and displayed. The visualization interface supports local viewing, parameter sliding query, and historical record switching, assisting users in real-time monitoring of machine tool operating status and potential risk changes.

[0086] In this embodiment, the acquisition of multi-source status parameter data during machine tool operation specifically includes:

[0087] Industrial sensors are installed on the machine tool spindle assembly, lead screw pair, tool post, coolant pump and power distribution control module to collect multi-source status parameter data during machine tool operation;

[0088] The multi-source state parameter data includes spindle axial and radial vibration signals, spindle current signals, spindle housing surface temperature signals, lead screw servo drive current signals, guide rail friction temperature rise signals, tool post impact vibration signals, cutting heat temperature signals, coolant temperature signals, hydraulic pump operating current signals, power distribution module power supply voltage signals, power grid fluctuation signals, and power factor change signals.

[0089] Based on industrial communication protocols, multi-source state parameter data is transmitted to edge computing nodes, and cached and timestamped.

[0090] The communication protocol includes Modbus, CAN, EtherCAT, or OPC-UA.

[0091] In this embodiment, the preprocessing of multi-source state parameter data at edge nodes to generate a standardized time-series dataset specifically involves:

[0092] Multi-source state parameter data are classified according to industrial sensor channels, and pulse anomalies in each channel are removed by using the sliding window mid-value filtering method. Duplicate sampled values ​​in each channel are removed by hash matching method. Signal interference data in each channel is identified and removed based on frequency domain energy threshold method. The data of each channel are time aligned according to the sampling frequency set by each industrial sensor to form an alignment parameter sequence.

[0093] The data of each channel of the alignment parameter sequence is mapped to the [0,1] interval by the maximum and minimum value normalization method to form a normalized parameter sequence;

[0094] The Z-Score standardization method is used to unify the dimensions of the normalized parameter sequence, forming a standardized parameter sequence.

[0095] By setting the time step and sliding window length, the standardized parameter sequence is reorganized into a three-dimensional tensor structure according to the sample batch size, the number of industrial sensor channels, and the time step, forming a standardized time series dataset.

[0096] In this embodiment, the improved TSMixer network includes a fluctuation sensing gating module, a channel-time bidirectional cross-modeling module, a local anomaly attention module, and a residual decomposition prediction module.

[0097] In this embodiment, the fluctuation sensing gating module performs fluctuation processing on the standardized time-series dataset by calculating and normalizing the fluctuation intensity score, and outputs a fluctuation time-series tensor, specifically as follows:

[0098] At each time step, a sliding window segment of the standardized time series dataset is extracted based on the sliding window length;

[0099] For each industrial sensor channel, the standard deviation of the sliding window segment is calculated, and the standard deviation of each industrial sensor channel is averaged to generate a fluctuation intensity score.

[0100] The fluctuation intensity scores of all time steps are normalized using the Softmax function to generate a sequence of gated weighted coefficients for the corresponding time steps;

[0101] Based on the gated weighted coefficient sequence, the standardized time series dataset is weighted according to the time step dimension to generate a fluctuating time series tensor.

[0102] This invention introduces a fluctuation-aware gating module, which calculates fluctuation intensity scores based on the standard deviation of local time-step segments and generates a gating weighted coefficient sequence through Softmax normalization. This enables dynamic weighting of the standardized time-series dataset, thereby highlighting the feature representation of key fluctuation moments. The fluctuation-aware gating module not only enhances the model's ability to perceive non-stationary signals such as sudden anomalies and spike disturbances during machine tool operation, but also effectively improves the robustness and response sensitivity of the overall modeling by suppressing interference from redundant stationary samples through guided attention distribution. This provides the improved TSMixer network with stronger fluctuation identification and feature focusing capabilities.

[0103] In this embodiment, the channel-time bidirectional cross-modeling module performs bidirectional cross-attention modeling on the fluctuating time series tensor in both the time step dimension and the industrial sensor channel dimension, and outputs a joint feature tensor, specifically as follows:

[0104] In the time step dimension, the fluctuation time series tensor is used to extract the time series evolution features of each industrial sensor channel through one-dimensional convolution, and the consistency of the time step dimension is maintained by symmetric zero-padding to generate a time feature tensor.

