A method and system for online control of machining accuracy based on multi-sensor data fusion

Abstract

This invention provides a method and system for online control of machining accuracy based on multi-sensor data fusion. The method includes the acquisition and preprocessing of a multimodal heterogeneous sensor dataset; weighted fusion using a weighted average method; extraction of aggregated feature vectors and time-frequency feature vectors using Local Feature Mean Aggregation (LFMA) and Wavelet Packet Transform (WPT); the introduction of a Fisher discriminant fusion algorithm to fuse the initial feature vectors; and prediction of machining error type and magnitude using an improved Extreme Learning Machine (ELM) model. Based on the error prediction results, an adaptive control strategy is constructed to control the machine tool online. This invention employs a multimodal data acquisition network using multiple types of sensors to acquire sensor parameters in real time, including dimensions, vibration, cutting force, temperature, and tool wear. Simultaneously, through a three-level hierarchical fusion architecture of data fusion, feature fusion, and decision fusion, the accuracy of error prediction is improved, thereby ensuring the accuracy of control.

Description

Technical Field

[0001] This invention relates to the field of intelligent machining control technology, and in particular to a method and system for online control of machining accuracy based on multi-sensor data fusion. Background Technology

[0002] In the field of modern machining, machining accuracy directly determines the core performance and market competitiveness of products. This is especially true in the manufacturing of precision parts, where the requirements for accuracy are becoming increasingly stringent. During the machining process, various factors such as machine tool vibration, changes in cutting temperature, tool wear, and fluctuations in cutting force can all cause machining errors. These errors are coupled with each other and are difficult to accurately capture using a single monitoring method.

[0003] Existing machining accuracy control methods mostly rely on single-sensor monitoring, such as using temperature sensors to monitor the temperature of the machining area, vibration sensors to monitor machine tool vibration, or vision sensors to detect workpiece dimensional deviations. However, single-sensor data has limitations: temperature sensors cannot reflect the impact of vibration on accuracy, and vibration sensors struggle to capture the cumulative errors caused by tool wear, resulting in incomplete monitoring information that cannot fully characterize the true state of the machining system. Even when some solutions use multiple sensors to collect data, they often involve simple data overlay processing without deep fusion, failing to effectively eliminate data redundancy and noise interference. Furthermore, they lack dynamic adaptive control mechanisms, making it difficult to cope with dynamic parameter changes in complex machining scenarios. This leads to control lag, inaccurate accuracy compensation, and an inability to meet the online control requirements of high-precision machining.

[0004] In addition, existing control methods are mostly based on fixed mathematical models, which are poorly adaptable to nonlinear and uncertain factors in the processing. When the processing materials, process parameters or environmental conditions change, the model's generalization ability is insufficient, and control failure is likely to occur, resulting in increased scrap rate and reduced production efficiency.

[0005] Therefore, there is an urgent need for a method that can integrate information from multiple sensor sources and achieve deep data fusion and dynamic, precise control. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for online control of machining accuracy based on multi-sensor data fusion. This invention constructs a hierarchical multi-sensor data fusion architecture and combines it with intelligent algorithms to achieve accurate identification of machining status and dynamic parameter control, thereby improving the stability of machining accuracy and the real-time performance of control.

[0007] To achieve the above objectives, the present invention adopts the following solution: In a first aspect, the present invention provides a method for online control of machining accuracy based on multi-sensor data fusion, specifically including the following steps: S1) Deploy multiple types of sensors at key locations on the machine tool to obtain a multimodal heterogeneous sensor dataset during the machining process, and preprocess the multimodal heterogeneous sensor dataset to obtain a standardized multimodal sensor dataset. S2) The preprocessed standardized multimodal sensor dataset is weighted and fused using a weighted average method to obtain a multimodal weighted fused dataset; S3) Use Local Feature Mean Aggregation (LFMA) to extract local features from the multimodal weighted fusion dataset and obtain the aggregated feature vector; S4) Extract the time-frequency feature vector of the multimodal weighted fusion dataset using wavelet packet transform (WPT); S5) The aggregated feature vector and the time-frequency feature vector are concatenated into an initial feature vector. The Fisher discriminant fusion algorithm is then introduced to fuse the initial feature vector to obtain the final fused feature vector. S6) Based on physical constraints, the results generated by the final fused feature vector input to the Extreme Learning Machine (ELM) model are corrected to obtain the error prediction results; S7) Based on the error prediction results, an adaptive control strategy is constructed to control the machine tool online.

