A time series prediction and decision-making method and system for multi-source data of CNC machine tools

By aligning multi-source data, identifying operating conditions, and constructing adaptive time windows, combined with a multi-scale prediction model, the problem of prediction stability and decision-making closure for CNC machine tools under complex operating conditions is solved, achieving high-precision prediction and decision-making integration, and improving production efficiency and quality.

CN123115701APending Publication Date: 2026-11-17QILU INST OF TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611610227.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing multi-source data prediction methods for CNC machine tools lack semantic information about operating conditions, making it difficult to maintain prediction stability when switching between complex operating conditions. Furthermore, they struggle to take into account multi-scale time-series information and lack a closed-loop mechanism between prediction results and actual decision-making.

Method used

By collecting and aligning multi-source heterogeneous data, a multi-source time series matrix under a unified time scale is constructed to identify processing conditions and extract differential features. An adaptive time window is built, and multi-scale prediction models of TCN, LSTM, and Transformer are combined to generate operation decision instructions.

Benefits of technology

It improves the prediction accuracy of CNC machine tool operating status and machining quality, realizes seamless connection from status perception to intelligent decision-making, reduces equipment failure rate, and improves production efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115701A_ABST
    Figure CN123115701A_ABST
Patent Text Reader

Abstract

This application discloses a time series prediction and decision-making method and system for multi-source data of CNC machine tools, relating to the field of data processing technology. The method includes: collecting multi-source heterogeneous data from CNC machine tools and performing time alignment and resampling to construct a multi-source time series matrix at a unified time scale; identifying machining conditions based on the multi-source time series matrix to obtain a condition label sequence; constructing differentiated time series features based on the condition label sequence to form a feature vector containing semantic information about the machining conditions; constructing an adaptive time window based on the feature vector, according to data volatility, spindle load rate, and vibration energy change rate; using the feature vector within the adaptive time window as input to construct a multi-scale prediction model and outputting the prediction results; and generating operation decision instructions based on the prediction results. This application improves the accuracy of CNC machine tool operation status prediction and the timeliness of decision-making.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a time series prediction and decision-making method and system for multi-source data of CNC machine tools. Background Technology

[0002] CNC machine tools are core equipment for discrete manufacturing and precision machining, and their operating status directly affects machining quality, equipment lifespan, and production efficiency. During machining, machine tools operate under high-speed, heavy-load, and complex conditions for extended periods, making them susceptible to factors such as tool wear, thermal deformation, and equipment aging. If abnormal conditions are not detected in time, they can lead to decreased machining accuracy, equipment failure, and production interruptions. Therefore, it is necessary to monitor the machine tool's operating status in real time and to make early warnings and decisions based on its changing trends to ensure stable machine tool operation.

[0003] With the development of the Industrial Internet and intelligent manufacturing, CNC machine tools are generally equipped with various sensors and monitoring modules, which can continuously collect data on spindle vibration, current power, temperature, and various process parameters and machining quality. This information accumulates continuously in the form of time series, forming typical multi-source industrial big data, providing a rich data foundation for the prediction and decision-making of machine tool operating status. Regarding machine tool status prediction, related methods have evolved from traditional statistical methods such as trend analysis, regression analysis, exponential smoothing, and grey prediction, to machine learning methods such as support vector machines, decision trees, and random forests, and then to deep learning methods such as RNNs, LSTMs, CNNs, and Transformers. Traditional methods are limited by the assumption of data stationarity, making it difficult to characterize complex nonlinear relationships; while machine learning methods have certain nonlinear modeling capabilities, they are heavily reliant on feature engineering and struggle to fully mine the deep information in multi-source time series data. In recent years, deep learning methods have demonstrated stronger capabilities in time-series dependency modeling. However, existing deep learning-based time series prediction methods for CNC machine tools still suffer from several problems: First, existing methods often directly model the raw, multi-source data, lacking characterization of data distribution differences under different machining conditions and failing to introduce semantic information about the working conditions before modeling, resulting in insufficient prediction stability when the model switches between complex working conditions. Second, most deep learning models use a single structure for time series modeling, making it difficult to simultaneously incorporate high-frequency local features such as kinetic energy, medium-term trend features such as temperature and current, and long-term dependency features such as equipment wear, leading to insufficient utilization of multi-scale time series information. Furthermore, existing research typically only focuses on time series prediction, lacking a mechanism to integrate prediction results with actual decisions such as machine tool operation and maintenance, and machining parameter adjustments, making it difficult to form a closed-loop decision-making system oriented towards production applications. Summary of the Invention

[0004] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: In a first aspect, embodiments of this application provide a time series prediction and decision-making method for multi-source data of CNC machine tools, including: Collect multi-source heterogeneous data from CNC machine tools, perform time alignment and resampling, and construct a multi-source time series matrix under a unified time scale; Based on the multi-source time series matrix, processing conditions are identified to obtain a processing condition label sequence; Based on the operating condition label sequence, differentiated time series features are constructed to form a feature vector containing operating condition semantic information; Based on the feature vector, an adaptive time window is constructed according to the data volatility, spindle load rate and vibration energy change rate. Using the feature vectors within the adaptive time window as input, a multi-scale prediction model is constructed, and the prediction results are output. Based on the prediction results, operational decision instructions are generated.

