A numerical control practical training effect evaluation method based on artificial intelligence

Abstract

The present application relates to the field of numerical control practical teaching and artificial intelligence technology, and particularly relates to a numerical control practical effect evaluation method based on artificial intelligence, which synchronously collects three types of multi-dimensional practical training data of numerical control equipment operation state, operation behavior and workpiece processing result through a multi-source sensing module and a numerical control system data interface; after the collected data are sequentially subjected to abnormal value cleaning, min-max standardization and time sequence alignment pretreatment, a multi-dimensional feature set containing operation behavior deviation features, processing effect deviation features and time sequence derived features of both is extracted; a practical effect evaluation model is constructed and trained based on a multi-scale convolution attention network, and a quantitative evaluation value of the practical effect is output, the corresponding evaluation grade is converted, and feature deviation and corresponding operation defects are output, and the model is iteratively updated in an incremental training mode. The present application realizes quantitative evaluation of the whole process of numerical control practical training, is suitable for different types of numerical control equipment and practical training scenes, and meets the precise evaluation requirements of numerical control practical training teaching.

Description

Technical Field

[0001] This invention relates to the fields of CNC training and artificial intelligence technology, specifically to an artificial intelligence-based method for evaluating the effectiveness of CNC training. Background Technology

[0002] With the rapid development of the intelligent manufacturing industry, CNC machining technology has become a core supporting technology in the machinery manufacturing field. The scale of CNC operator skills training is expanding year by year, and CNC practical training, as a core component of CNC skills training, has seen its evaluation system continuously optimized with technological advancements. Currently, the evaluation of CNC practical training effectiveness mainly suffers from the following technical deficiencies: 1. Single evaluation dimension: Existing solutions mostly collect only two types of basic data: trainee operation instructions and workpiece processing dimensions. They do not include CNC equipment operating status data, which cannot reflect the impact of the entire training operation process on the processing effect, resulting in one-sided evaluation results. 2. Insufficient assessment accuracy and large subjective bias: Most schemes use manual scoring or simple size comparison for assessment. Manual scoring is influenced by the teacher's experience and has subjective bias. Simple size comparison can only reflect the final processing result and cannot capture the timing deviation in the operation process. The assessment results have large errors and are difficult to meet the needs of precise teaching. 3. Lack of corresponding operational item location capability: The existing assessment scheme can only output the final score result, but cannot correspond to the specific operational steps and parameters that cause processing deviations, and cannot provide trainees with targeted operational references; 4. Insufficient model stability: A few evaluation schemes that introduce artificial intelligence use general machine learning models and do not design a dedicated network structure for the temporal and multi-dimensional characteristics of CNC training. They cannot adapt to the needs of different models of CNC equipment and different training scenarios, and there is no model update mechanism. After long-term use, the evaluation accuracy decreases.

[0003] To address the aforementioned shortcomings, patent application CN119494472A discloses a method for evaluating vocational education training results based on machine learning. However, this method is only applicable to general vocational education scenarios and does not have a dedicated technical solution for CNC training, thus failing to address the core pain points of CNC training evaluation. Existing CNC simulation training platforms can only record the operation process and provide simple scoring, lacking quantitative evaluation and corresponding operation item location capabilities, which cannot meet the actual needs of CNC training teaching. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide an artificial intelligence-based method for evaluating the effect of CNC training. By using multi-dimensional time-series synchronous data acquisition, multi-scale convolutional attention network evaluation model, quantitative evaluation system and incremental update mechanism, this invention solves the technical problems of single evaluation dimension, insufficient accuracy, large subjective bias, inability to locate corresponding operation items and poor model stability in the prior art, and realizes the full-process evaluation of CNC training effect.