[0008] Preferably, in step S1), the multimodal heterogeneous sensor dataset includes workpiece size deviation data, vibration acceleration signal data, cutting force data, machining area temperature data, and tool wear data.

[0009] Preferably, in step S2), the signal-to-noise ratio of each sensor is used as a basis. We dynamically assign weighted fusion weights to each sensor and perform weighted fusion on the standardized multimodal sensor dataset based on the weighted fusion weights to obtain a multimodal weighted fusion dataset.

[0010] Preferably, step S3) includes the following steps: S31) Partition the multimodal weighted fusion dataset into local windows, dividing it according to the time series. Non-overlapping windows; S32) Extract the local temporal features of each window, and finally obtain Local temporal feature matrix of each window; S33) The mean and standard deviation of each local temporal feature across all windows are calculated using the Local Feature Mean Aggregation (LFMA) algorithm to compress the dimension of the local temporal feature matrix. S34) The mean and standard deviation of each local time domain feature are concatenated sequentially to obtain the aggregated feature vector matrix.

[0011] Preferably, step S4) includes the following steps: S41) Wavelet packet decomposition initialization, using db4 wavelet basis pairs for... Layered wavelet packet decomposition yields multiple frequency band components; S42) Calculate the frequency band energy ratio, center frequency, and frequency band entropy of multiple frequency band components, and concatenate the frequency band energy ratio, center frequency, and frequency band entropy to obtain the frequency domain feature vector; S43) Extract global time-domain features from reconstructed signals with multiple frequency band components and multimodal weighted fusion datasets; S44) The global time-domain feature vector is concatenated with the frequency-domain feature vector to obtain the time-frequency feature vector.

[0012] Preferably, step S5) includes the following steps: S51) Based on historical processing data, the initial feature vector is classified into error levels. kind; S52) Calculate the intraclass scatter matrix The inter-class scatter matrix reflects the degree of dispersion of features among samples of the same class, while the intra-class scatter matrix reflects the degree of separation of features among samples of different classes. S53) Optimize the Fisher discriminant objective function by using the inverse of the intra-class scatter matrix and the transpose of the inter-class scatter matrix; S54) Find the extremum of the optimized Fisher discriminant objective function by solving the generalized eigenvalue problem; obtain the corresponding eigenvalues; and select the eigenvalues. The high-value feature values ​​are obtained by removing low-value features that are less than the average feature value; the removed low-value features are then sorted from largest to smallest and added to the high-value feature values; the weight coefficients of each feature dimension are obtained. S55) The filtered feature values ​​are weighted and fused according to their corresponding weights to obtain the final fused feature vector.

[0013] Secondly, the present invention provides an online machining accuracy control system based on multi-sensor data fusion, comprising: The data acquisition module is used to acquire multimodal heterogeneous sensor datasets during the machining process of the machine tool, and to preprocess the multimodal heterogeneous sensor datasets to obtain standardized multimodal sensor datasets. The dataset fusion module uses a weighted average method to perform weighted fusion on the preprocessed standardized multimodal sensor dataset to obtain a multimodal weighted fused dataset. The feature aggregation module uses Local Feature Mean Aggregation (LFMA) to extract local features from the multimodal weighted fusion dataset, resulting in an aggregated feature vector. The time-frequency feature extraction module uses wavelet packet transform (WPT) to extract the time-frequency feature vector of the multimodal weighted fusion dataset; The feature fusion module is used to concatenate the aggregated feature vector and the time-frequency feature vector into an initial feature vector. The Fisher discriminant fusion algorithm is introduced to fuse the initial feature vector to obtain the final fused feature vector. The error prediction module calls the Extreme Learning Machine (ELM) model and combines the physical constraints constructed by finite element analysis to predict the final fused feature vector, and outputs the processing error type and error amount. The control module constructs an adaptive control strategy based on error prediction results to control the machine tool online.

[0014] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-mentioned online control method for processing accuracy based on multi-sensor data fusion.