[0005] In one possible implementation, the step of acquiring multi-source heterogeneous data from CNC machine tools, performing time alignment and resampling, and constructing a multi-source time series matrix at a unified time scale includes: Multi-source heterogeneous data is collected during the machine tool operation process. The multi-source heterogeneous data includes spindle vibration signal, spindle current and power signal, temperature sensor data, process parameters, machine tool operating status signal and machining quality detection data. Set a uniform time interval, establish a standard time axis, and obtain the standard time axis sequence on the standard time axis; Based on the standard time axis sequence, windowed feature extraction is performed on the main shaft vibration signal to obtain a low-dimensional vibration feature vector. For the spindle current and power signals, temperature sensor data, and machining quality inspection data, using the standard time axis sequence as a reference, missing values ​​are filled in using linear interpolation, with the following formula: in, The current time to be interpolated on the standard time axis. The current time to be interpolated Previously, and with Adjacent sampled times, The current time to be interpolated Afterwards, and with Adjacent sampled times, These are the spindle current signals at time [time]. and The sampled values, Temperature sensor data at time 1 and The sampled values, Temperature sensor data at time 1 and The sampled values, , and Do not use the spindle current, temperature sensor data, and machining quality inspection data obtained through linear interpolation at time... Data values; For the process parameters and the machine tool operating status signals, using the standard time axis sequence as a reference, missing values ​​are filled in using a forward hold method, as shown in the formula: in, For process parameters, This is a machine tool operating status signal. For process parameters at time The sampled values, For machine tool operating status signals at time The sampled values; After aligning various features by time, they are concatenated to form a multi-source time series matrix with a unified time scale.

[0006] In one possible implementation, based on the standard timeline sequence, windowed feature extraction is performed on the principal shaft vibration signal to obtain a low-dimensional vibration feature vector, including: Within the time interval window corresponding to each moment on the standard time axis, the original high-frequency vibration sampling sequence of the spindle vibration signal is extracted; The statistical features of the original high-frequency sampling sequence are extracted using the following formula: in, for The mean value of the vibration signal within the time window. This represents the number of sampling points within the window. for Standard deviation of vibration signal within time window for Peak value of vibration signal within the time window For the frequency domain sequence, the first... Each frequency component for The total frequency domain energy of the vibration signal within the time window. for Within the time window Each vibration signal sample value, for The first vibration signal sample value within the time window, for The second vibration signal sample value within the time window. for Within the time window One vibration signal sample value; The mean, standard deviation, peak value, and total frequency domain energy of the vibration signal are concatenated to obtain the low-dimensional vibration feature vector at the current moment.

[0007] In one possible implementation, processing condition identification is performed based on the multi-source time series matrix to obtain a processing condition label sequence, including: Obtain a multi-source time series matrix, and extract the working condition identification basis parameters from the multi-source time series matrix; Set the cutting current threshold, current fluctuation judgment threshold, and vibration sudden change judgment threshold; Iterate through each moment in the multi-source time series matrix, determine the working condition based on the working condition identification parameters and the set threshold, and generate a working condition label for each moment; When the spindle speed is zero, it is determined to be an unloaded working condition; When the spindle current exceeds the cutting current threshold and the vibration fluctuation is stable, it is determined to be a cutting condition; When the change in current exceeds the current fluctuation judgment threshold and the change in vibration amplitude exceeds the vibration sudden change judgment threshold, it is judged as an abnormal fluctuation condition. When the PLC status signal is at the preset value, it is determined to be a tool changing condition; When the spindle current exceeds the cutting current threshold, and the current fluctuates slightly periodically without sudden changes in vibration amplitude, it is determined to be a normal fluctuating machining condition. When the spindle current exceeds the cutting current threshold and the vibration continues to fluctuate at a high frequency without a sudden change in amplitude, it is determined to be a high-frequency vibration machining condition. By combining the operating condition labels at each time point in chronological order, a sequence of operating condition labels is obtained.

[0008] In one possible implementation, differentiated time-series features are constructed based on the operating condition label sequence to form a feature vector containing operating condition semantic information, including: Obtain the operating condition label sequence and the corresponding multi-source time series matrix; Set local time series data at the current moment, match the corresponding working condition type according to the working condition label at the current moment, and construct differentiated time series features according to the working condition type; When the operating condition label represents a stable operating condition, extract the steady-state feature subset; When the operating condition label indicates a fluctuating operating condition, extract the fluctuating feature subset; When the operating condition label represents a high-frequency vibration operating condition, extract the high-frequency vibration feature subset; When the operating condition label indicates an abnormal operating condition, extract the abnormal feature subset; The corresponding feature subset is matched based on the working condition label, and the matched feature subset is combined with the corresponding working condition semantic information to obtain a feature vector containing the working condition semantic information.

[0009] In one possible implementation, an adaptive time window is constructed based on the feature vector, according to the data volatility, spindle load rate, and vibration energy change rate, including: Obtain feature vectors containing semantic information about operating conditions; The vibration fluctuation rate, current change rate, and vibration energy change rate of the adaptive time window are calculated based on the feature vector. The length of the adaptive time window is dynamically adjusted based on the vibration fluctuation rate, current change rate, and vibration energy change rate, using the following formula: in, To adapt to the time window length, This is the proportionality coefficient. To prevent tiny constants with a denominator of zero, For vibrational fluctuation rate, The rate of change of current, The rate of change of vibrational energy.