[0005] To solve the above-mentioned technical problems, the technical solution provided by the present invention is as follows: An artificial intelligence-based method for evaluating the effectiveness of CNC training includes the following steps: S1. CNC training data acquisition: Multi-dimensional training data during the training process is acquired through the multi-source sensor module and the CNC system data interface. The multi-dimensional training data includes CNC equipment operating status data, operation behavior data and workpiece processing result data. The three types of data are acquired synchronously in time sequence. S2. Data preprocessing: Cleaning, standardizing, and time-series alignment of multi-dimensional training data; S3. Feature Extraction: Extract a multi-dimensional feature set from the preprocessed data. The multi-dimensional feature set includes operational behavior deviation features, processing effect deviation features, and time-series derived features of both. S4. Evaluation Model Construction and Training: A training effect evaluation model is constructed based on a multi-scale convolutional attention network. The multi-dimensional feature set is input into the training effect evaluation model, and the mean squared error loss function and Adam optimization algorithm are used to train the trained training effect evaluation model. S5. Training effect evaluation: After data preprocessing and feature extraction, the training data to be evaluated is input into the trained training effect evaluation model, and the quantitative evaluation value of the training effect is output. S6. Evaluation Result Output: Convert the quantitative evaluation value of the training effect into the training effect evaluation level, and output the data of each feature dimension, key operational defects, and corresponding processing parameters. S7. Model Iteration and Update: Regularly collect new training data and corresponding manual evaluation results to form a new training sample set. Use incremental training to iteratively update the training effect evaluation model. Verify the accuracy of the training effect evaluation model through the test set. If the accuracy is not up to standard, retrain the training effect evaluation model.

[0006] Furthermore, the multi-dimensional training data specifically includes: CNC equipment operation behavior data, workpiece machining result data, and CNC equipment operating status data; CNC equipment operation behavior data includes operation command timing data, feed rate adjustment timing data, spindle speed adjustment timing data, and tool path control data; workpiece machining result data includes workpiece dimensional deviation data, surface roughness data, and machining allowance data; CNC equipment operating status data includes equipment vibration data, cutting force data, and spindle temperature data.

[0007] Furthermore, in step S1, the multi-source sensing module includes a vibration sensor, a force sensor, and a temperature sensor. The CNC system data interface establishes a communication connection with the CNC training equipment control system through an industrial Ethernet. The timing synchronization accuracy of the acquired data is consistent with the clock reference of the CNC system, and the data acquisition time granularity matches the interpolation cycle of the CNC equipment.

[0008] Furthermore, in step S2, data preprocessing includes the following steps: S21. Outlier handling: Outliers are identified using the Grubbs criterion, and linear interpolation is used to complete the identified outliers. ; in, For the first Each data sample value, The mean of the data sample. This is the Grubbs critical value. At the significance level, For the number of data samples, The standard deviation of the data sample; S22. Data Standardization: The min-max standardization method is used to map the data after outlier processing to the 0-1 range; S23. Timing Alignment: Using the start time of the CNC training operation as the time reference, align the timing data from different acquisition sources according to the timestamp. After alignment, the time step of the timing data is consistent with the granularity of the data acquisition time.

[0009] Furthermore, in step S3, during feature extraction, the operational behavior deviation feature and the processing effect deviation feature are calculated using a preset quantization formula, and the time-series derived features are the first-order difference and second-order difference of the feature values ​​in the time dimension. ; in, This represents the characteristic value of operational behavior deviation. This represents the total number of time steps for the practical training operation. For the first Actual values ​​of the operation parameters for the time step. For the first Standard values ​​for the operating parameters of the time step; The calculation formula for extracting processing effect features is as follows: ; in, This represents the characteristic value of the processing effect deviation. The number of dimensions to be inspected on the workpiece. For the first The actual measured values ​​of each detection dimension For the first Standard values ​​for each detection dimension; the multi-dimensional feature set consists of... and It consists of the temporal derivative features of both, which are calculated by the adjacent time step difference of the feature values.

[0010] Further, in step S4, the multi-scale convolutional attention network model includes an input layer, a multi-scale convolutional layer, a channel attention layer, a temporal attention layer, a fully connected layer, and an output layer; the multi-scale convolutional layer uses three different odd-sized convolutional kernels, the channel attention layer fuses channel-dimensional features by calculating feature channel weight coefficients, the temporal attention layer assigns attention weights to temporal features, and the fully connected layer maps the fused features to the evaluation value space; ; in, For the first The weight coefficients of each feature channel, It is the Sigmoid activation function. This is a global average pooling operation. For the first Feature maps of each feature channel This represents the total number of feature channels; the output layer outputs a quantitative evaluation value of the training effect.