[0015] The beneficial effects of this invention are as follows: 1. This invention employs multiple types of sensors to construct a multimodal data acquisition network to acquire sensor parameters such as size, vibration, cutting force, temperature, and tool wear in real time. At the same time, through a three-level hierarchical fusion architecture of data fusion, feature fusion, and decision fusion, the accuracy of error prediction is improved, thereby ensuring the accuracy of control. 2. This invention dynamically calculates weights based on sensor signal-to-noise ratio (SNR), giving higher weights to high-precision data such as laser displacement sensors and tool wear sensors, while suppressing interference from low-SNR data such as temperature sensors. At the same time, it fuses standardized multimodal sensor datasets using a weighted average method, reducing data dimensionality while retaining core information, thus reducing the computational load for subsequent feature extraction. 3. This invention combines time-domain statistical features extracted by Local Feature Mean Aggregation (LFMA) with time-frequency features extracted by Wavelet Packet Transform (WPT) to achieve comprehensive coverage of instantaneous change features and global time-frequency features. LFMA captures instantaneous features such as cutting force peak and vibration change, while WPT covers full-band information, avoiding the limitations of single feature dimension representation and further improving the correlation between features and machining errors. 4. This invention introduces the Fisher discriminant fusion algorithm, which filters features by maximizing inter-class discriminability and minimizing intra-class dispersion. At the same time, through feature weight allocation, the weight of features strongly correlated with processing errors accounts for more than 65%, suppressing interference from redundant features. Furthermore, by compressing the local feature matrix, the data dimensionality is reduced while retaining key information, thereby improving inference speed. 5. The decision fusion of the present invention accurately predicts errors through the Extreme Learning Machine (ELM) model, realizes precise location of error sources, and adopts different control strategies for different error sources, thereby effectively improving the stability of processing accuracy. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention; Figure 2 This is a schematic diagram of the process for extracting the aggregated feature vector matrix in an embodiment of the present invention; Figure 3 This is a schematic diagram of the process for calculating the initial feature vector in an embodiment of the present invention; Figure 4 This is a schematic diagram of the system framework according to an embodiment of the present invention; In the diagram, 100 is the data acquisition module; 200 is the dataset fusion module; 300 is the feature aggregation module; 400 is the time-frequency feature extraction module; 500 is the feature fusion module; 600 is the error prediction module; and 700 is the control module. Detailed Implementation

[0017] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1 like Figure 1 As shown, this embodiment provides a method for online control of machining accuracy based on multi-sensor data fusion, which specifically includes the following steps: S1) Deploy multiple types of sensors at key locations on the machine tool to obtain a multimodal heterogeneous sensor dataset during the machining process, and preprocess the multimodal heterogeneous sensor dataset to obtain a standardized multimodal sensor dataset. In this embodiment, the various types of sensors include a laser displacement sensor (e.g., Keyence LK-G80), a triaxial vibration sensor (e.g., PCB 356A16), a cutting force sensor (e.g., Kistler 9257B), an infrared temperature sensor (e.g., FLIR TG267), and a tool wear sensor (using a Basler acA2040-180km vision sensor and a Keyence GT2-H12K resistive sensor).

[0018] The laser displacement sensor (model Keyence LK-G80) is deployed on the side of the spindle, with a measurement accuracy of ±0.001mm, and collects workpiece size deviation data; The three-dimensional vibration sensor (model PCB 356A16) is deployed on the bed, spindle box and worktable to collect vibration acceleration signal data; The cutting force sensor (model Kistler 9257B) is integrated into the grinding wheel holder to measure cutting force data; An infrared temperature sensor (model FLIR TG267) is positioned over the grinding area to collect temperature data of the processing area. The tool wear sensor uses a combination of a Basler acA2040-180km vision sensor (with a 50mm telecentric lens) and a Keyence GT2-H12K resistive sensor to collect tool wear data. The vision sensor is responsible for capturing the tool wear pattern, while the resistive sensor accurately measures the wear amount and monitors the wear status of the grinding wheel. The acquisition frequency is set to 200Hz.

[0019] In this embodiment, a multimodal heterogeneous sensor dataset is used. Preprocessing yields a standardized multimodal sensor dataset, as follows: S11) An adaptive wavelet thresholding algorithm is used to eliminate high-frequency noise in machine tool vibration acceleration signal data and cutting force data; specifically: S111) Decomposes the original information into low-frequency approximate components and multi-scale high-frequency detail components; S112) For the decomposed high-frequency detail components, the dynamic threshold is calculated using the quantization formula, i.e.: In the formula, The optimal threshold; The standard deviation of noise; The length of the high-frequency detail component; Wavelet coefficients of high-frequency detail components; This represents the median function, used to calculate the median of the high-frequency coefficient sequence after wavelet decomposition; S113) An improved soft threshold function is used to correct the high-frequency wavelet coefficients, namely: ; In the formula, These are the adjusted coefficients; It is a symbolic function; When | The value is determined to be the noise figure, and after correction, it approaches 0, thus achieving noise suppression. when It is determined to be the dominant factor of the signal. The signal details are preserved through nonlinear correction, while the amplitude attenuation of the traditional soft threshold is compensated to improve the signal fidelity. S114) Perform inverse wavelet transform on the processed high-frequency detail components and the unmodified low-frequency approximation components to reconstruct the denoised signal, thus completing the high-frequency noise elimination.

