Processing parameter adaptive control method and system for multi-source heterogeneous data fusion

Abstract

The invention provides a multi-source heterogeneous data fusion processing parameter adaptive control method and system, and belongs to the technical field of intelligent manufacturing, and the method comprises the steps: obtaining multi-source processing data, and carrying out the data preprocessing; inputting the preprocessed multi-source processing data into a multi-modal feature fusion model, synchronously extracting local abnormal features and global correlation features in the processing process, and fusing to obtain a fused feature vector; inputting the fusion feature vector into a machining quality prediction model to obtain a surface roughness prediction value and a size error prediction value, and further obtaining a comprehensive machining error; on the basis of the comprehensive machining error, the parameter optimization model is used for optimizing the machining parameters, a parameter adjusting instruction is generated and sent to a machine tool numerical control system to drive the machining parameters to be updated, workpiece material characteristics and environment parameters are included, machining of various materials is supported, intelligent and precise regulation and control of the machining process are achieved, and the machining efficiency is improved. The problems of poor adaptability, one-sided information and no self-optimization capability in the prior art are solved.

Description

Technical Field

[0001] This invention relates to a method and system for adaptive control of processing parameters based on multi-source heterogeneous data fusion, belonging to the field of intelligent manufacturing technology. Background Technology

[0002] In machining processes, abnormal conditions such as tool wear, excessive vibration, and excessive load are key factors affecting product quality, production efficiency, and equipment lifespan. Industry data shows that approximately 30% of scrap rates and 20% of equipment failures stem from undetected machining abnormalities.

[0003] With the development of intelligent manufacturing technology, processing anomaly monitoring systems are gradually upgrading to automation and intelligence. Traditional processing monitoring relies on manual inspection or simple threshold alarms, which is difficult to adapt to complex and ever-changing processing conditions such as different materials, workpiece types, and tool status. Processing monitoring systems with poor adaptability are prone to false alarms, missed alarms, or excessive downtime, resulting in production losses.

[0004] Several existing technologies offer solutions, such as patent CN116184960A, which describes an adaptive control method for CNC machine tool process parameters that considers tool status. This method primarily uses RBF and RNN neural networks to predict tool status and spindle power, combined with fuzzy control to adjust the feed rate. However, this approach only adjusts the feed rate based on tool status and power prediction, without considering spindle speed optimization, and relies on a fuzzy rule base. Another example is patent CN117250912A, a multi-neural network fusion method that uses RBF, stacked autoencoders, and random forests to adjust feed rate and spindle speed, while also incorporating chatter suppression. Although it employs multi-neural network fusion, it requires extensive pre-training data, limiting real-time performance, and it lacks a physical mechanism model.

[0005] In summary, existing technologies for intelligent control of machining still have the following shortcomings:

[0006] (1) Poor adaptability: The fixed threshold is used and the differences in processing materials and workpiece size under different processing scenarios are not considered, which easily leads to false alarms or missed alarms.

[0007] (2) Incomplete information: The monitoring data is too simple and it is difficult to distinguish the causes of abnormalities, resulting in a lack of targeted control strategies.

[0008] (3) Passive response: Most existing systems only alarm after an anomaly occurs, rather than actively intervening in the early stages of an anomaly, which often leads to small anomalies escalating into major faults.

[0009] (4) No self-optimization capability: Control parameters need to be preset manually and cannot be optimized by learning from historical processing data. With long-term use, the accuracy will decrease as the equipment ages and the performance of the cutting tools deteriorates. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive control method for processing parameters based on the fusion of multi-source heterogeneous data, thereby solving the problems.

[0011] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0012] In a first aspect, the present invention provides an adaptive control method for processing parameters of multi-source heterogeneous data fusion, comprising:

[0013] S1: Acquire multi-source machining data and perform data preprocessing; wherein, the multi-source machining data includes: tool status data, workpiece material spectral data, machining environment data, cutting force time-domain data, and workpiece surface roughness;

[0014] S2: Input the preprocessed multi-source processing data into the pre-constructed multimodal feature fusion model, simultaneously extract local abnormal features and global correlation features during the processing, and fuse them to obtain a fusion feature vector;

[0015] S3: Input the fused feature vector into the trained processing quality prediction model to obtain the surface roughness prediction value and the dimensional error prediction value, and then obtain the comprehensive processing error;

[0016] S4: Based on the comprehensive machining error, the machining parameters are optimized using a parameter optimization model to generate parameter adjustment instructions;

[0017] S5: Send the parameter adjustment command to the machine tool CNC system to drive the machining parameters to update, and return to execute S1 based on the updated machining status to form a closed-loop control.