[0010] In one possible implementation, the multi-scale prediction model includes a TCN layer, an LSTM layer, and a Transformer layer. The outputs of the TCN layer, LSTM layer, and Transformer layer are connected to the input of the feature concatenation layer. The output of the feature concatenation layer is connected to the input of the fully connected layer. The output of the fully connected layer is connected to the prediction output layer.

[0011] In one possible implementation, a multi-scale prediction model is constructed using the feature vectors within the adaptive time window as input, and the prediction results are output, including: Obtain the feature vectors within the adaptive time window and the length of the adaptive time window; At each time step, construct a time series subsequence of adaptive window length; The time series subsequences are constructed according to the prediction targets to form corresponding input feature subsets, including vibration prediction subsequences, temperature and current prediction subsequences, power prediction subsequences, and processing quality prediction subsequences; A parallel multi-scale prediction model combining TCN, LSTM, and Transformer is constructed. The vibration prediction subsequence is input into the TCN layer, and high-frequency local correlation features of vibration are extracted through multi-layer one-dimensional dilated causal convolution to obtain the local feature vector. The formula is as follows: in, for Vibration prediction subsequence within the time window; for The local feature vector output by the layer. For feature extraction of vibration prediction subsequences network; The temperature and current prediction subsequences and the power prediction subsequences are input into an LSTM layer to extract the intermediate trend features, resulting in the intermediate trend feature vector, as shown in the formula: in, for Temperature and current prediction subsequences within the time window; for Power prediction subsequence within the time window; for The layer extracts the mid-term trend feature vector based on temperature and current prediction subsequences. for The layer extracts the mid-term trend feature vector based on the power prediction subsequence; To extract mid-term trend features from the temperature, current, and power prediction subsequences. network; For mid-term feature fusion function, This represents the feature vector of the medium-term trend. The processing quality prediction subsequence is input into the Transformer layer, and long-term dependencies are captured through a self-attention mechanism to obtain the long-term dependency feature vector, as shown in the formula: in, for Processing quality prediction subsequence within the time window; for The long-term dependency feature vector output by the layer; To extract long-term dependency features for processing quality prediction subsequences network; The local feature vector, intermediate trend feature vector, and long-term dependency feature vector are concatenated to form a multi-scale fused feature vector: The multi-scale fused feature vector is input into a fully connected layer for adaptive weighted fusion, and the prediction result is output as follows: in, For multi-scale feature fusion function, for The prediction result at any given time; This is the weight matrix of the fully connected layer; For bias terms of fully connected layers; For the mapping function of the fully connected layer, This is a multi-scale fused feature vector.

[0012] In one possible implementation, generating an operational decision instruction based on the prediction result includes: Obtain the prediction results output by the multi-scale prediction model, and pre-set safety thresholds and trend judgment rules for each prediction indicator; When the prediction result meets the corresponding threshold condition, the corresponding operation decision instruction is generated; When the predicted vibration exceeds the vibration safety threshold or the rate of change is abnormal, a spindle maintenance early warning command is generated. When the predicted temperature exceeds the temperature safety threshold or the heating rate exceeds the heating rate threshold, a cooling or shutdown recommendation is generated. When the predicted power exceeds the power fluctuation threshold or the rate of change exceeds the rate of change threshold, processing parameter adjustment suggestions are generated. When the predicted quality fluctuation exceeds the quality fluctuation threshold, suggestions for tool change or machining parameter optimization are generated.

[0013] Secondly, embodiments of this application provide a time series prediction and decision-making system for multi-source data of CNC machine tools, including: The data acquisition and processing module is used to acquire multi-source heterogeneous data from CNC machine tools and perform time alignment and resampling to construct a multi-source time series matrix under a unified time scale. The working condition identification module is used to identify the processing working conditions based on the multi-source time series matrix and obtain a working condition label sequence. The construction module is used to construct differentiated time series features based on the working condition label sequence, forming a feature vector containing working condition semantic information; The time window module is used to construct an adaptive time window based on the feature vector, according to the data volatility, spindle load rate and vibration energy change rate. The prediction module is used to construct a multi-scale prediction model with the feature vectors within the adaptive time window as input, and output the prediction results. The decision instruction generation module is used to generate execution decision instructions based on the prediction results.

[0014] Compared with the prior art, the beneficial effects of this application are as follows: This application solves the problems of inconsistent scales and poor adaptability of multi-source heterogeneous data of CNC machine tools by coordinating multi-source data time alignment and resampling, working condition identification and adaptive time window construction. It effectively improves the regularity and effectiveness of time series data and provides a high-quality data foundation for subsequent predictive modeling.

[0015] This application constructs a multi-scale prediction model that integrates TCN, LSTM, and Transformer in parallel. It extracts local vibration features, medium-term trend features, and long-term dependency features in a targeted manner. Through adaptive weight learning, it achieves efficient fusion of multi-scale features, which overcomes the limitation of a single model in taking into account the time series features of different scales and improves the prediction accuracy of CNC machine tool operating status and machining quality.