[0020] S12) Through The criteria for eliminating outlier data in the processing area temperature data and workpiece dimensional deviation data are as follows: First calculate the mean of the data with standard deviation Eliminate those that meet the requirements outliers; The corresponding sensor data is used; if the proportion of abnormal data exceeds 5%, linear interpolation is used to complete the data. S13) Adopt The algorithm normalizes the data from each sensor to the [0,1] interval, eliminating dimensional differences; S14) Based on timestamp synchronization technology, the spatiotemporal alignment of multi-source data is realized to ensure the corresponding matching of data from different sensors at the same time and output a standardized multimodal sensor dataset.

[0021] S2) The preprocessed standardized multimodal sensor dataset is weighted and fused using a weighted average method to obtain a multimodal weighted fused dataset. The details are as follows: Based on the signal-to-noise ratio of each sensor Dynamically assign weighted fusion weights to each sensor And based on the weighted fusion weight Weighted fusion of standardized multimodal sensor datasets, i.e.: ; ; In the formula, For the first The weighted fusion weights of each sensor; For the first The signal-to-noise ratio of each sensor; This represents the total number of sensors; For the first A standardized sensor dataset for each sensor.

[0022] S3) Utilize the Local Feature Mean Aggregation (LFMA) algorithm to perform multimodal weighted fusion on the dataset. Local feature extraction is performed to obtain an aggregated feature vector matrix; such as Figure 2 As shown, the details are as follows: S31) Multimodal weighted fusion dataset Perform local window partitioning, divided according to time series. Non-overlapping windows; S32) Extract the local temporal features of each window, and finally obtain Local temporal feature matrix of each window ; The local time-domain features of each window include six dimensions: mean, variance, peak value, kurtosis, waveform factor, and impulse factor. S33) Calculate the mean of each dimension's feature across all windows using the Local Feature Mean Aggregation (LFMA) algorithm. with standard deviation To compress the local time domain feature matrix The dimension; that is: In the formula, For the first The mean of the dimensional features; For the first Standard deviation of dimensional features; Indicates the first The first local window Dimensional features; ; S34) The mean and standard deviation of each dimension are concatenated sequentially to obtain a 12-dimensional aggregated feature vector matrix. .

[0023] S4) Extract the multimodal weighted fusion dataset using wavelet packet transform (WPT). Time-frequency eigenvectors Specifically, the steps include the following: S41) Wavelet packet decomposition initialization, using db4 wavelet basis pairs for... Layered wavelet packet decomposition yields 8 frequency band components. ; S42) Calculate multiple frequency band components Frequency band energy percentage Center frequency Frequency band entropy By adjusting the frequency band energy ratio Center frequency Frequency band entropy Concatenation forms frequency domain feature vector ; Among them, the frequency band energy ratio It is the ratio of the energy of each frequency band component to the total energy of all frequency band components.

[0024] In this embodiment, the top two frequency band components by energy percentage are selected, and their center frequencies are calculated. The center frequency of a frequency band component is the ratio of the product of the amplitudes of all discrete frequency points and their corresponding amplitudes within that frequency band component to the amplitudes of all frequency points; that is: In the formula, Indicates the first The center frequency of each frequency band component; Indicates the first The first frequency band component A discrete frequency point; For the first The first frequency band component The amplitude of a signal at a discrete frequency point; For the first The number of discrete frequency points in each frequency band component; The frequency band component numbers are the top two in terms of energy percentage; In this embodiment, the two frequency band components with the highest energy percentage are selected, and their frequency band entropy is calculated. ,Right now: In the formula, For the first The frequency band entropy of each frequency band component; For the first The first frequency band component The normalized probability of each discrete amplitude; For the first The number of discrete amplitudes within each frequency band component.

[0025] S43), based on multiple frequency band components Reconstructed signal and multimodal weighted fusion dataset Extracting global temporal features, the global temporal features include: Extract basic statistical features, including global peak value, global mean, global variance, and global kurtosis; Extract morphological features, including skewness, waveform factor, impulse factor, and peak factor; Extract trend features, including the mean amplitude difference between adjacent frequency bands, amplitude change rate, frequency of peak occurrence, and cumulative amplitude sum.