[0105] The channel dimension and time step dimension of the fluctuating time series tensor of industrial sensors are permuted to form a permuted time series tensor;

[0106] At each time step, the channel features of the permutation time series tensor are extracted, and the channel features are input into the channel hybrid MLP to perform nonlinear mapping and feature compression of the channel dimension, generating the channel feature tensor.

[0107] The channel feature tensor of an industrial sensor is permuted by its channel dimension and time step dimension to generate a permuted channel feature tensor.

[0108] Based on the temporal feature tensor and the permuted channel feature tensor, the temporal-channel attention weight matrix and the channel-temporal attention weight matrix are calculated respectively.

[0109] The time-channel attention weight matrix and the time feature tensor are weighted and modulated to generate a cross-time feature tensor;

[0110] The channel-time attention weight matrix and the permutation channel feature tensor are weighted and modulated to generate the cross-channel feature tensor;

[0111] The cross-temporal feature tensor and the cross-channel feature tensor are fused element-wise with weights to form a cross-fused feature tensor.

[0112] The time feature tensor and the permutation channel feature tensor are fused element-wise with weights to form the residual feature tensor;

[0113] The cross-fusion feature tensor and the residual feature tensor are weighted and combined based on the residual fusion coefficients, and a joint feature tensor is generated through linear transformation and activation function. :

[0114] ;

[0115] in, For the cross-time feature tensor, For the cross-channel feature tensor, For time feature tensors, To replace the channel feature tensor, The residual fusion coefficient is... Let be the projection transformation matrix. For bias terms, This is the activation function.

[0116] In this embodiment, the local anomaly attention module performs perturbation enhancement processing on the joint feature tensor through a wavelet transform layer, and outputs an enhanced feature tensor, specifically as follows:

[0117] At each time step, a sliding window segment of the joint feature tensor is extracted based on the sliding window length;

[0118] On each industrial sensor channel, high-frequency detail coefficients are obtained by performing a discrete wavelet transform on each time feature in the sliding window segment, and the mean square value of the high-frequency detail coefficients is calculated as a perturbation score.

[0119] The perturbation scores of each industrial sensor channel are averaged to obtain the average perturbation score at the current time step.

[0120] The average perturbation scores of all time steps are normalized using the Softmax function and then weighted step-by-step with the joint feature tensor to generate an enhanced feature tensor.

[0121] In this invention, the local anomaly attention module significantly enhances the model's ability to identify and represent high-frequency anomalous disturbance features. Feature fragments are extracted at each time step using a sliding window, and discrete wavelet transform is performed on the temporal features within each industrial sensor channel to obtain high-frequency detail coefficients. Their mean square values ​​are then calculated as disturbance scores, effectively extracting high-frequency information from short-term sudden disturbances. The disturbance scores from all channels are averaged and normalized to generate a time-weighted coefficient sequence with disturbance sensing capabilities. This achieves time-step enhancement processing of the joint feature tensor, enabling the model to focus more on weak but critical disturbance signals during state prediction.

[0122] Compared to the fluctuation perception gating module, both modules employ a sliding window approach and extract local segments step-by-step in their structural design, but their focuses differ. The fluctuation perception gating module primarily extracts abrupt change amplitude information through standard deviation, focusing more on global fluctuation trends and amplitude changes, and exhibits stronger perception of continuous changes. In contrast, the local anomaly attention module introduces wavelet transform to focus on high-frequency detail signals, enhancing the characterization of high-frequency anomalies such as sharp disturbances and short-term anomalies, thus compensating for the fluctuation gating module's blind spot in short-term high-frequency responses. The synergistic effect of both modules improves the model's ability to perceive multi-scale dynamic disturbances, ensures contextual robustness and anomaly interpretability during feature processing, and significantly enhances the modeling accuracy and diagnostic performance of the entire prediction model for complex machine tool operating conditions.

[0123] In this embodiment, the residual decomposition prediction module performs trend modeling and perturbation modeling on the enhanced feature tensor, and outputs the predicted value of the machine tool operating status and the time series response score map, specifically as follows:

[0124] The enhanced feature tensor is used to extract time trend features through three layers of causal convolution to generate a trend feature tensor;

[0125] Perform global average pooling on the trend feature tensor along the time step dimension to obtain the trend prediction sequence of the machine tool's operating status;

[0126] The trend prediction sequence is averaged through channels to output the average trend prediction value of the machine tool's operating status.