[0016] This application establishes an integrated closed-loop mechanism for prediction and decision-making, which directly transforms multi-scale prediction results into actionable decision commands such as spindle maintenance, parameter adjustment, and shutdown testing. This achieves seamless integration of CNC machine tools from state perception and trend prediction to intelligent decision-making, effectively reducing equipment failure rates and improving production efficiency and processing quality. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a time series prediction and decision-making method for multi-source data of CNC machine tools provided in this application embodiment; Figure 2 A schematic diagram illustrating the working condition identification and feature enhancement provided in the embodiments of this application; Figure 3 This is a schematic diagram of the multi-scale prediction model provided in the embodiments of this application. Detailed Implementation

[0018] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0019] Figure 1 This application provides a flowchart illustrating a time series prediction and decision-making method for multi-source data of CNC machine tools, as illustrated in the embodiments of this application. Figure 1 This embodiment provides a time series prediction and decision-making method for multi-source data of CNC machine tools, including: S101 collects multi-source heterogeneous data from CNC machine tools and performs time alignment and resampling to construct a multi-source time series matrix under a unified time scale.

[0020] In this embodiment, multi-source heterogeneous data is collected during machine tool operation. This data includes spindle vibration signals, spindle current and power signals, temperature sensor data, process parameters, machine tool operating status signals, and machining quality inspection data. Specifically, the spindle vibration signal is acquired via an accelerometer with a sampling frequency of 1kHz to 10kHz; the spindle current and power signals are acquired through an electrical parameter acquisition module with a sampling frequency of 1Hz to 10Hz; temperature sensor data includes spindle temperature and motor temperature, with a sampling frequency of 1Hz; process parameters, including rotational speed and feed rate, are read by the CNC system with a sampling frequency of 1Hz; machine tool operating status signals, including start / stop signals, alarm signals, and tool change status, are acquired through PLC system event triggering; and machining quality inspection data, including dimensional deviations and surface roughness, are periodically collected by the quality inspection system.

[0021] Due to differences in sampling frequencies across different data sources, a data structure with a unified time scale is constructed. A standard time axis is established by setting a uniform time interval, resulting in a standard time axis sequence. Based on the standard time axis sequence, windowed feature extraction is performed on the principal axis vibration signal to obtain a low-dimensional vibration feature vector.

[0022] In this embodiment, within the time interval window corresponding to each moment on the standard time axis, the original high-frequency vibration sampling sequence of the spindle vibration signal is extracted, and statistical features are extracted from the original high-frequency sampling sequence using the following formula: in, for The mean value of the vibration signal within the time window. This represents the number of sampling points within the window. for Standard deviation of vibration signal within time window for Peak value of vibration signal within the time window For the frequency domain sequence, the first... Each frequency component for The total frequency domain energy of the vibration signal within the time window. for Within the time window Each vibration signal sample value, for The first vibration signal sample value within the time window, for The second vibration signal sample value within the time window. for Within the time window Each vibration signal sample value is used. The mean, standard deviation, peak value, and total frequency domain energy of the vibration signal are concatenated to obtain the low-dimensional vibration feature vector at the current moment, as shown in the formula: in, For the corresponding time The low-dimensional vibration eigenvectors.

[0023] For spindle current and power signals, temperature sensor data, and machining quality inspection data, based on a standard time axis sequence, missing values ​​are filled in using linear interpolation. The formula is as follows: in, The current time to be interpolated on the standard time axis. The current time to be interpolated Previously, and with Adjacent sampled times, The current time to be interpolated Afterwards, and with Adjacent sampled times, These are the spindle current signals at time [time]. and The sampled values, Temperature sensor data at time 1 and The sampled values, Temperature sensor data at time 1 and The sampled values, , and Do not use the spindle current, temperature sensor data, and machining quality inspection data obtained through linear interpolation at time... The data value.

[0024] For process parameters and machine tool operating status signals, using a standard time axis sequence as a reference, missing values ​​are filled in using a forward hold method, with the following formula: in, For process parameters, This is a machine tool operating status signal. For process parameters at time The sampled values, For machine tool operating status signals at time The sampled values ​​are then concatenated after aligning various features by time to form a multi-source time series matrix with a unified time scale.

[0025] S102, based on the multi-source time series matrix, the processing condition is identified to obtain the condition label sequence.

[0026] Figure 2 The schematic diagram of working condition identification and feature enhancement provided in the embodiments of this application shows, from top to bottom, multi-source data input, working condition judgment result, and working condition driving feature enhancement; the multi-source data input module includes spindle current, speed, vibration amplitude, and operating status data items; the working condition judgment result includes four types of working conditions: no-load working condition, cutting working condition, abnormal fluctuation working condition, and tool changing working condition; the working condition driving feature enhancement module includes high-frequency vibration feature extraction, fluctuation feature extraction, steady-state feature extraction, and sudden change feature extraction.