[0026] S44) Global temporal features With frequency domain eigenvectors By concatenating the vectors, we obtain the time-frequency feature vector. .

[0027] S5) Concatenate the aggregated feature vector and the time-frequency feature vector to form the initial feature vector. And introduce the Fisher discriminant fusion algorithm to process the initial feature vector. The feature vectors are then fused to obtain the final fused feature vector. ;like Figure 3As shown, the specific steps include the following: S51) Based on historical processing data, the initial feature vector Error levels are classified as follows kind; S52) Calculate the intraclass scatter matrix Inter-class scatter matrix Intra-class scatter matrix It reflects the degree of dispersion of characteristics of samples of the same type, that is: ; In the formula, Total number of categories; For the first Number of samples per class; For the first The first in the class The initial feature vector of each sample; For the first The feature mean vector of the class samples; the inner layer accumulates and traverses the first... All samples of each class are summed, and the outermost layer iterates through all classes. This indicates the transpose operation.

[0028] The inter-class scatter matrix The degree of separation of features of different categories of samples, i.e.: ; In the formula, This is the global feature mean vector for all samples; This represents the total number of samples. S53) The Fisher discriminant objective function is optimized by transposing the inter-class scatter matrix and using the inverse of the intra-class scatter matrix. The optimized Fisher discriminant objective function is... Represented as: In the formula, The inter-class scatter matrix transpose; It is the inverse of the intra-class scatter matrix; This is the matrix trace operation, which is the sum of the elements on the main diagonal of the matrix; S54) The optimized Fisher discriminant objective function Finding extrema by solving the generalized eigenvalue problem ; obtain the corresponding eigenvalues ; For the initial feature vector The Middle The eigenvalues ​​corresponding to each eigenvector; The number of initial feature vectors; And filter feature values High-value characteristic values, eliminated The low-value features are identified; these low-value features are then sorted from largest to smallest and added to the high-value feature values; finally, the weight coefficients for each feature dimension are obtained. ; S55) The filtered feature values ​​are weighted and fused according to their corresponding weights to obtain the final fused feature vector. ,Right now: In the formula, For the initial feature vector The 1 eigenvector.

[0029] S6) Based on physical constraints, the results generated by the final fused feature vector input to the Extreme Learning Machine (ELM) model are corrected to obtain the error prediction results; In this embodiment, the Extreme Learning Machine (ELM) model includes an input layer, a hidden layer, and an output layer. A Genetic Algorithm (GA) is used to optimize the connection weights between the input and hidden layers and the threshold of the hidden layer in the ELM model. The GA parameters are set as follows: population size 50 (i.e., generating 50 parameter combinations per iteration), number of iterations 100, crossover probability 0.7 (controlling the crossover and mutation probability of parameter combinations), and mutation probability 0.05 (avoiding the algorithm getting trapped in local optima). The optimization objective is to minimize the model's prediction error. ,Right now: ; In the formula, This represents actual processing error data; For the predicted results; The optimal cutting parameter combination is obtained through iterative optimization and assigned to the Extreme Learning Machine (ELM) model, which is based on the 2-norm. Finite element simulation analysis is performed based on the machining scenario to clarify the physical constraint boundaries in the machining process and form a set of constraint conditions. Taking bearing ring grinding as an example, the upper limit constraint of grinding force F≤3000N (exceeding this value can easily cause workpiece deformation) and the upper limit constraint of machining area temperature T≤700℃ are obtained through finite element analysis (excessive temperature leads to changes in material properties and aggravates errors). At the same time, combined with the mechanical properties of the workpiece material, a reasonable range constraint of error is added (such as the grinding error of bearing ring ≤0.02mm). The constraint conditions will serve as the basis for correcting the prediction results of the Extreme Learning Machine (ELM) model.

[0030] The final fused feature vector is obtained through the input layer. The input is an Extreme Learning Machine (ELM) model. The model is then propagated forward through its hidden layers, and the output layer outputs preliminary error type codes and predicted error values. Let the optimal connection weight matrix between the input layer and the hidden layer be... The optimal threshold vector of the hidden layer is Then the first Output of each hidden layer node for: In the formula, For the first The hidden layer node and the first The connection weights of each input layer node; For the final fused feature vector The One feature; Use the Sigmoid activation function; Let the connection weight matrix between the hidden layer and the output layer be... The initial prediction results of the output layer Represented as: In the formula, For the first The hidden layer node and the first The connection weights of each output layer node; Encode the initial error type; This is the preliminary error prediction value.