[0018] Furthermore, the multimodal feature fusion model includes: an input layer, a dual-channel feature extraction layer, an attention fusion layer, and an output layer; the dual-channel feature extraction layer includes a CNN branch and a Transformer branch; the CNN branch takes tool state data, cutting force time-domain data, and workpiece surface roughness as input to extract local anomaly features during the machining process; the Transformer branch takes workpiece material spectral data and machining environment data as input to extract global correlation features during the machining process; the attention fusion layer adopts a scaled dot product attention mechanism, dynamically assigns weights according to the degree of influence of local anomaly features and global correlation features on machining quality, and obtains a fused feature vector through weighted fusion.

[0019] Furthermore, the process of obtaining the comprehensive machining error includes: obtaining the surface roughness error based on the predicted surface roughness value and the target surface roughness value required by the workpiece design; obtaining the dimensional error margin based on the predicted dimensional error value and the target dimensional error value; normalizing the surface roughness error and the dimensional error margin, and then obtaining the comprehensive machining error by weight superposition.

[0020] The method for determining the target value of the dimensional error is as follows: A basic target value of the dimensional error is obtained based on the dimensional tolerance range indicated on the workpiece design drawings; the basic target value of the dimensional error is dynamically corrected based on the deformation characteristics of the processed material to obtain the target value of the dimensional error. The specific formula is as follows:

[0021] ;

[0022] In the formula, The target value for dimensional error. This is the basic target value for dimensional error. This is the material deformation coefficient.

[0023] Furthermore, the processing quality prediction model adopts a multi-output GBDT model architecture; the weight update method of the processing quality prediction model includes: sampling and detecting the actual value of the size error using a laser interferometer, and calculating the prediction accuracy of the multi-output GBDT model based on the actual value of the size error; in response to the prediction accuracy being lower than a preset threshold, immediately triggering the weight update of the multi-output GBDT model; and updating the weights of the multi-output GBDT model every m control cycles using an online gradient descent algorithm.

[0024] Furthermore, the optimization parameters of the parameter optimization model include: feed rate, spindle speed, depth of cut, and cutting fluid concentration. Its objective function is a dual-objective fusion function formed by weighted fusion of the comprehensive error objective function and the efficiency objective function. The comprehensive error objective function aims to minimize the comprehensive machining error, and the efficiency objective function aims to maximize the material removal rate.

[0025] The formula for calculating the material removal rate is:

[0026] ;

[0027] In the formula, For material removal rate, For feed rate, For cutting depth, This represents the cutting width.

[0028] Furthermore, the constraints of the parameter optimization model include: feed rate constraint: the feed rate is not greater than the maximum rated speed of the machine tool feed system; spindle power constraint: the actual cutting power is not greater than the maximum rated power of the machine tool spindle; spindle torque constraint: the actual cutting torque is not greater than the maximum rated torque of the machine tool spindle.

[0029] Surface roughness error constraint: The surface roughness predicted by the multi-output GBDT model shall not exceed the target value of surface roughness required by the workpiece design; Dimensional error constraint: The dimensional error predicted by the multi-output GBDT model shall not exceed the target value of dimensional error corresponding to the workpiece design tolerance; Tool life constraint: The measured value of tool wear shall not exceed the scrap threshold corresponding to the tool material.

[0030] It also includes: cutting speed constraints and cutting force constraints set according to different processing materials.

[0031] Furthermore, it also includes: using a particle swarm optimization algorithm with adaptive inertia weight optimization to solve the parameter optimization model; wherein, the cutting fluid concentration is a discrete variable, which is first optimized into a continuous value in the algorithm iteration, and then the optimization result is mapped back to the original discrete value after the iteration is completed; when the number of iterations reaches the preset number, the optimal parameter combination is output.