[0027] In this embodiment, a multi-source time series matrix is ​​obtained, and operating condition identification parameters are extracted from the matrix. These parameters include: whether the spindle speed is zero, the magnitude and trend of the spindle current, the vibration amplitude and fluctuation level, the PLC status signal, and the current change rate. Cutting current thresholds, current fluctuation judgment thresholds, and vibration abrupt change judgment thresholds are set. The process iterates through each moment in the multi-source time series matrix, and operating condition is determined based on the identification parameters and the set thresholds, generating an operating condition label for each moment. When the spindle speed is zero, it is determined to be an unloaded working condition. When the spindle current exceeds the cutting current threshold and the vibration fluctuation is stable, it is determined to be a cutting working condition. When the current change exceeds the current fluctuation judgment threshold and the vibration amplitude change exceeds the vibration sudden change judgment threshold, it is determined to be an abnormal fluctuation working condition. When the PLC status signal is at the preset value, it is determined to be a tool changing working condition. When the spindle current exceeds the cutting current threshold, the current fluctuates slightly periodically and there is no sudden change in vibration amplitude, it is determined to be a normal fluctuation machining working condition. When the spindle current exceeds the cutting current threshold, the vibration fluctuates continuously at high frequency and there is no sudden change in amplitude, it is determined to be a high-frequency vibration machining working condition. The working condition labels at each time point are combined in chronological order to obtain the working condition label sequence.

[0028] S103, construct differentiated time series features based on the working condition label sequence to form a feature vector containing working condition semantic information.

[0029] In this embodiment, the operating condition label sequence and the corresponding multi-source time series matrix are obtained. Local time series data for the current moment are set. The corresponding operating condition type is matched according to the operating condition label at the current moment. Differentiated time series features are constructed according to the operating condition type. When the operating condition label represents a stable operating condition, a steady-state feature subset is extracted, including mean, moving average trend, and first-order difference features. When the operating condition label represents a fluctuating operating condition, a fluctuating feature subset is extracted, including variance, range, and rate of change features. When the operating condition label represents a high-frequency vibration operating condition, a high-frequency vibration feature subset is extracted, including frequency domain energy, peak factor, and kurtosis features. When the operating condition label represents an abnormal operating condition, an abnormal feature subset is extracted, including abrupt change amplitude and second-order difference features. The corresponding feature subset is matched according to the operating condition label. The matched feature subset is combined with the corresponding operating condition semantic information to obtain a feature vector containing the operating condition semantic information.

[0030] S104, based on the feature vector, constructs an adaptive time window according to the data volatility, spindle load rate and vibration energy change rate.

[0031] In this embodiment, a feature vector containing semantic information of working conditions is obtained. Based on the feature vector, the vibration fluctuation rate, current change rate and vibration energy change rate of the adaptive time window are calculated. The vibration fluctuation rate is the standard deviation of the vibration feature within the historical window, the current change rate is the absolute difference between the current and the current at the previous moment, and the vibration energy change rate is the absolute value of the difference between the current vibration frequency domain energy and the energy at the previous moment.

[0032] The length of the adaptive time window is dynamically adjusted based on the vibration fluctuation rate, current change rate, and vibration energy change rate, using the following formula: in, To adapt to the time window length, This is the proportionality coefficient. To prevent tiny constants with a denominator of zero, For vibrational fluctuation rate, The rate of change of current, The rate of change of vibrational energy.

[0033] When the operating condition is stable, the vibration fluctuation rate Current change rate and vibrational energy change rate The vibration fluctuation rate is relatively small, but the time window length is increased. When the operating status fluctuates or becomes abnormal, the vibration fluctuation rate is... Current change rate and vibrational energy change rate Increase the window size, and the time window will automatically shrink. You can also set the upper and lower limits of the window. in, This is the minimum allowable value for the adaptive time window length, used to limit the minimum range of the time window. This is the maximum allowed value for the adaptive time window length, used to limit the maximum range of the time window. The adaptive time window length is calculated based on the vibration fluctuation rate, current change rate, and vibration energy change rate.

[0034] This mechanism enables dynamic adjustment of the time series modeling scale under different operating conditions.

[0035] S105 uses the feature vectors within the adaptive time window as input to construct a multi-scale prediction model and outputs the prediction results.

[0036] See Figure 3 In this embodiment, the multi-scale prediction model is specifically used to predict four core indicators: spindle vibration trend, temperature change trend, power change trend, and machining quality fluctuation trend of CNC machine tools. It includes a TCN layer, an LSTM layer, and a Transformer layer. The outputs of the TCN layer, LSTM layer, and Transformer layer are connected to the input of the feature splicing layer. The output of the feature splicing layer is connected to the input of the fully connected layer. The output of the fully connected layer is connected to the prediction output layer.

[0037] Obtain the feature vector and the length of the adaptive time window. At each time step, construct a time series subsequence with a length equal to the adaptive window length. For each time series subsequence, construct a corresponding input feature subset according to the prediction target, including vibration prediction subsequence, temperature and current prediction subsequence, power prediction subsequence, and processing quality prediction subsequence. Construct a parallel multi-scale prediction model using TCN, LSTM, and Transformer. Input the vibration prediction subsequence into the TCN layer, and extract high-frequency local correlation features of vibration through multi-layer one-dimensional dilated causal convolution to obtain the local feature vector, as shown in the formula: in, for Vibration prediction subsequence within the time window; for The local feature vector output by the layer. For feature extraction of vibration prediction subsequences network.