[0127] Calculate the residual between the enhancement feature tensor and the trend feature tensor to obtain the perturbation residual tensor;

[0128] The perturbation residual tensor is averaged along the channel dimension of the industrial sensor to obtain the time-series response score map.

[0129] The timing response score map is averaged along the time step dimension to output the average disturbance response value of the machine tool operating status.

[0130] By setting trainable fusion coefficients, the average trend prediction value and the average disturbance response value are fused to generate the machine tool operating status prediction value.

[0131] In this embodiment, the step of comparing the predicted machine tool operating status with the operating status threshold to determine the machine tool operating status, generating a status assessment label and a warning level label, and analyzing the disturbance based on the time-series response scoring chart specifically involves:

[0132] Set operating status thresholds and abnormal disturbance thresholds, wherein the operating status thresholds include a lower threshold and an upper threshold;

[0133] The predicted machine tool operating status is compared with the operating status threshold, specifically:

[0134] If the predicted value of the machine tool's operating status is lower than the lower limit of the threshold, it is determined to be in an underload state, and a status assessment label of L0 is generated, the warning level label is normal, and the status color is marked as green.

[0135] If the predicted value of the machine tool's operating status is not lower than the lower threshold and not higher than the upper threshold, it is determined to be in a stable state, a status assessment label L1 is generated, the warning level label is "attention", and the status color is marked as blue.

[0136] If the predicted value of the machine tool's operating status is higher than the upper limit of the threshold, it is determined to be in an overload state, and a status assessment label of L2 is generated, a warning level label of high risk is generated, and the status color is marked as red.

[0137] The time-series response scoring graph is mapped to a two-dimensional line graph, where the horizontal axis of the two-dimensional line graph is the time step index and the vertical axis is the disturbance response value.

[0138] If there is a time step where the disturbance response value is higher than the abnormal disturbance threshold, then the time step that meets the condition will be highlighted in the two-dimensional line graph to form a local disturbance abnormality interval.

[0139] For example, during the continuous operation of a certain intelligent machining center, the predicted value of the machine tool's operating status measured at the 28th second is 0.294, which is lower than the set lower threshold of 0.35, and can be judged as an underload state. The corresponding status assessment label is L0, the warning level label is "normal", and it is represented in green in the visualization interface. At the 52nd second, the predicted value is 0.612, which is within the set threshold range [0.35, 0.75], and can be judged as a stable state. The corresponding status assessment label is L1, the warning level label is "attention", and it is marked in blue in the visualization interface. At the 75th second, the predicted value reaches 0.834, which exceeds the set upper threshold of 0.75, and can be judged as an overload state. The corresponding status assessment label is L2, the warning level label is "high risk", and it is displayed in red. Furthermore, by mapping the time-series response score map to a two-dimensional line graph, the disturbance response values ​​at time steps 73, 75, and 77 are 0.91, 0.89, and 0.94, respectively, all of which are higher than the abnormal disturbance threshold of 0.85. These three time steps are highlighted and identified as local disturbance abnormal intervals, which can provide accurate basis for state intervention and fault prediction.

[0140] In this embodiment, the step of executing a response control strategy based on the warning level label, collecting operational status feedback data, and performing incremental updates and model parameter optimization on the improved TSMixer network specifically involves:

[0141] If the warning level label is normal, the machine tool operating parameters are kept unchanged, and the current status is recorded as a low-load operating condition benchmark sample.

[0142] If the warning level label is "Attention", the cutting speed, feed rate or coolant flow rate will be adjusted according to the preset parameter adjustment rules, and the operating status feedback data of each industrial sensor channel after adjustment will be collected.

[0143] If the warning level label is high risk, immediately issue a pause command or reduce load control, execute emergency stop protection or cooling enhancement operations, and use the adjustment process and post-execution operation status feedback data as abnormal samples of high-risk operating conditions.

[0144] The collected operational status feedback data is cached and tagged with time, and then input to the edge nodes to perform incremental updates and model parameter optimization on the improved TSMixer network;

[0145] The model parameters include fluctuation-aware gating coefficients, channel-time cross-modeling weight matrix, residual fusion coefficients, trend perturbation fusion coefficients, and activation function bias terms.