[0032] Furthermore, a penalty function method is used to incorporate the constraints into the objective function of the parameter optimization model to ensure that the algorithm's search direction conforms to the constraints. Specifically:

[0033] If any particle violates a constraint, a fitness penalty is added to the violating particle to correct the objective function, specifically: ;

[0034] In the formula: The objective function after penalty, For the unpenalized objective function, The penalty coefficient is... To what extent it violates the rules.

[0035] Furthermore, it also includes: obtaining the error change rate within the control cycle based on the comprehensive machining error; inputting the comprehensive machining error and the error change rate into the fuzzy PID controller, obtaining real-time optimized PID parameters through fuzzy inference and PID control, and then calculating the adjustment amount of the optimized parameters; sending the adjustment amount of the optimized parameters to the machine tool CNC system to drive the machine tool to perform parameter updates.

[0036] In a second aspect, the present invention provides a machining parameter adaptive control system based on multi-source heterogeneous data fusion, used to implement the machining parameter adaptive control system method as described in any of the first aspects, comprising:

[0037] The multi-source data acquisition module is used to acquire multi-source heterogeneous data during the machining process, including tool status data, workpiece material spectral data, machining environment data, cutting force time-domain data, and workpiece surface roughness.

[0038] The data preprocessing module is used to preprocess multi-source heterogeneous data;

[0039] The multimodal feature fusion module is used to synchronously extract local abnormal features and global correlation features during the processing based on the preprocessed data, and fuse them to obtain a fused feature vector;

[0040] The machining quality prediction module is used to obtain the predicted values ​​of surface roughness and dimensional error using the machining quality prediction model, and then obtain the comprehensive machining error.

[0041] The parameter optimization module is used to optimize the processing parameters using a parameter optimization model and generate parameter adjustment instructions;

[0042] A fuzzy PID controller is used to dynamically adjust the proportional, integral, and derivative parameters of a PID controller through fuzzy inference, and output the real-time adjustment of the optimized parameters.

[0043] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0044] (1) This invention achieves intelligent and precise control of the processing process through a full-link technical architecture of "multi-source data acquisition and preprocessing → multi-modal feature fusion → quality prediction and parameter optimization → real-time closed-loop control", which solves the problems of poor adaptability, passive response, one-sided information and lack of self-optimization capability of existing technologies.

[0045] (2) This invention incorporates workpiece material properties and environmental parameters, supports the processing of various materials such as steel, aluminum, and titanium alloys, and improves the stability of processing quality by more than 25%. At the same time, it realizes the linkage control of feed rate, spindle speed, cutting depth and cutting fluid concentration, taking into account both accuracy and efficiency. During the processing, the processing quality is also fed back in real time through a laser interferometer, and the surface roughness control accuracy reaches ±0.02μm, avoiding small abnormalities from escalating into major failures. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the adaptive control method for processing parameters of multi-source heterogeneous data fusion in Embodiment 1 of the present invention. Detailed Implementation

[0047] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or apparatus. The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0048] Example 1

[0049] This embodiment provides an adaptive control method for processing parameters based on multi-source heterogeneous data fusion, including:

[0050] Step 1: Obtain multi-source processing data and perform data preprocessing.

[0051] like Figure 1 As shown, in this embodiment, a multi-dimensional sensor array is deployed to comprehensively capture multi-source machining data, including: tool status data, workpiece material spectral data, machining environment data, cutting force time-domain data, and workpiece surface roughness. Specifically, the tool status data includes the temperature and wear of the cutting edge acquired by a fiber optic grating sensor; the workpiece material spectral data includes the material hardness and density distribution detected in real time using an infrared spectrometer; the machining environment data includes the ambient temperature and humidity acquired by a temperature and humidity sensor and the cutting fluid flow rate monitored by a flow sensor; the cutting force time-domain data is acquired by a triaxial force sensor at a sampling frequency of 1 kHz; and the surface roughness of the machined workpiece is obtained by detection using a laser interferometer with a resolution of 0.01 μm.