[0011] Furthermore, in step S4, the training effect evaluation model uses the mean squared error loss function to calculate the loss value, and uses the Adam optimization algorithm to update the training effect evaluation model parameters. The learning rate is adaptively adjusted according to the exponential decay law with each training round. ; in, The loss value. The number of training samples. For the first The actual evaluation value of each training sample. For the first The training effect evaluation model predicts the evaluation value of each training sample; the learning rate is adaptively adjusted with each training round, following an exponential decay law, and the decay formula is: ; in, Let be the learning rate for the t-th training round. The initial learning rate, t is the decay coefficient, and t is the training round.

[0012] Furthermore, in step S5, the quantitative evaluation value of the training effect is calculated by weighting the operational behavior characteristics, processing effect characteristics, and time-series derived characteristics according to preset weights, and the value range is 0 to 1; ; in, To quantify the evaluation value of the practical training effect, These are the weighting coefficients for operational behavior characteristics. The weighting coefficients for processing effect characteristics. Let be the weight coefficients of the time-series derived features, and satisfy . , The composite value of the time-series derived features is calculated from the mean of the first-order and second-order differences.

[0013] Furthermore, in step S6, the evaluation level includes excellent, good, qualified, and unqualified, which is determined by the interval division of the quantitative evaluation value based on the training effect; the output feature dimension data includes the actual value of each feature and the percentage of deviation from the standard value, and the key operational defects and corresponding processing parameters are determined by sorting through temporal attention weights. ; in, The percentage of deviation in the feature dimension. The actual value of the feature dimension. The standard values ​​for the feature dimensions are used; key operation steps and processing parameters are determined by sorting the attention weights output by the temporal attention layer in descending order. Operation steps and processing parameters with attention weight values ​​greater than a preset weight threshold after sorting are selected as key items. The preset weight threshold is the average of all attention weights, i.e. ,in, To preset the weight threshold, For the first Temporal attention weights for time steps This represents the total number of attention weights.

[0014] Furthermore, in step S7, the batch size of incremental training is consistent with the batch size of initial training, and the preset threshold is determined by combining the verification accuracy of the initial training of the training effect evaluation model with the error allowable coefficient; the verification accuracy of the training effect evaluation model and the preset threshold are calculated by the corresponding quantization formulas respectively. When the verification accuracy of the training effect evaluation model is lower than the preset threshold, the training effect evaluation model can be retrained after adjusting the combination of the multi-scale convolutional layer kernel size and the timing attention layer weight calculation logic. ; in, To evaluate the effectiveness of the practical training, the model's accuracy was verified. To predict the number of correctly predicted samples in the test set, The total number of samples in the test set; the formula for calculating the preset threshold is: ,in, For the preset threshold, To evaluate the validation accuracy of the initial training of the model in order to assess the effectiveness of the practical training. This is the tolerance coefficient for error.

[0015] The advantages of this invention compared to the prior art are: 1. This invention integrates three types of multi-dimensional data—CNC equipment operation behavior, workpiece processing results, and equipment operating status—through time-synchronous acquisition of data from multi-source sensing modules and CNC system data interfaces. This achieves complete coverage of data throughout the entire CNC training process. Time-synchronous acquisition ensures accurate time correspondence between data from different sources, solving the problem of one-sided evaluation results caused by single evaluation dimensions and misaligned data timing.

[0016] 2. This invention constructs a training effect evaluation model through a multi-scale convolutional attention network. It combines multi-scale convolution, channel attention, and time-series attention to process multi-dimensional temporal features, thereby achieving quantitative evaluation of training effects. This eliminates the subjective bias of manual scoring. The multi-scale convolutional structure can extract feature information from different time dimensions, improving the accuracy of the evaluation results.

[0017] 3. This invention uses deviation ratio calculation and temporal attention weight ranking to realize the corresponding positioning of operation steps and processing parameters that affect the evaluation results, and maps the final evaluation results to specific operation links, providing clear operation improvement references for practical training and teaching, and solving the problem that the existing technology can only output scoring results and cannot locate operation defects.

[0018] 4. This invention iteratively updates the evaluation model through incremental training, incorporating new training data while retaining the model's original evaluation capabilities, thus maintaining the model's evaluation stability for long-term use. The accuracy verification and retraining mechanism ensures that the model's evaluation accuracy remains within an usable range, and it can adapt to parameter differences in different CNC equipment and training scenarios. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0020] In the attached diagram: Figure 1 This is a flowchart of an artificial intelligence-based numerical control training effect evaluation method in Example 1.