[0146] Example 1:

[0147] To verify the feasibility of this invention in practice, it was applied to an edge deployment prediction system in a CNC machine tool processing workshop. This workshop is equipped with four vertical machining centers, which are used year-round for high-intensity batch cutting of parts. During machine tool operation, there are significant non-stationary factors such as thermal expansion disturbances, tool impact vibrations, sudden spindle current changes, and coolant pressure fluctuations. These disturbances often interfere with prediction accuracy and early warning timeliness, thus affecting operational stability and maintenance efficiency.

[0148] In actual deployment, technicians installed industrial sensors for vibration, temperature, current, and voltage at the spindle, leadscrew, coolant pump, tool post, and power distribution module. The sampling frequency was uniformly set to 500Hz. The collected data was transmitted in real-time to the edge computing node via a hybrid Modbus and OPC-UA protocol. The edge node hardware consisted of an embedded industrial control host with a GPU acceleration module, running the method proposed in this invention.

[0149] Preprocessing of multi-source state parameter data is performed at the edge, including outlier removal, duplicate sample removal, time alignment, and standardization, to generate a standardized time-series dataset. This standardized time-series dataset is then input into an improved TSMixer network deployed on edge nodes for modeling and prediction. The network architecture integrates a fluctuation-aware gating module, a channel-time bidirectional cross-modeling module, a local anomaly attention module, and a residual decomposition prediction module. It is sensitive to abrupt changes and outputs interpretable disturbance score maps and predicted machine tool operating status values.

[0150] To comprehensively evaluate the performance advantages of this invention, three other commonly used industrial prediction models were selected for comparison, including the original TSMixer model, the lightweight GRU model, and the traditional LSTM model. The prediction accuracy, response latency, and disturbance localization capability were tested in the same dataset and environment. The experimental results are shown in Table 1.

[0151] Table 1. Performance Comparison of Different Prediction Methods in Industrial Machine Tool Condition Prediction Tasks

[0152]

[0153] As shown in Table 1, the method of this invention outperforms the comparative models in multiple performance indicators. Regarding state prediction accuracy, the method of this invention achieves a state prediction MAE of 0.127 and a state prediction RMSE of 0.201, representing reductions of 30.6% and 24.2% compared to the TSMixer model, 47.7% and 38.7% compared to the GRU model, and 52.6% and 42.7% compared to the LSTM model, significantly improving prediction accuracy. In terms of anomaly warning capability, the method of this invention can issue warnings an average of 9.7 seconds in advance, representing improvements of 86.5%, 185.3%, and 234.5% compared to TSMixer, GRU, and LSTM, respectively, enhancing the sensitivity to high-risk states.

[0154] In terms of disturbance identification, the method of this invention achieves a disturbance localization accuracy of 91.3%, far exceeding that of the comparison models, demonstrating stronger capabilities in mutation modeling and temporal anomaly identification. Regarding model interpretability, the method of this invention achieves an interpretability score of 4.85, representing a 66.7% improvement compared to TSMixer, and improvements of 160.8% and 180.3% compared to GRU and LSTM, respectively, facilitating on-site operators' understanding of the prediction basis and warning reasons. In terms of execution efficiency, the method of this invention has an average prediction time of 82ms at edge nodes, slightly higher than TSMixer and GRU models, but a 24.8% reduction compared to the LSTM model, indicating that the method of this invention can balance accuracy and interpretability while meeting real-time requirements.

[0155] The method of this invention demonstrates significant advantages in terms of accuracy, responsiveness, interpretability, and edge deployment adaptability, and has good engineering implementation value and potential for widespread application.