[0031] The preliminary prediction results are compared and verified with the finite element physical constraints. If the prediction error is within a reasonable range of the physical constraints and the error type is consistent with the error trend derived from the physical constraints, the final prediction result is output. If the prediction result exceeds the range of the physical constraints, the error is corrected based on the constraint boundary, and the coding credibility is adjusted in combination with the error type.

[0032] The physical constraints are as follows: Grinding force constraint: During grinding, the cutting force of the grinding wheel on the workpiece must be less than the critical force corresponding to the yield strength of the workpiece material to avoid workpiece deformation. The expression is: ,in The grinding force is measured by a cutting force sensor. The critical grinding force; Temperature constraints in the machining zone: The temperature in the grinding zone must be lower than the material's phase transformation temperature to prevent material property degradation. ,in The actual temperature measured by the infrared temperature sensor. This refers to the critical temperature of the material's phase transformation. Prediction error constraints: ; The prediction error of the Extreme Learning Machine (ELM) model; The maximum allowable error; Grinding wheel-workpiece contact constraint: Based on finite element contact mechanics analysis, the contact pressure and contact area between the grinding wheel and the workpiece are limited, expressed as follows: ;in For contact pressure; To allow for high contact pressure; Contact area; This represents the minimum permissible contact area.

[0033] S7) Based on the error prediction results, construct an adaptive control strategy; In this embodiment, if the error source is machine tool vibration, the machine tool damping parameters are adjusted by a PID controller, i.e.: In the formula, To control the output; This is the proportionality coefficient; This refers to the deviation in vibration amplitude. The integral time constant; The differential time constant; At the same time, the machine tool spindle is finely adjusted, namely: ; In the formula, The preset safety threshold for machine tool vibration amplitude; The vibration amplitude of the machine tool is monitored in real time; This is the speed adjustment coefficient.

[0034] In this embodiment, if the error source is that the cutting temperature is greater than a preset temperature threshold... The cutting fluid flow rate and injection angle are dynamically adjusted. The formula for adjusting the cutting fluid flow rate is: ; In the formula, The adjusted cutting fluid flow rate; This represents the initial cutting fluid flow rate; This is the flow rate adjustment coefficient; For real-time monitoring of cutting temperature; The flow rate should be adjusted within the range of 5-20 L / min; if it exceeds the range, the boundary value should be used. The amount of injection angle adjustment is directly proportional to the temperature deviation, as shown in the formula: ; In the formula, This is the amount of angle adjustment; This is the angle adjustment coefficient.

[0035] In this embodiment, if the error source is tool wear or cutting force fluctuation, the particle swarm optimization algorithm is used to solve for the optimal combination of cutting parameters. , This refers to the cutting speed; For feed rate; This refers to the depth of cut. PSO algorithm parameters: number of particles 30, number of iterations 50, learning factor Inertial weight The objective function of the PSO algorithm decreases linearly from 0.9 to 0.4. for: ; In the formula, , , For weighted systems; This represents the total machining error. , , These are respectively dimensional error, positional error, and shape error; in, ; ; ; In the formula, All are fitting coefficients; , These are the inherent positioning error of the machine tool and the inherent shape error of the workpiece material, respectively. The amplitude of machine tool vibration; Particle Update and Optimal Solution Iteration: Initialize the positions and velocities of 30 particles. During the iteration process, each particle updates its position and velocity based on its own optimal position and the group's optimal position; that is: Speed ​​updates: ; Location update: ; In the formula, Number the particles; , respectively corresponding ; , A random number in the range [0,1]. This is the optimal position for the particle itself; The optimal position for the group; To indicate the first During the nth iteration, the 1st Particles in the first Speed ​​on the dimensional parameter; For the first During the nth iteration, the 1st Particles in the first Position on the dimension parameter.

[0036] After 50 iterations, the parameter combination corresponding to the optimal position of the population is the optimal cutting parameter.

[0037] In this embodiment, the constraints are set as follows: , , .