[0021] Figure 2 This is a flowchart of data preprocessing in an AI-based numerical control training effect evaluation method in Example 1. Detailed Implementation

[0022] The detailed description of the following embodiments is used to illustrate the principles of this application, but should not be used to limit the scope of this application. That is, the artificial intelligence-based CNC training effect evaluation method of this application is not limited to the described embodiments.

[0023] The present invention will be further described below with reference to embodiments.

[0024] This specific implementation method uses the "CNC Lathe Machining Training" course for the CNC Technology major in the School of Mechanical Engineering as the application scenario. The training task is to machine the outer diameter, end face, and steps of a 45# high-quality carbon structural steel shaft part. The CNC machining equipment used is a Shenyang Machine Tool CAK6150 CNC horizontal lathe, equipped with a Siemens SINUMERIK 828D CNC system, supporting Profinet industrial Ethernet communication. The system interpolation cycle is 1ms, which is the default parameter of the equipment. The multi-source sensing module is equipped with three types of industrial-grade sensors. The vibration sensor is a PCB Piezotronics 352C22 IEPE vibration accelerometer, with a maximum sampling frequency of 10kHz. In this embodiment, the sampling frequency is set to 1000Hz. It is installed at the front end of the lathe spindle box to collect vibration data in the X and Z axes. The force sensor is an HBM U9B tensile and compressive force sensor with a range of 0-5kN. In this embodiment, the sampling frequency is set to 1000Hz. It is installed at the bottom of the lathe turret to collect cutting force data in the X and Z axes. The temperature sensor is an industrial-grade PT100 platinum resistance contact temperature sensor. In this embodiment, the sampling frequency is set to 100Hz. It is installed at the spindle bearing end cover to collect spindle operating temperature data.

[0025] The data acquisition and computing hardware includes an industrial edge computer and a model training workstation. The industrial edge computer is the Advantech UNO-2484G industrial edge computing device, mass-produced by Advantech, equipped with an Intel Celeron J3455 processor, 8GB DDR3L memory, a 128GB SSD, and pre-installed with the Ubuntu 20.04 LTS operating system. It can establish communication with the CNC system via the Profinet protocol and is equipped with an 8-channel analog input module to complete the acquisition of sensor data. The model training workstation is a Dell Precision 5820 tower workstation, equipped with an Intel Xeon W-2245 processor, 32GB DDR4 memory, an NVIDIA RTX A4000 professional graphics card, a 1TB NVMe SSD, and pre-installed with the Windows 10 Professional operating system. It is used to complete the training, inference, and result output of the model. The workpiece inspection equipment is equipped with two types of industrial-grade measuring instruments: a Zeiss CONTURA G2 coordinate measuring machine is used for dimensional inspection to detect workpiece dimensional deviations; and a Tokyo Seimitsu SURFCOM 130A surface roughness meter is used for surface roughness inspection to detect workpiece surface roughness.

[0026] like Figure 1As shown, step S1 involves CNC training data acquisition. The training task in this embodiment is the machining of shaft-type parts from 45# steel bars. The machining includes face turning, rough turning of the outer diameter, finish turning of the outer diameter, step turning, grooving, and chamfering. The total machining time is 120 seconds, corresponding to a total of 120,000 acquisition steps, with an acquisition cycle of 1ms. Using the PLC system clock of the Siemens SINUMERIK 828D CNC system as the global synchronization reference, the system clock of the industrial edge computer is synchronized with the CNC system clock via the NTP protocol, with a synchronization error ≤100μs. The timestamps of all sensor data and CNC system data acquisition are based on this synchronized clock, with an acquisition time granularity of 1ms, perfectly matching the CNC system interpolation cycle. This ensures accurate time correspondence between the three types of data: operational behavior, equipment operating status, and machining results, avoiding deviations in feature extraction due to time misalignment. Real-time operational data is acquired from the CNC system via the Profinet protocol, including NC code operation command timing data, X / Z axis feed rate adjustment timing data, spindle speed adjustment timing data, and tool path coordinate data at the tool tip point, with a sampling frequency of 1000Hz. Equipment operating status data is acquired through the analog input module of the edge computer, including X / Z axis vibration acceleration data, X / Z axis cutting force data, and spindle temperature data, with vibration and cutting force sampling frequencies of 1000Hz and temperature sampling frequency of 100Hz. After the training, workpiece machining result data is obtained by measuring the coordinate measuring machine and surface roughness tester, including external diameter deviation, step length deviation, end face runout deviation, cylindricity deviation, surface roughness deviation, and radial machining allowance deviation.