[0052] For raw data collected from different sensors, data preprocessing is first required to transform heterogeneous, non-specific, and asynchronous data into consistent, reliable data that can be directly used by subsequent algorithms. In this embodiment, multi-source data preprocessing includes: first, using a wavelet threshold denoising algorithm to eliminate vibration interference signals; then, using Kalman filtering to complete missing data to avoid data gaps affecting analysis; and finally, using normalization to map all data to the [0,1] interval to achieve data standardization and unification.

[0053] Step 2: Input the preprocessed multi-source processing data into the pre-constructed multimodal feature fusion model, simultaneously extract local abnormal features and global correlation features during the processing, and fuse them to obtain a fusion feature vector.

[0054] Specifically, the multimodal feature fusion model adopts a four-level architecture of "input layer → dual-channel feature extraction layer → attention fusion layer → output layer". The dual-channel feature extraction layer includes a CNN branch and a Transformer branch. Its core function is to simultaneously extract the local features of the cutting force time-domain signal and the global features of the workpiece material spectral signal during the machining process, and output a 256-dimensional comprehensive feature vector through weighted fusion, which provides the core input for the subsequent GBDT quality prediction model.

[0055] In some specific embodiments, the CNN branch adopts a 3-layer structure, with each layer containing convolution and pooling. The CNN branch takes tool state data, cutting force time-domain data and workpiece surface roughness as input, and focuses on capturing local abnormal features in the cutting force signal, such as instantaneous impact force abrupt changes and periodic vibration peaks.

[0056] The Transformer branch adopts a structure of two encoder layers plus a fully connected layer. It takes workpiece material spectral data and processing environment data as input, and focuses on capturing the correlation features of different wavelength points in the spectral data, such as the correspondence between material hardness and specific wavelengths, and density distribution anomalies.

[0057] The attention fusion layer adopts a scaled dot product attention mechanism, which dynamically allocates weights based on the degree of influence of local anomaly features and global correlation features on processing quality, and obtains a fused feature vector through weighted fusion.

[0058] Specifically, to ensure that the feature vector C extracted by the CNN branch and the feature vector T extracted by the Transformer branch have completely identical dimensions, a dimension alignment verification is required before fusion calculation to avoid dimension mismatch during fusion. After dimension alignment verification, attention weights are calculated to obtain the feature weights of the CNN branch and the feature weights of the Transformer branch, and it must be ensured that the sum of the two weights is 1. Finally, a weighted fusion operation is performed to obtain the fused feature vector.

[0059] The following describes the training parameter settings for the multimodal feature fusion model in this embodiment. First, the dataset is divided into a training set:validation set:test set ratio of 7:2:1, with a sample size of no less than 5000 sets, covering processing data for various materials such as steel, aluminum, and titanium alloys. Second, the optimizer uses the Adam optimizer, with an initial learning rate of 0.001. The learning rate decay strategy is to decrease to 0.9 of the current value every 10 iterations to accommodate gradient fine-tuning in the later stages of training. Third, the loss function uses mean squared error, with a total of 100 iterations. Finally, an early stopping strategy is implemented: training stops when the validation set loss does not decrease for 5 consecutive iterations to prevent overfitting. In actual training, the model converges after 80-90 iterations, with a validation set MAPE ≤ 2.8%, meeting the requirement of "prediction error ≤ 3%".

[0060] Step 3: Input the fused feature vector into the trained machining quality prediction model to obtain the surface roughness prediction value and the dimensional error prediction value, and then obtain the comprehensive machining error.

[0061] In this embodiment, the processing quality prediction model adopts a multi-output GBDT model architecture to achieve high-precision quality prediction. By predicting surface roughness and dimensional error simultaneously with a single model, the accumulation of errors caused by independent training of multiple models is avoided.

[0062] The multi-output GBDT model uses CART regression decision trees as base learners. A 256-dimensional fused feature vector is sequentially fed into 100 trained CART decision trees to obtain the predicted values ​​for "roughness residual" and "size error residual" from each tree. The final predicted values ​​are then calculated based on the weights of each tree, outputting the surface roughness Ra and size error ΔD. The maximum depth of the decision trees is 8, the minimum number of sample splits is 20, and the learning rate is 0.05.