[0156] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine tool operating status prediction method based on edge computing, characterized in that, Includes the following steps: Step 1: Collect multi-source status parameter data during machine tool operation; Step 2: Preprocess the multi-source state parameter data at the edge nodes to generate a standardized time-series dataset, specifically as follows: Multi-source state parameter data are classified according to industrial sensor channels, and pulse anomalies in each channel are removed by using the sliding window mid-value filtering method. Duplicate sampled values ​​in each channel are removed by hash matching method. Signal interference data in each channel is identified and removed based on frequency domain energy threshold method. The data of each channel are time aligned according to the sampling frequency set by each industrial sensor to form an alignment parameter sequence. The data of each channel of the alignment parameter sequence is mapped to the [0,1] interval by the maximum and minimum value normalization method to form a normalized parameter sequence; The Z-Score standardization method is used to unify the dimensions of the normalized parameter sequence, forming a standardized parameter sequence. By setting the time step and sliding window length, the standardized parameter sequence is reorganized into a three-dimensional tensor structure according to the sample batch size, the number of industrial sensor channels, and the time step, forming a standardized time series dataset. Step 3: Input the standardized time series dataset into the improved TSMixer network deployed on the edge nodes to model and predict the machine tool operating status, and output the predicted machine tool operating status value and time series response score map; the improved TSMixer network includes a fluctuation sensing gating module, a channel-time bidirectional cross modeling module, a local anomaly attention module, and a residual decomposition prediction module; The fluctuation sensing gating module calculates and normalizes fluctuation intensity scores to perform fluctuation processing on the standardized time-series dataset, outputting a fluctuation time-series tensor, specifically: At each time step, a sliding window segment of the standardized time series dataset is extracted based on the sliding window length; For each industrial sensor channel, the standard deviation of the sliding window segment is calculated, and the standard deviation of each industrial sensor channel is averaged to generate a fluctuation intensity score. The fluctuation intensity scores of all time steps are normalized using the Softmax function to generate a sequence of gated weighted coefficients for the corresponding time steps; Based on the gated weighted coefficient sequence, the standardized time series dataset is weighted according to the time step dimension to generate a fluctuating time series tensor; The channel-time bidirectional cross-modeling module performs bidirectional cross-attention modeling on the fluctuating time series tensor in both the time step dimension and the industrial sensor channel dimension, outputting a joint feature tensor, specifically: In the time step dimension, the fluctuation time series tensor is used to extract the time series evolution features of each industrial sensor channel through one-dimensional convolution, and the consistency of the time step dimension is maintained by symmetric zero-padding to generate a time feature tensor. The channel dimension and time step dimension of the fluctuating time series tensor of industrial sensors are permuted to form a permuted time series tensor; At each time step, the channel features of the permutation time series tensor are extracted, and the channel features are input into the channel hybrid MLP to perform nonlinear mapping and feature compression of the channel dimension, generating the channel feature tensor. The channel feature tensor of an industrial sensor is permuted by its channel dimension and time step dimension to generate a permuted channel feature tensor. Based on the temporal feature tensor and the permuted channel feature tensor, the temporal-channel attention weight matrix and the channel-temporal attention weight matrix are calculated respectively. The time-channel attention weight matrix and the time feature tensor are weighted and modulated to generate a cross-time feature tensor; The channel-time attention weight matrix and the permutation channel feature tensor are weighted and modulated to generate the cross-channel feature tensor; The cross-temporal feature tensor and the cross-channel feature tensor are fused element-wise with weights to form a cross-fused feature tensor. The time feature tensor and the permutation channel feature tensor are fused element-wise with weights to form the residual feature tensor; The cross-fusion feature tensor and the residual feature tensor are weighted and combined based on the residual fusion coefficients, and a joint feature tensor is generated through linear transformation and activation function. The local anomaly attention module performs perturbation enhancement processing on the joint feature tensor through a wavelet transform layer, and outputs an enhanced feature tensor, specifically: At each time step, a sliding window segment of the joint feature tensor is extracted based on the sliding window length; On each industrial sensor channel, high-frequency detail coefficients are obtained by performing a discrete wavelet transform on each time feature in the sliding window segment, and the mean square value of the high-frequency detail coefficients is calculated as a perturbation score. The perturbation scores of each industrial sensor channel are averaged to obtain the average perturbation score at the current time step. The average perturbation scores of all time steps are normalized using the Softmax function and then weighted step-by-step with the joint feature tensor to generate the enhanced feature tensor. The residual decomposition prediction module performs trend modeling and perturbation modeling on the enhanced feature tensor, and outputs predicted machine tool operating status values ​​and time-series response score maps, specifically: The enhanced feature tensor is used to extract time trend features through three layers of causal convolution to generate a trend feature tensor; Perform global average pooling on the trend feature tensor along the time step dimension to obtain the trend prediction sequence of the machine tool's operating status; The trend prediction sequence is averaged through channels to output the average trend prediction value of the machine tool's operating status. Calculate the residual between the enhancement feature tensor and the trend feature tensor to obtain the perturbation residual tensor; The perturbation residual tensor is averaged along the channel dimension of the industrial sensor to obtain the time-series response score map. The timing response score map is averaged along the time step dimension to output the average disturbance response value of the machine tool operating status. Set trainable fusion coefficients to fuse the average trend prediction value and the average disturbance response value to generate machine tool operating status prediction value; Step 4: Compare the predicted machine tool operating status with the operating status threshold to determine the machine tool operating status, generate status assessment labels and warning level labels, and analyze the disturbance situation based on the time-series response scoring chart; Step 5: Based on the warning level label, execute the response control strategy, collect operational status feedback data, and perform incremental updates and model parameter optimization on the improved TSMixer network; Step Six: In the visualization interface of the edge node, the predicted value of machine tool operation status, status assessment label, early warning level label, time sequence response score chart and response control strategy are visualized.