[0027] like Figure 2 As shown, in step S2, data preprocessing, the collected time-series data is preprocessed. First, Grubbs' criterion is used to identify outliers, and a significance level is set. Sample size By consulting the Grubbs' critical value table in the national standard GB / T4883-2008 "Statistical Processing and Interpretation of Data - Judgment and Handling of Outliers in Normal Samples", we can obtain... Through formula Outliers are identified and imputed using linear interpolation to remove invalid data generated during data acquisition due to sensor electromagnetic interference and system communication fluctuations, thus preventing outliers from interfering with feature extraction and model evaluation. A min-max normalization method is employed to normalize all dimensions of data, mapping them to the 0-1 interval to eliminate dimensional differences and ensure that each feature contributes equally to the model training process. Using the timestamp of the M03 spindle forward rotation command output by the CNC system as time reference 0, all sensor and CNC system data are aligned by timestamp. After alignment, the time step of all data is unified to 1ms. Missing time steps are imputed using forward padding, resulting in a standardized time-series dataset. This ensures complete temporal correspondence between data from different acquisition sources, providing a unified time reference for time-series feature extraction.

[0028] Step S3, Feature Extraction: Extract a multi-dimensional feature set from the preprocessed time-series data. First, divide the total processing time into 6 time intervals corresponding to the processing steps, with 20,000 steps in each interval. Calculate the operational behavior deviation feature value for each interval, and then take the average to obtain the overall operational behavior deviation feature value. The calculation formula is: ,in This represents the total number of time steps for the practical training operation. For the first Actual values ​​of the operation parameters for the time step. For the first The standard values ​​of the operating parameters for each time step quantify the overall deviation between the trainee's actual operation and the standard process requirements, reflecting the standardization of the trainee's operation. This is achieved through formulas. Calculate the characteristic value of the processing effect deviation ,in The number of dimensions to be inspected on the workpiece. For the first The actual measured values ​​of each detection dimension For the first Standard values ​​for each detection dimension quantify the deviation between the final processed quality of the workpiece and the requirements of the drawing, reflecting the final result of the practical training operation. Calculate the characteristic values ​​of operational behavior deviations. Processing effect deviation characteristic value For each process time interval, the first and second differences are used as the average of the first and second differences to obtain the comprehensive value of the time-series derived features. This captures the changing trends of operational behaviors and processing effects over time, reflecting the stability of trainees' operations. The operational behavior deviation characteristic values ​​are then analyzed. Processing effect deviation characteristic value Time-series derived feature composite value The feature set is obtained by combining the feature values ​​and time-series difference features of each process interval, and serves as the input to the evaluation model.

[0029] Step S4: Evaluate model construction and training. A multi-scale convolutional attention network model is constructed based on the TensorFlow 2.10.0 deep learning framework. The model structure consists of an input layer, a multi-scale convolutional layer, a concatenation layer, a channel attention layer, a temporal attention layer, a global average pooling layer, a fully connected layer, and an output layer. The input layer receives a multi-dimensional feature set. The multi-scale convolutional layer has three parallel convolutional branches with kernel sizes of 3×1, 5×1, and 7×1, each with 16 output channels. Padding is set to "same," and the activation function is ReLU. Temporal features at different time scales are extracted, covering three different time dimensions: short-term operation details, medium-term process features, and long-term processing trends. The concatenation layer concatenates the output features of the three convolutional branches along the channel dimension to obtain fused features. The channel attention layer performs global average pooling on the concatenated features, outputs channel weights through two fully connected layers, normalizes them using the Sigmoid function, and multiplies them with the original features to achieve channel-dimensional feature recalibration. The weight calculation formula is... ,in For the first The weight coefficients of each feature channel, It is the Sigmoid activation function. This is a global average pooling operation. For the first Feature maps of each feature channel The total number of feature channels. The temporal attention layer calculates the attention weight for each time step of the channel attention output features. After normalization, the weights are multiplied by the original features. Weights are assigned to features of different feature channels and time steps to enhance the extraction of feature information with high correlation to the evaluation results. The global average pooling layer compresses the temporal features into a one-dimensional feature vector. The fully connected layer has two fully connected layers with hidden layer dimensions of 24 and 8 respectively, using ReLU activation function. The output layer has one neuron with Sigmoid activation function, outputting a quantitative evaluation value of the training effect in the range of 0-1. Data from 1200 groups of trainees who have completed training were collected. Each group of data includes corresponding multi-dimensional training data and manual evaluation values ​​jointly given by three training teachers with intermediate or higher professional titles. The data were divided into a training set of 960 groups and a test set of 240 groups in an 8:2 ratio to construct the training sample set.