[0063] The dataset used to train the multi-output GBDT model covers three typical processing materials: steel, aluminum alloy, and titanium alloy. Each material corresponds to five workpiece sizes. 100 sets of "fusion features + actual quality" samples are collected under each working condition, for a total of 1,500 sets. Subsequently, the dataset is expanded to 5,000 sets by adding ±5% feature noise to simulate environmental fluctuations, ensuring that the dataset covers complex working conditions.

[0064] The dataset was partitioned using a time series method to simulate the time-series characteristics of actual production. The first 3500 groups were used as the training set, the middle 1000 as the validation set, and the last 500 as the test set, to avoid inflated model performance due to data leakage.

[0065] The actual value of surface roughness was measured using a laser interferometer with a resolution of 0.01 μm, and the actual value of dimensional error was measured using a coordinate measuring machine with an accuracy of 0.0001 mm.

[0066] The regression loss function for model training is mean squared error. Through 100 rounds of iterative training, combined with feature selection and cross-validation optimization, the prediction error is ≤3%.

[0067] The weight update method for the processing quality prediction model includes: sampling and detecting the actual value of the dimensional error using a laser interferometer, and calculating the prediction accuracy of the multi-output GBDT model based on the actual value of the dimensional error; immediately triggering the weight update of the multi-output GBDT model in response to the prediction accuracy being lower than a preset threshold; and updating the weights of the multi-output GBDT model every m control cycles using an online gradient descent algorithm.

[0068] In addition, the process of obtaining the comprehensive machining error includes: obtaining the surface roughness error based on the predicted surface roughness value and the target surface roughness value required by the workpiece design; obtaining the dimensional error allowance based on the predicted dimensional error value and the target dimensional error value; normalizing the surface roughness error and the dimensional error allowance, and then obtaining the comprehensive machining error by weighting and superimposing them.

[0069] It should be noted that the target value for dimensional error is determined as follows: The basic target value for dimensional error is obtained based on the dimensional tolerance range indicated on the workpiece design drawings; the basic target value for dimensional error is then dynamically corrected based on the deformation characteristics of the processed material to obtain the final target value for dimensional error. The specific formula is as follows:

[0070] ;

[0071] In the formula, The target value for dimensional error. This is the basic target value for dimensional error. This is the material deformation coefficient.

[0072] Step 4: Based on the comprehensive machining error, optimize the machining parameters using the parameter optimization model and generate parameter adjustment instructions.

[0073] In this embodiment, the quality and efficiency requirements in the processing are transformed into mathematical optimization objectives. Combined with actual constraints such as machine tool performance and material properties, a solvable parameter optimization model is constructed to provide the algorithm with a clear direction for optimization and parameter boundaries, and finally output the optimal parameter combination that conforms to the processing scenario.

[0074] Specifically, a parameter optimization model is constructed with the dual objectives of minimizing machining error and maximizing material removal rate. The optimization parameters include feed rate, spindle speed, depth of cut, and cutting fluid concentration. Among these, feed rate, spindle speed, and depth of cut are continuous parameters, while cutting fluid concentration is a discrete parameter.

[0075] The objective function of the parameter optimization model is a dual-objective fusion function formed by weighted fusion of the comprehensive error objective function and the efficiency objective function. The comprehensive error objective function aims to minimize the comprehensive processing error, while the efficiency objective function aims to maximize the material removal rate. The priority of the two is balanced by weights to adapt to different processing scenarios, such as prioritizing the minimum error in precision processing and prioritizing the highest efficiency in rough processing.

[0076] The formula for calculating the material removal rate in this embodiment is:

[0077] ;

[0078] In the formula, For material removal rate, For feed rate, For cutting depth, This represents the cutting width.

[0079] Constraints are key to ensuring that optimization results can be implemented. They need to cover three dimensions: machine tool performance, material processing, and quality and safety. The limitations in actual processing should be transformed into mathematical inequality constraints to avoid optimizing parameters from exceeding the actual capability range.

[0080] Specifically, the constraints of the parameter optimization model include three aspects: machine tool performance constraints, material processing constraints, and quality and safety constraints.