[0038] Example 2 like Figure 4 As shown, this embodiment provides an online machining accuracy control system based on multi-sensor data fusion, including: The data acquisition module 100 is used to acquire multimodal heterogeneous sensor datasets during the machining process of the machine tool, and to preprocess the multimodal heterogeneous sensor datasets to obtain standardized multimodal sensor datasets. The dataset fusion module 200 uses a weighted average method to perform weighted fusion on the preprocessed standardized multimodal sensor dataset to obtain a multimodal weighted fused dataset. The feature aggregation module 300 uses Local Feature Mean Aggregation (LFMA) to extract local features from the multimodal weighted fusion dataset and obtain an aggregated feature vector. The time-frequency feature extraction module 400 uses wavelet packet transform (WPT) to extract the time-frequency feature vector of the multimodal weighted fusion dataset; The feature fusion module 500 is used to concatenate the aggregated feature vector and the time-frequency feature vector into an initial feature vector, and introduce the Fisher discriminant fusion algorithm to fuse the initial feature vector to obtain the final fused feature vector; The error prediction module 600 calls the Extreme Learning Machine (ELM) model and combines the physical constraints constructed by finite element analysis to predict the final fused feature vector, and outputs the processing error type and error amount. The control module 700 constructs an adaptive control strategy based on error prediction results to control the machine tool online.

[0039] Example 3 This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the online control method described in Embodiment 1.

[0040] In this embodiment, the memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. A processor, coupled to the memory, is used to execute computer programs stored in the memory.

[0041] The computer program includes computer program code, which may be in the form of source code, object code, executable file, or some intermediate form.

[0042] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A method for online control of machining accuracy based on multi-sensor data fusion, characterized in that, Includes the following steps: S1) Obtain the multimodal heterogeneous sensor dataset during the machining process of the machine tool, and preprocess the multimodal heterogeneous sensor dataset to obtain a standardized multimodal sensor dataset; S2) The preprocessed standardized multimodal sensor dataset is weighted and fused using a weighted average method to obtain a multimodal weighted fused dataset; S3) Use Local Feature Mean Aggregation (LFMA) to extract local features from the multimodal weighted fusion dataset and obtain the aggregated feature vector; S4) Extract the time-frequency feature vector of the multimodal weighted fusion dataset using wavelet packet transform (WPT); S5) The aggregated feature vector and the time-frequency feature vector are concatenated into an initial feature vector. The Fisher discriminant fusion algorithm is then introduced to fuse the initial feature vector to obtain the final fused feature vector. S6) Based on physical constraints, the results generated by the final fused feature vector input to the Extreme Learning Machine (ELM) model are corrected to obtain the error prediction results; S7) Based on the error prediction results, an adaptive control strategy is constructed to control the machine tool online.

2. The method for online control of machining accuracy based on multi-sensor data fusion according to claim 1, characterized in that: In step S2), based on the signal-to-noise ratio of each sensor... We dynamically assign weighted fusion weights to each sensor and perform weighted fusion on the standardized multimodal sensor dataset based on the weighted fusion weights to obtain a multimodal weighted fusion dataset.

3. The method for online control of machining accuracy based on multi-sensor data fusion according to claim 2, characterized in that: Step S3) specifically includes the following steps: S31) Partition the multimodal weighted fusion dataset into local windows, dividing it according to the time series. Non-overlapping windows; S32) Extract the local temporal features of each window, and finally obtain Local temporal feature matrix of each window; S33) The mean and standard deviation of each local temporal feature across all windows are calculated using the Local Feature Mean Aggregation (LFMA) algorithm to compress the dimension of the local temporal feature matrix. S34) The mean and standard deviation of each local time domain feature are concatenated sequentially to obtain the aggregated feature vector matrix.

4. The method for online control of machining accuracy based on multi-sensor data fusion according to claim 3, characterized in that: Step S4) specifically includes the following steps: S41) Wavelet packet decomposition initialization, using db4 wavelet basis pairs for... Layered wavelet packet decomposition yields multiple frequency band components; S42) Calculate the frequency band energy ratio, center frequency, and frequency band entropy of multiple frequency band components, and concatenate the frequency band energy ratio, center frequency, and frequency band entropy to obtain the frequency domain feature vector; S43) Extract global time-domain features from reconstructed signals with multiple frequency band components and multimodal weighted fusion datasets; S44) The global time-domain feature vector is concatenated with the frequency-domain feature vector to obtain the time-frequency feature vector.