[0030] Model training uses the mean squared error loss function to calculate the loss value. The formula for the loss function is as follows: ,in The loss value. The number of training samples. For the first The actual evaluation value of each training sample. For the first The model prediction evaluation values ​​for each training sample. The model parameters are updated using the Adam optimizer, with the parameters set to... , , Initial learning rate attenuation coefficient After each round of training, follow the formula Update the learning rate, batch size = 32, epochs = 200, and terminate training when the test set loss value does not decrease for 10 consecutive epochs or when the training epochs reach 200. After training, save the model for evaluation of training effectiveness.

[0031] Step S5, training effectiveness evaluation: The training data of the trainees to be evaluated, after being processed in steps S2 and S3, is input into the trained evaluation model, which outputs basic evaluation values, and then uses a weighted formula. Final calibration is performed to obtain a quantitative evaluation value of the training effect. ,in These are the weighting coefficients for operational behavior characteristics. The weighting coefficients for processing effect characteristics. Let be the weight coefficients of the time-series derived features, and satisfy . , This is the comprehensive value of time-series derived features. The weighted summation calculation method can adjust the weights of different features according to the focus of practical training, corresponding to the evaluation needs of different practical training tasks.

[0032] Step S6: Output the evaluation results, quantifying the evaluation values ​​according to the preset range. Convert to evaluation grades, which include Excellent, Good, Satisfactory, and Unsatisfactory, with a range set as follows: For excellence, For good, ) is qualified. This is considered unqualified. Calculate the deviation percentage for each feature dimension. The formula for calculating the deviation percentage is: ,in The percentage of deviation in the feature dimension. The actual value of the feature dimension. For each feature dimension, output the actual value, standard value, and deviation percentage. Extract the attention weights for each time step from the temporal attention layer output, sort them in descending order of weight value, and calculate the mean of all attention weights as the preset weight threshold. The calculation formula is: ,in To preset the weight threshold, For the first Temporal attention weights for time steps The total number of attention weights is used to select weights greater than 1. The time step is mapped to specific processing procedures and operating parameters, and output as key operation items. The final evaluation results are then mapped to specific operational steps, providing clear operational improvement references for practical training.

[0033] Step S7: Model Iteration and Update. A fixed iteration cycle is set at the end of each semester, with one model iteration and update at the end of each cycle. Collect 300 new sets of training data added during the cycle, along with corresponding manual evaluation values ​​provided by three training instructors. After processing in steps S2 and S3, this forms a new incremental training sample set. Load the previously trained model, freeze the parameters of the multi-scale convolutional layers, and train only the parameters of the channel attention layer, temporal attention layer, and fully connected layer. The batch size is set to 16, and the epoch count is 50. All other parameters remain the same as the initial training. This process incorporates new training data while preserving the model's original evaluation capabilities, avoiding computational overhead and accuracy fluctuations caused by retraining. After iteration, a fixed test set of 240 sets is used to verify the model's accuracy. The accuracy calculation formula is: ,in To verify the accuracy of the model, To predict the number of correctly predicted samples in the test set, This represents the total number of samples in the test set. (Preset threshold) The error tolerance coefficient is determined by combining the validation accuracy of the initial model training with the error allowable coefficient. The calculation formula is as follows: ,in For the preset threshold, This is to verify the accuracy of the initial training of the model. This is the tolerance coefficient for error. If verifying accuracy... If so, then save the iterated model as the new evaluation model; if Then, all network layers are unfrozen, the convolution kernel size and the weight calculation logic of the timing attention layer are adjusted, and the complete model training process is re-executed to keep the model's evaluation accuracy within an usable range.