[0081] Firstly, there are machine tool performance constraints. The machine tool's drive system and spindle system have maximum load and performance limits, requiring optimization variables to be limited to within the machine tool's rated capacity. Specifically, this includes: feed rate constraints (the feed rate must not exceed the maximum rated speed of the machine tool's feed system); spindle power constraints (the actual cutting power must not exceed the maximum rated power of the machine tool's spindle); and spindle torque constraints (the actual cutting torque must not exceed the maximum rated torque of the machine tool's spindle).

[0082] Secondly, there are constraints related to material processing. Different materials have different physical properties, such as hardness and toughness, which require limiting cutting parameters to avoid processing defects such as built-up edge and tool breakage. Specifically, this includes cutting speed constraints and cutting force constraints.

[0083] For common aluminum alloys, the cutting speed should not exceed 300 m / min to avoid built-up edge caused by high temperature; for 45# steel, the cutting speed should not exceed 150 m / min to avoid tool wear caused by excessive cutting force; for titanium alloys, the cutting speed should not exceed 80 m / min to avoid tool adhesion due to high temperature.

[0084] Thirdly, there are quality and safety constraints. The optimization results must ensure that the processing quality is not lower than the design requirements to avoid the production of unqualified products. Specifically, this includes: surface roughness error constraints, i.e., the surface roughness predicted by the multi-output GBDT model does not exceed the surface roughness target value required by the workpiece design; dimensional error constraints, i.e., the dimensional error predicted by the multi-output GBDT model does not exceed the dimensional error target value corresponding to the workpiece design tolerance; and tool life constraints, i.e., the measured value of tool wear does not exceed the scrap threshold corresponding to the tool material to avoid tool breakage leading to equipment failure.

[0085] The algorithm for solving the parameter optimization model adopts an improved particle swarm optimization algorithm, which introduces a two-factor adaptive inertial weight based on the number of iterations and particle fitness to balance global exploration and local development. The optimal parameter combination is found within ≤50 iterations, and the material removal rate is improved by 18-22%.

[0086] It should be noted that the optimized parameter, cutting fluid concentration, in this embodiment is a discrete value. During the algorithm iteration, it is first optimized according to continuous values, such as mapping 5%, 8%, and 10% to 1, 2, and 3. After the iteration is completed, the optimization results are mapped back to the original discrete values, such as mapping the optimized value 1.2 to 8% and the optimized value 2.7 to 10%, to ensure that the variables conform to the actual values.

[0087] In some specific embodiments, a penalty function method is used to integrate constraints into the objective function of the parameter optimization model to ensure that the algorithm's search direction conforms to the constraints. Specifically, if any particle violates a constraint, a fitness penalty is added to the violating particle to correct the objective function. ;

[0088] In the formula: The objective function after penalty, For the unpenalized objective function, The penalty coefficient is... To what extent it violates the rules.

[0089] Step 5: Send the parameter adjustment command to the machine tool CNC system to drive the update of machining parameters, and based on the updated machining status, return to execute step 1 to form a closed-loop control.

[0090] In this embodiment, a fuzzy PID controller is used to achieve dynamic parameter adjustment. The fuzzy PID controller is a key execution module that connects the parameter optimization target with the machine tool execution control. Its core function is to dynamically adjust the proportional, integral, and derivative parameters of the PID controller through fuzzy inference based on the surface roughness error, dimensional error, and error change rate predicted by the multi-output GBDT model. It outputs the real-time adjustment of machining parameters such as feed rate and spindle speed, and finally sends them to the machine tool CNC system through industrial Ethernet. This realizes closed-loop control from error detection to parameter adjustment and then to quality correction. The error change rate is obtained based on the comprehensive machining error.

[0091] The closed-loop execution flow of the fuzzy PID controller is as follows:

[0092] Data acquisition: Every 10ms control cycle, receive the quality prediction error and error change rate output by the GBDT model;

[0093] Parameter tuning: The incremental adjustments of proportional, integral, and derivative parameters are calculated using fuzzy inference to update the real-time PID parameters;

[0094] Control quantity calculation: The adjustment amounts of feed rate, spindle speed, and depth of cut are calculated based on the position-based PID formula;

[0095] Command transmission: Adjustment commands are sent to the machine tool CNC system via industrial Ethernet;

[0096] Execution feedback: After the machine tool performs the adjustment, it collects new machining status data through sensors in the next control cycle, repeats the above process, and forms a closed loop.