[0038] The temperature and current prediction subsequences and the power prediction subsequences are input into the LSTM layer to extract the intermediate trend features, resulting in the intermediate trend feature vector, as shown in the formula: in, for Temperature and current prediction subsequences within the time window; for Power prediction subsequence within the time window; for The layer extracts the mid-term trend feature vector based on temperature and current prediction subsequences. for The layer extracts the mid-term trend feature vector based on the power prediction subsequence; To extract mid-term trend features from the temperature, current, and power prediction subsequences. network, For mid-term feature fusion function, This represents the feature vector of the medium-term trend.

[0039] The processing quality prediction subsequence is input into the Transformer layer, and long-term dependencies are captured through a self-attention mechanism to obtain the long-term dependency feature vector, as shown in the formula: in, for Processing quality prediction subsequence within the time window; for The long-term dependency feature vector output by the layer; To extract long-term dependency features for processing quality prediction subsequences network.

[0040] Calculate the learnable scale weights for the three types of features: In the formula, , , These are element-wise adaptive weight vectors for short-term, medium-term, and long-term features, respectively. , , For learnable weight matrix, for The activation function is used to map the weights to the 0~1 interval, thereby achieving adaptive weighting of three types of multi-scale features.

[0041] The local feature vector, intermediate trend feature vector, and long-term dependency feature vector are concatenated to form a multi-scale fused feature vector. This multi-scale fused feature vector is then input into a fully connected layer for adaptive weighted fusion. Trainable weight parameters adaptively learn the contribution of features at different scales to the prediction result, and the prediction result is output as follows: in, For multi-scale feature fusion function, for The prediction result at any given time; This is the weight matrix of the fully connected layer; For bias terms of fully connected layers; For the mapping function of the fully connected layer, This is a multi-scale fused feature vector.

[0042] S106, Generate operational decision instructions based on the prediction results.

[0043] In this embodiment, the prediction results output by the multi-scale prediction model are obtained, and safety thresholds and trend judgment rules are pre-set for each prediction index, including vibration safety threshold, temperature safety threshold and heating rate threshold, power fluctuation threshold and rate of change threshold, and mass fluctuation threshold. For the prediction sequence within the prediction time range, its maximum value, average value, and rate of change are calculated and compared with the corresponding thresholds to determine whether an early warning or optimization condition is triggered.

[0044] When the prediction results meet the corresponding threshold conditions, a corresponding operation decision instruction is generated. When the predicted vibration exceeds the vibration safety threshold or the rate of change is abnormal, a spindle maintenance warning instruction is generated. When the predicted temperature exceeds the temperature safety threshold or the heating rate exceeds the heating rate threshold, a cooling or shutdown suggestion is generated. When the predicted power exceeds the power fluctuation threshold or the rate of change exceeds the rate of change threshold, a machining parameter adjustment suggestion is generated. When the predicted quality fluctuation exceeds the quality fluctuation threshold, a tool change or machining parameter optimization suggestion is generated.

[0045] Decision-making instructions may include reducing feed rate, adjusting depth of cut, optimizing spindle speed parameters, activating enhanced cooling mode, scheduling maintenance, or performing safe shutdown operations. The generated instructions are sent to the machine tool control system, production management system, or equipment maintenance management system via an industrial communication interface, and are also displayed through a human-machine interface. In automatic control mode, the system can automatically perform parameter adjustments or safety control operations according to the set strategy; in manual confirmation mode, the corresponding instructions are executed after confirmation by the operator.

[0046] Corresponding to the time series prediction and decision-making method for multi-source data of CNC machine tools provided in the above embodiments, this application also provides a time series prediction and decision-making system for multi-source data of CNC machine tools.

[0047] This embodiment of a time series prediction and decision-making system for multi-source data of CNC machine tools includes: The data acquisition and processing module is used to acquire multi-source heterogeneous data from CNC machine tools and perform time alignment and resampling to construct a multi-source time series matrix under a unified time scale.

[0048] The working condition identification module is used to identify the processing working conditions based on a multi-source time series matrix and obtain a working condition label sequence.

[0049] The construction module is used to construct differentiated time series features based on the working condition label sequence, forming a feature vector containing working condition semantic information.

[0050] The time window module is used to construct an adaptive time window based on the feature vector, according to the data volatility, spindle load rate, and vibration energy change rate.

[0051] The prediction module is used to construct a multi-scale prediction model with feature vectors within an adaptive time window as input and output the prediction results.

[0052] The decision instruction generation module is used to generate execution decision instructions based on the prediction results.

[0053] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0054] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. A time series forecasting and decision-making method for multi-source data of CNC machine tools, characterized in that, include: Collect multi-source heterogeneous data from CNC machine tools, perform time alignment and resampling, and construct a multi-source time series matrix under a unified time scale; Based on the multi-source time series matrix, processing conditions are identified to obtain a processing condition label sequence; Based on the operating condition label sequence, differentiated time series features are constructed to form a feature vector containing operating condition semantic information; Based on the feature vector, an adaptive time window is constructed according to the data volatility, spindle load rate and vibration energy change rate. Using the feature vectors within the adaptive time window as input, a multi-scale prediction model is constructed, and the prediction results are output. Based on the prediction results, operational decision instructions are generated.