2. The machine tool operating status prediction method based on edge computing according to claim 1, characterized in that, The multi-source status parameter data collected during the machine tool operation process are specifically as follows: Industrial sensors are installed on the machine tool spindle assembly, lead screw pair, tool post, coolant pump and power distribution control module to collect multi-source status parameter data during machine tool operation; The multi-source state parameter data includes spindle axial and radial vibration signals, spindle current signals, spindle housing surface temperature signals, lead screw servo drive current signals, guide rail friction temperature rise signals, tool post impact vibration signals, cutting heat temperature signals, coolant temperature signals, hydraulic pump operating current signals, power distribution module power supply voltage signals, power grid fluctuation signals, and power factor change signals. Based on industrial communication protocols, multi-source state parameter data is transmitted to edge computing nodes, and cached and timestamped. The communication protocol includes Modbus, CAN, EtherCAT, or OPC-UA.

3. The machine tool operating status prediction method based on edge computing according to claim 1, characterized in that, The process involves comparing the predicted machine tool operating status with the operating status threshold to determine the machine tool operating status, generating a status assessment label and a warning level label, and analyzing the disturbance based on the time-series response scoring chart. Specifically: Set operating status thresholds and abnormal disturbance thresholds, wherein the operating status thresholds include a lower threshold and an upper threshold; The predicted machine tool operating status is compared with the operating status threshold, specifically: If the predicted value of the machine tool's operating status is lower than the lower limit of the threshold, it is determined to be in an underload state, and a status assessment label of L0 is generated, the warning level label is normal, and the status color is marked as green. If the predicted value of the machine tool's operating status is not lower than the lower threshold and not higher than the upper threshold, it is determined to be in a stable state, a status assessment label L1 is generated, the warning level label is "attention", and the status color is marked as blue. If the predicted value of the machine tool's operating status is higher than the upper limit of the threshold, it is determined to be in an overload state, and a status assessment label of L2 is generated, a warning level label of high risk is generated, and the status color is marked as red. The time-series response scoring graph is mapped to a two-dimensional line graph, where the horizontal axis of the two-dimensional line graph is the time step index and the vertical axis is the disturbance response value. If there are two or more time steps where the disturbance response value is higher than the abnormal disturbance threshold, the time steps that meet the condition will be highlighted in the two-dimensional line graph to form a local disturbance abnormality interval.

4. The machine tool operating status prediction method based on edge computing according to claim 1, characterized in that, The process involves executing a response control strategy based on the warning level label, collecting operational status feedback data, and performing incremental updates and model parameter optimization on the improved TSMixer network. Specifically: If the warning level label is normal, the machine tool operating parameters are kept unchanged, and the current status is recorded as a low-load operating condition benchmark sample. If the warning level label is "Attention", the cutting speed, feed rate or coolant flow rate will be adjusted according to the preset parameter adjustment rules, and the operating status feedback data of each industrial sensor channel after adjustment will be collected. If the warning level label is high risk, immediately issue a pause command or reduce load control, execute emergency stop protection or cooling enhancement operations, and use the adjustment process and post-execution operation status feedback data as abnormal samples of high-risk operating conditions. The collected operational status feedback data is cached and tagged with time, and then input to the edge nodes to perform incremental updates and model parameter optimization on the improved TSMixer network; The model parameters include fluctuation-aware gating coefficients, channel-time cross-modeling weight matrix, residual fusion coefficients, trend perturbation fusion coefficients, and activation function bias terms.

Images