[0097] Example 2

[0098] This embodiment provides a machining parameter adaptive control system based on multi-source heterogeneous data fusion, used to implement the machining parameter adaptive control system method as described in Embodiment 1, including:

[0099] The multi-source data acquisition module is used to acquire multi-source heterogeneous data during the machining process, including tool status data, workpiece material spectral data, machining environment data, cutting force time-domain data, and workpiece surface roughness.

[0100] The data preprocessing module is used to preprocess multi-source heterogeneous data;

[0101] The multimodal feature fusion module is used to synchronously extract local abnormal features and global correlation features during the processing based on the preprocessed data, and fuse them to obtain a fused feature vector;

[0102] The machining quality prediction module is used to obtain the predicted values ​​of surface roughness and dimensional error using the machining quality prediction model, and then obtain the comprehensive machining error.

[0103] The parameter optimization module is used to optimize the processing parameters using a parameter optimization model and generate parameter adjustment instructions;

[0104] A fuzzy PID controller is used to dynamically adjust the proportional, integral, and derivative parameters of a PID controller through fuzzy inference, and output the real-time adjustment of the optimized parameters.

[0105] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0106] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for adaptive control of processing parameters based on multi-source heterogeneous data fusion, characterized in that, include: S1: Acquire multi-source machining data and perform data preprocessing; wherein, the multi-source machining data includes: tool status data, workpiece material spectral data, machining environment data, cutting force time-domain data, and workpiece surface roughness; S2: Input the preprocessed multi-source processing data into the pre-constructed multimodal feature fusion model, simultaneously extract local abnormal features and global correlation features during the processing, and fuse them to obtain a fusion feature vector; S3: Input the fused feature vector into the trained processing quality prediction model to obtain the surface roughness prediction value and the dimensional error prediction value, and then obtain the comprehensive processing error; S4: Based on the comprehensive machining error, the machining parameters are optimized using a parameter optimization model to generate parameter adjustment instructions; S5: Send the parameter adjustment command to the machine tool CNC system to drive the machining parameters to update, and return to execute S1 based on the updated machining status to form a closed-loop control.

2. The adaptive control method for processing parameters based on multi-source heterogeneous data fusion according to claim 1, characterized in that, Also includes: The multimodal feature fusion model includes: an input layer, a dual-channel feature extraction layer, an attention fusion layer, and an output layer; The dual-channel feature extraction layer includes a CNN branch and a Transformer branch; the CNN branch takes tool state data, cutting force time-domain data and workpiece surface roughness as input to extract local anomaly features in the machining process; the Transformer branch takes workpiece material spectral data and machining environment data as input to extract global correlation features in the machining process. The attention fusion layer adopts a scaled dot product attention mechanism, which dynamically allocates weights based on the degree of influence of local anomaly features and global correlation features on processing quality, and obtains a fused feature vector through weighted fusion.

3. The adaptive control method for processing parameters based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The process of obtaining the comprehensive processing error includes: The surface roughness error is obtained based on the predicted surface roughness value and the target surface roughness value required by the workpiece design. The dimensional error margin is obtained based on the predicted dimensional error value and the target dimensional error value; The surface roughness error and dimensional error allowance are normalized, and then the comprehensive processing error is obtained by weighting and superimposing. The method for determining the target value of the size error is as follows: The basic target value of dimensional error is obtained based on the dimensional tolerance range indicated in the workpiece design drawings; The basic target value of the dimensional error is dynamically corrected based on the deformation characteristics of the processed material to obtain the target value of the dimensional error. The specific formula is as follows: ； In the formula, The target value for dimensional error. This is the basic target value for dimensional error. This is the material deformation coefficient.

4. The adaptive control method for processing parameters based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The processing quality prediction model adopts a multi-output GBDT model architecture; The weight update method of the processing quality prediction model includes: The actual value of the size error is detected by sampling with a laser interferometer, and the prediction accuracy of the multi-output GBDT model is calculated based on the actual value of the size error. In response to a prediction accuracy falling below a preset threshold, the weights of the multi-output GBDT model are immediately updated. The weights of the multi-output GBDT model are updated every m control cycles using an online gradient descent algorithm.