2. The time series prediction and decision-making method for multi-source data of CNC machine tools according to claim 1, characterized in that, The process of collecting multi-source heterogeneous data from CNC machine tools, performing time alignment and resampling, and constructing a multi-source time series matrix at a unified time scale includes: Multi-source heterogeneous data is collected during the machine tool operation process. The multi-source heterogeneous data includes spindle vibration signal, spindle current and power signal, temperature sensor data, process parameters, machine tool operating status signal and machining quality detection data. Set a uniform time interval, establish a standard time axis, and obtain the standard time axis sequence on the standard time axis; Based on the standard time axis sequence, windowed feature extraction is performed on the main shaft vibration signal to obtain a low-dimensional vibration feature vector. For the spindle current and power signals, temperature sensor data, and machining quality inspection data, using the standard time axis sequence as a reference, missing values ​​are filled in using linear interpolation, with the following formula: in, The current time to be interpolated on the standard time axis. The current time to be interpolated Previously, and with Adjacent sampled times, The current time to be interpolated Afterwards, and with Adjacent sampled times, These are the spindle current signals at time [time]. and The sampled values, Temperature sensor data at time 1 and The sampled values, Temperature sensor data at time 1 and The sampled values, , and Do not use the spindle current, temperature sensor data, and machining quality inspection data obtained through linear interpolation at time... Data values; For the process parameters and the machine tool operating status signals, using the standard time axis sequence as a reference, missing values ​​are filled in using a forward hold method, as shown in the formula: in, For process parameters, This is a machine tool operating status signal. For process parameters at time The sampled values, For machine tool operating status signals at time The sampled values; After aligning various features by time, they are concatenated to form a multi-source time series matrix with a unified time scale.

3. The time series prediction and decision-making method for multi-source data of CNC machine tools according to claim 2, characterized in that, Based on the standard timeline sequence, windowed feature extraction is performed on the principal shaft vibration signal to obtain a low-dimensional vibration feature vector, including: Within the time interval window corresponding to each moment on the standard time axis, the original high-frequency vibration sampling sequence of the spindle vibration signal is extracted; The statistical features of the original high-frequency sampling sequence are extracted using the following formula: in, for The mean value of the vibration signal within the time window. This represents the number of sampling points within the window. for Standard deviation of vibration signal within time window for Peak value of vibration signal within the time window For the frequency domain sequence, the first... Each frequency component for The total frequency domain energy of the vibration signal within the time window. for Within the time window Each vibration signal sample value, for The first vibration signal sample value within the time window, for The second vibration signal sample value within the time window. for Within the time window One vibration signal sample value; The mean, standard deviation, peak value, and total frequency domain energy of the vibration signal are concatenated to obtain the low-dimensional vibration feature vector at the current moment.

4. The time series prediction and decision-making method for multi-source data of CNC machine tools according to claim 1, characterized in that, Based on the multi-source time series matrix, processing condition identification is performed to obtain a processing condition label sequence, including: Obtain a multi-source time series matrix, and extract the working condition identification basis parameters from the multi-source time series matrix; Set the cutting current threshold, current fluctuation judgment threshold, and vibration sudden change judgment threshold; Iterate through each moment in the multi-source time series matrix, determine the working condition based on the working condition identification parameters and the set threshold, and generate a working condition label for each moment; When the spindle speed is zero, it is determined to be an unloaded working condition; When the spindle current exceeds the cutting current threshold and the vibration fluctuation is stable, it is determined to be a cutting condition; When the change in current exceeds the current fluctuation judgment threshold and the change in vibration amplitude exceeds the vibration sudden change judgment threshold, it is judged as an abnormal fluctuation condition. When the PLC status signal is at the preset value, it is determined to be a tool changing condition; When the spindle current exceeds the cutting current threshold, and the current fluctuates slightly periodically without sudden changes in vibration amplitude, it is determined to be a normal fluctuating machining condition. When the spindle current exceeds the cutting current threshold and the vibration continues to fluctuate at a high frequency without a sudden change in amplitude, it is determined to be a high-frequency vibration machining condition. By combining the operating condition labels at each time point in chronological order, a sequence of operating condition labels is obtained.

5. The time series prediction and decision-making method for multi-source data of CNC machine tools according to claim 1, characterized in that, Based on the operating condition label sequence, differentiated time series features are constructed to form a feature vector containing operating condition semantic information, including: Obtain the operating condition label sequence and the corresponding multi-source time series matrix; Set local time series data at the current moment, match the corresponding working condition type according to the working condition label at the current moment, and construct differentiated time series features according to the working condition type; When the operating condition label represents a stable operating condition, extract the steady-state feature subset; When the operating condition label indicates a fluctuating operating condition, extract the fluctuating feature subset; When the operating condition label represents a high-frequency vibration operating condition, extract the high-frequency vibration feature subset; When the operating condition label indicates an abnormal operating condition, extract the abnormal feature subset; The corresponding feature subset is matched based on the working condition label, and the matched feature subset is combined with the corresponding working condition semantic information to obtain a feature vector containing the working condition semantic information.