[0034] It should be noted that the combination of the technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0035] It should also be noted that the embodiments listed above are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and similar changes or modifications made thereto are those that can be directly derived or easily conceived by those skilled in the art from the content disclosed in the present invention, and should all fall within the protection scope of the present invention.

Claims

1. A method for evaluating the effectiveness of CNC training based on artificial intelligence, characterized in that, Includes the following steps: S1. CNC training data acquisition: Multi-dimensional training data during the training process is acquired through the multi-source sensing module and the CNC system data interface. The multi-dimensional training data includes CNC equipment operating status data, operation behavior data and workpiece processing result data. The three types of data are acquired synchronously in time sequence. S2. Data preprocessing: Cleaning, standardizing, and time-series alignment of the multi-dimensional training data; S3. Feature extraction: Extract a multi-dimensional feature set from the preprocessed data. The multi-dimensional feature set includes operational behavior deviation features, processing effect deviation features, and time-series derived features of both. S4. Evaluation Model Construction and Training: A training effect evaluation model is constructed based on a multi-scale convolutional attention network. The multi-dimensional feature set is input into the training effect evaluation model, and the mean squared error loss function and Adam optimization algorithm are used to train the trained training effect evaluation model. S5. Training effect evaluation: After the training data to be evaluated has undergone the data preprocessing and feature extraction processing, it is input into the trained training effect evaluation model, and the quantitative evaluation value of the training effect is output. S6. Evaluation Result Output: Convert the quantitative evaluation value of the training effect into a training effect evaluation level, and output the data of each feature dimension, key operational defects, and corresponding processing parameters. S7. Model Iteration Update: Regularly collect new training data and corresponding manual evaluation results to form a new training sample set. Use incremental training to iteratively update the training effect evaluation model. Verify the accuracy of the training effect evaluation model through the test set. If the accuracy does not meet the standard, retrain the training effect evaluation model.

2. The method for evaluating the effectiveness of CNC training based on artificial intelligence according to claim 1, characterized in that, The multi-dimensional training data specifically includes: CNC equipment operation behavior data, workpiece machining result data, and CNC equipment operating status data; the CNC equipment operation behavior data includes operation command timing data, feed rate adjustment timing data, spindle speed adjustment timing data, and tool path control data; the workpiece machining result data includes workpiece dimensional deviation data, surface roughness data, and machining allowance data; the CNC equipment operating status data includes equipment vibration data, cutting force data, and spindle temperature data.

3. The method for evaluating the effectiveness of CNC training based on artificial intelligence according to claim 2, characterized in that, In step S1, the multi-source sensing module includes a vibration sensor, a force sensor, and a temperature sensor. The CNC system data interface establishes a communication connection with the CNC training equipment control system through an industrial Ethernet. The timing synchronization accuracy of the acquired data is consistent with the CNC system clock reference, and the data acquisition time granularity matches the CNC equipment interpolation cycle.

4. The method for evaluating the effectiveness of CNC training based on artificial intelligence according to claim 3, characterized in that, In step S2, the data preprocessing includes the following steps: S21. Outlier handling: Outliers are identified using the Grubbs criterion, and linear interpolation is used to complete the identified outliers. ； in, For the first Each data sample value, The mean of the data sample is... This is the Grubbs critical value. At the significance level, The number of data samples, The standard deviation of the data sample; S22. Data standardization: Using the min-max standardization method, the data after outlier processing is mapped to the 0-1 range; S23. Timing Alignment: Using the start time of the CNC training operation as the time reference, align the timing data from different acquisition sources according to the timestamp. After alignment, the time step of the timing data is consistent with the granularity of the data acquisition time.

5. The method for evaluating the effectiveness of CNC training based on artificial intelligence according to claim 4, characterized in that, In step S3, during feature extraction, the operational behavior deviation feature and the processing effect deviation feature are calculated using a preset quantization formula, and the time-series derived feature is the first-order difference and second-order difference of the feature value in the time dimension. ； in, The characteristic value of the operational behavior deviation. This represents the total number of time steps for the practical training operation. For the first Actual values ​​of the operation parameters for the time step. For the first Standard values ​​for the operating parameters of the time step; The calculation formula for the extraction of processing effect features is as follows: ； in, The characteristic value of the processing effect deviation is... The number of dimensions to be inspected on the workpiece. For the first The actual measured values ​​of each detection dimension For the first Standard values ​​for each detection dimension; the multi-dimensional feature set consists of... and The time-series derived features of both are composed of the time-series derived features, which are calculated by the adjacent time step difference values ​​of the feature values.