5. The online control method for machining accuracy based on multi-sensor data fusion according to claim 4, characterized in that: Step S5) specifically includes the following steps: S51) Based on historical processing data, the initial feature vector is classified into error levels. kind; S52) Calculate the intraclass scatter matrix The inter-class scatter matrix reflects the degree of dispersion of features among samples of the same class, while the intra-class scatter matrix reflects the degree of separation of features among samples of different classes. S53) Optimize the Fisher discriminant objective function by using the inverse of the intra-class scatter matrix and the transpose of the inter-class scatter matrix; S54) Find the extremum of the optimized Fisher discriminant objective function by solving the generalized eigenvalue problem; Obtain the corresponding eigenvalues; And filter feature values High-value feature values ​​are selected, while low-value features that are less than the average feature value are removed. The low-value features that were removed are sorted from largest to smallest according to their feature values ​​and then added to the high-value feature values; thus, the weight coefficients of each feature dimension are obtained. S55) The filtered feature values ​​are weighted and fused according to their corresponding weights to obtain the final fused feature vector.

6. The method for online control of machining accuracy based on multi-sensor data fusion according to claim 5, characterized in that: In step S6), the Extreme Learning Machine (ELM) model includes an input layer, a hidden layer, and an output layer; and the Genetic Algorithm (GA) is used to optimize the connection weights between the input layer and the hidden layer and the threshold of the hidden layer of the ELM model; the final fused feature vector is input into the ELM model through the input layer, and the forward propagation of the hidden layer of the ELM model is used for calculation, and the preliminary error type code and error prediction value are output through the output layer.

7. The method for online control of machining accuracy based on multi-sensor data fusion according to claim 6, characterized in that: In step S6), let the optimal connection weight matrix between the input layer and the hidden layer be... The optimal threshold vector of the hidden layer is Then the first Output of each hidden layer node for: In the formula, For the first The hidden layer node and the first The connection weights of each input layer node; For the final fused feature vector The One feature; Use the Sigmoid activation function; Let the connection weight matrix between the hidden layer and the output layer be... The initial prediction results of the output layer Represented as: In the formula, For the first The hidden layer node and the first The connection weights of each output layer node; Encode the initial error type; This is the preliminary error prediction value; The physical constraints are as follows: Grinding force constraint: ,in The grinding force is measured by a cutting force sensor. The critical grinding force; Temperature constraints in the processing area: ,in The actual temperature measured by the infrared temperature sensor. This refers to the critical temperature of phase transformation in the material. Prediction error constraints: ; The prediction error of the Extreme Learning Machine (ELM) model; The maximum allowable error; Grinding wheel-workpiece contact constraint: ;in For contact pressure; To allow for high contact pressure; Contact area; This represents the minimum permissible contact area.

8. The method for online control of machining accuracy based on multi-sensor data fusion according to claim 7, characterized in that: In step S7), if the error source is machine tool vibration, the machine tool damping parameters are adjusted by the PID controller, and the machine tool spindle is fine-tuned at the same time. If the error source is that the cutting temperature is greater than the preset temperature threshold, the cutting fluid flow rate and spray angle will be dynamically adjusted. If the error source is tool wear or cutting force fluctuation, then the particle swarm optimization algorithm is used to solve for the optimal combination of cutting parameters.

9. A machining accuracy online control system based on multi-sensor data fusion, characterized in that, include: The data acquisition module (100) is used to acquire the multimodal heterogeneous sensor dataset during the machining process of the machine tool, and to preprocess the multimodal heterogeneous sensor dataset to obtain a standardized multimodal sensor dataset. The dataset fusion module (200) uses a weighted average method to perform weighted fusion on the preprocessed standardized multimodal sensor dataset to obtain a multimodal weighted fused dataset; The feature aggregation module (300) uses Local Feature Mean Aggregation (LFMA) to extract local features from the multimodal weighted fusion dataset and obtains the aggregated feature vector. The time-frequency feature extraction module (400) uses wavelet packet transform (WPT) to extract the time-frequency feature vector of the multimodal weighted fusion dataset; The feature fusion module (500) is used to concatenate the aggregated feature vector and the time-frequency feature vector into an initial feature vector, and introduce the Fisher discriminant fusion algorithm to fuse the initial feature vector to obtain the final fused feature vector; The error prediction module (600) calls the Extreme Learning Machine (ELM) model and combines the physical constraints constructed by finite element analysis to predict the final fused feature vector, and outputs the processing error type and error amount. The control module (700) constructs an adaptive control strategy based on the error prediction results.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the online control method for machining accuracy based on multi-sensor data fusion as described in any one of claims 1-8.

Images