6. The time series prediction and decision-making method for multi-source data of CNC machine tools according to claim 1, characterized in that, Based on the aforementioned feature vector, an adaptive time window is constructed according to data volatility, spindle load rate, and vibration energy change rate, including: Obtain feature vectors containing semantic information about operating conditions; The vibration fluctuation rate, current change rate, and vibration energy change rate of the adaptive time window are calculated based on the feature vector. The length of the adaptive time window is dynamically adjusted based on the vibration fluctuation rate, current change rate, and vibration energy change rate, using the following formula: in, To adapt to the time window length, This is the proportionality coefficient. To prevent tiny constants with a denominator of zero, For vibrational fluctuation rate, The rate of change of current, The rate of change of vibrational energy.

7. The time series prediction and decision-making method for multi-source data of CNC machine tools according to claim 1, characterized in that, The multi-scale prediction model includes a TCN layer, an LSTM layer, and a Transformer layer. The outputs of the TCN layer, LSTM layer, and Transformer layer are connected to the input of the feature splicing layer. The output of the feature splicing layer is connected to the input of the fully connected layer. The output of the fully connected layer is connected to the prediction output layer.

8. The time series prediction and decision-making method for multi-source data of CNC machine tools according to claim 7, characterized in that, Using the feature vectors within the adaptive time window as input, a multi-scale prediction model is constructed, and the prediction results are output, including: Obtain the feature vectors within the adaptive time window and the length of the adaptive time window; At each time step, construct a time series subsequence of adaptive window length; The time series subsequences are constructed according to the prediction targets to form corresponding input feature subsets, including vibration prediction subsequences, temperature and current prediction subsequences, power prediction subsequences, and processing quality prediction subsequences; A parallel multi-scale prediction model combining TCN, LSTM, and Transformer is constructed. The vibration prediction subsequence is input into the TCN layer, and high-frequency local correlation features of vibration are extracted through multi-layer one-dimensional dilated causal convolution to obtain the local feature vector. The formula is as follows: in, for Vibration prediction subsequence within the time window; for The local feature vector output by the layer. For feature extraction of vibration prediction subsequences network; The temperature and current prediction subsequences and the power prediction subsequences are input into an LSTM layer to extract the intermediate trend features, resulting in the intermediate trend feature vector, as shown in the formula: in, for Temperature and current prediction subsequences within the time window; for Power prediction subsequence within the time window; for The layer extracts the mid-term trend feature vector based on temperature and current prediction subsequences. for The layer extracts the mid-term trend feature vector based on the power prediction subsequence; To extract mid-term trend features from the temperature, current, and power prediction subsequences. network, For mid-term feature fusion function, This represents the feature vector of the medium-term trend. The processing quality prediction subsequence is input into the Transformer layer, and long-term dependencies are captured through a self-attention mechanism to obtain the long-term dependency feature vector, as shown in the formula: in, for Processing quality prediction subsequence within the time window; for The long-term dependency feature vector output by the layer; To extract long-term dependency features for processing quality prediction subsequences network; The local feature vector, intermediate trend feature vector, and long-term dependency feature vector are concatenated to form a multi-scale fused feature vector: The multi-scale fused feature vector is input into a fully connected layer for adaptive weighted fusion, and the prediction result is output as follows: in, For multi-scale feature fusion function, for The prediction result at any given time; This is the weight matrix of the fully connected layer; For bias terms of fully connected layers; For the mapping function of the fully connected layer, This is a multi-scale fused feature vector.

9. The time series prediction and decision-making method for multi-source data of CNC machine tools according to claim 1, characterized in that, Based on the prediction results, operational decision instructions are generated, including: Obtain the prediction results output by the multi-scale prediction model, and pre-set safety thresholds and trend judgment rules for each prediction indicator; When the prediction result meets the corresponding threshold condition, the corresponding operation decision instruction is generated; When the predicted vibration exceeds the vibration safety threshold or the rate of change is abnormal, a spindle maintenance early warning command is generated. When the predicted temperature exceeds the temperature safety threshold or the heating rate exceeds the heating rate threshold, a cooling or shutdown recommendation is generated. When the predicted power exceeds the power fluctuation threshold or the rate of change exceeds the rate of change threshold, processing parameter adjustment suggestions are generated. When the predicted quality fluctuation exceeds the quality fluctuation threshold, suggestions for tool change or machining parameter optimization are generated.

10. A time series prediction and decision-making system for multi-source data of CNC machine tools, characterized in that, include: The data acquisition and processing module is used to acquire multi-source heterogeneous data from CNC machine tools and perform time alignment and resampling to construct a multi-source time series matrix under a unified time scale. The working condition identification module is used to identify the processing working conditions based on the multi-source time series matrix and obtain a working condition label sequence. The construction module is used to construct differentiated time series features based on the working condition label sequence, forming a feature vector containing working condition semantic information; The time window module is used to construct an adaptive time window based on the feature vector, according to the data volatility, spindle load rate and vibration energy change rate. The prediction module is used to construct a multi-scale prediction model with the feature vectors within the adaptive time window as input, and output the prediction results. The decision instruction generation module is used to generate execution decision instructions based on the prediction results.