6. The method for evaluating the effectiveness of CNC training based on artificial intelligence according to claim 5, characterized in that, In step S4, the multi-scale convolutional attention network model includes an input layer, a multi-scale convolutional layer, a channel attention layer, a temporal attention layer, a fully connected layer, and an output layer. The multi-scale convolutional layer uses three different odd-sized convolutional kernels. The channel attention layer fuses channel-dimensional features by calculating feature channel weight coefficients. The temporal attention layer assigns attention weights to temporal features. The fully connected layer maps the fused features to the evaluation value space. ； in, For the first The weight coefficients of each feature channel, It is the Sigmoid activation function. This is a global average pooling operation. For the first Feature maps of each feature channel The total number of feature channels; the output layer outputs the quantitative evaluation value of the training effect.

7. The method for evaluating the effectiveness of CNC training based on artificial intelligence according to claim 6, characterized in that, In step S4, the training effect evaluation model is trained by calculating the loss value using the mean square error loss function, and the Adam optimization algorithm is used to update the parameters of the training effect evaluation model. The learning rate is adaptively adjusted according to the exponential decay law with each training round. ； in, The loss value is... The number of training samples. For the first The actual evaluation value of each of the training samples. For the first The training effect evaluation model predicts the evaluation value of each training sample; the learning rate is adaptively adjusted with each training round, following an exponential decay law, and the decay formula is: ； in, Let be the learning rate for the training in round t. The initial learning rate, t is the decay coefficient, and t is the training round.

8. The method for evaluating the effectiveness of CNC training based on artificial intelligence according to claim 7, characterized in that, In step S5, the quantitative evaluation value of the training effect is calculated by weighting the operational behavior characteristics, the processing effect characteristics, and the time-series derived characteristics according to a preset weight, and the value range is 0 to 1; ； in, The evaluation value for the training effect is quantified. These are the weighting coefficients for the operational behavior characteristics. The weighting coefficients for the processing effect features are as follows: The weight coefficients of the time-series derived features are, and satisfy the following conditions: , The comprehensive value of the time-series derived features is calculated from the mean of the first-order difference and the second-order difference.

9. The method for evaluating the effectiveness of CNC training based on artificial intelligence according to claim 8, characterized in that, In step S6, the evaluation level includes excellent, good, qualified, and unqualified, which is determined by the interval division of the quantitative evaluation value based on the training effect; the output feature dimension data includes the actual value of each feature and the percentage of deviation from the standard value, and the key operational defects and corresponding processing parameters are determined by the temporal attention weight sorting. ； in, The deviation percentage of the aforementioned feature dimension. The actual value of the feature dimension. The standard value of the feature dimension is used; the key operation steps and the processing parameters are determined by sorting the attention weights output by the temporal attention layer in descending order, and the operation steps and processing parameters whose attention weight values ​​are greater than a preset weight threshold after sorting are selected as key items, where the preset weight threshold is the average of all attention weights, i.e. ,in, The preset weight threshold, For the first The temporal attention weights mentioned in the time step, This represents the total number of attention weights.

10. The method for evaluating the effectiveness of CNC training based on artificial intelligence according to claim 9, characterized in that, In step S7, the batch size of the incremental training is the same as the initial training batch size. The preset threshold is determined by combining the verification accuracy of the initial training of the training effect evaluation model with the error allowable coefficient. The verification accuracy of the training effect evaluation model and the preset threshold are calculated by the corresponding quantization formulas. When the verification accuracy of the training effect evaluation model is lower than the preset threshold, the training effect evaluation model can be retrained after adjusting the combination of the convolution kernel size of the multi-scale convolutional layer and the weight calculation logic of the temporal attention layer. ； in, To verify the accuracy of the training effect evaluation model, To predict the number of correctly predicted samples in the test set, The total number of samples in the test set; the formula for calculating the preset threshold is: ,in, The preset threshold, The verification accuracy of the initial training of the model is used to evaluate the training effect. This is the error tolerance coefficient.

Images