5. The adaptive control method for processing parameters based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The optimization parameters of the parameter optimization model include: feed rate, spindle speed, depth of cut and cutting fluid concentration. Its objective function is a dual-objective fusion function formed by weighted fusion of comprehensive error objective function and efficiency objective function. The overall error objective function is to minimize the overall processing error; the efficiency objective function is to maximize the material removal rate. The formula for calculating the material removal rate is as follows: ； In the formula, For material removal rate, For feed rate, For cutting depth, This represents the cutting width.

6. The adaptive control method for processing parameters based on multi-source heterogeneous data fusion according to claim 5, characterized in that, The constraints of the parameter optimization model include: Feed rate constraint: The feed rate shall not exceed the maximum rated speed of the machine tool feed system; Spindle power constraint: The actual cutting power shall not exceed the maximum rated power of the machine tool spindle; Spindle torque constraint: The actual cutting torque shall not exceed the maximum rated torque of the machine tool spindle; Surface roughness error constraint: The surface roughness predicted by the multi-output GBDT model shall not exceed the target surface roughness value required by the workpiece design; Dimensional error constraint: The dimensional error predicted by the multi-output GBDT model shall not exceed the target dimensional error value corresponding to the workpiece design tolerance; Tool life constraint: The measured value of tool wear shall not exceed the scrap threshold corresponding to the tool material; It also includes: cutting speed constraints and cutting force constraints set according to different processing materials.

7. The adaptive control method for processing parameters based on multi-source heterogeneous data fusion according to claim 5, characterized in that, It also includes: solving the parameter optimization model using a particle swarm optimization algorithm with adaptive inertia weight optimization; Among them, the cutting fluid concentration is a discrete variable. In the algorithm iteration, it is first optimized into a continuous value, and after the iteration is completed, the optimization result is mapped back to the original discrete value. Once the preset number of iterations is reached, the optimal parameter combination is output.

8. The adaptive control method for processing parameters based on multi-source heterogeneous data fusion according to claim 7, characterized in that, The penalty function method is used to incorporate constraints into the objective function of the parameter optimization model to ensure that the algorithm's search direction conforms to the constraints. Specifically: If any particle violates a constraint, a fitness penalty is added to the violating particle to correct the objective function, specifically: ; In the formula: The objective function after penalty, For the unpenalized objective function, The penalty coefficient is... To what extent it is violated.

9. The adaptive control method for processing parameters based on multi-source heterogeneous data fusion according to claim 5, characterized in that, Also includes: The error change rate within the control cycle is obtained based on the comprehensive processing error. The comprehensive processing error and error change rate are input into the fuzzy PID controller. The real-time optimized PID parameters are obtained through fuzzy inference and PID control, and then the adjustment amount of the optimized parameters is calculated. The adjustment amount of the optimized parameters is sent to the machine tool CNC system to drive the machine tool to perform parameter updates.

10. A machining parameter adaptive control system based on multi-source heterogeneous data fusion, used to implement the machining parameter adaptive control system method as described in any one of claims 1-9, characterized in that, include: The multi-source data acquisition module is used to acquire multi-source heterogeneous data during the machining process, including tool status data, workpiece material spectral data, machining environment data, cutting force time-domain data, and workpiece surface roughness. The data preprocessing module is used to preprocess multi-source heterogeneous data; The multimodal feature fusion module is used to synchronously extract local abnormal features and global correlation features during the processing based on the preprocessed data, and fuse them to obtain a fused feature vector; The machining quality prediction module is used to obtain the predicted values ​​of surface roughness and dimensional error using the machining quality prediction model, and then obtain the comprehensive machining error. The parameter optimization module is used to optimize the processing parameters using a parameter optimization model and generate parameter adjustment instructions; A fuzzy PID controller is used to dynamically adjust the proportional, integral, and derivative parameters of a PID controller through fuzzy inference, and output the real-time adjustment of the optimized parameters.

Images