A dynamic prediction method for cutting tool wear based on multi-model fusion

Abstract

The present invention relates to a method for dynamically predicting cutting tool wear based on multi-model fusion, which belongs to the field of lean manufacturing technology. It solves the technical problems of poor accuracy and poor practicality in current prediction of cutting tool wear. The method of the present invention includes selecting a method for collecting machine tool working condition data during the cutting process, conducting a tool wear processing test and collecting tool wear degradation data and corresponding cutting process working condition data, pre-processing the wear degradation data, establishing a wear increment prediction model, establishing a wear amount prediction model, constructing a hybrid prediction model and predicting tool wear. The method of the present invention uses a hybrid prediction model based on cutting process working condition data to derive degradation increments in multiple time periods, and predicts the degree of degradation based on the accumulation of degradation increments, thereby achieving dynamic prediction of tool wear for cutting tools in the future, and the model is closer to the actual situation. The implementation of the method of the present invention does not require the addition of external sensors or additional wiring, and is more economical, practical and feasible.

Description

Technical Field

[0001] The present invention relates to the technical field of lean manufacturing, and in particular to a multi-model fusion method for dynamically predicting cutting tool wear. Background Art

[0002] As a crucial piece of equipment directly involved in cutting operations, the wear state of cutting tools has a direct impact on the quality of workpiece machining. Predicting their wear state plays a crucial role in ensuring machining quality. Currently, there are two main approaches to predicting tool wear state. One involves direct manual measurement of the tool after machining, using the acquired tool wear to predict future tool wear. This approach is only valuable when the cutting process is completely stable, ignoring the impact of abnormal conditions on tool wear during complex cutting processes. Another approach involves online acquisition of dynamic operating parameters during the cutting process and using artificial intelligence algorithms to predict the current tool wear state. These approaches often rely on a "signal analysis + feature extraction" approach followed by model building. These approaches typically use time-domain features, which increases the impact of machining parameter changes on model predictions and limits the model's generalization capabilities. Furthermore, selecting effective features is a challenging process. Furthermore, current predictions of tool wear degradation typically use real-time signals to fit the current degradation level. This approach results in non-monotonic tool wear degradation curves that are inconsistent with the performance degradation process of mechanical products. Furthermore, these prediction methods ignore the time effect, resulting in suboptimal prediction results. In actual production, predicting tool wear over a specific period of time is more practical and practicable than predicting long-term tool wear.

[0003] Therefore, the present invention considers predicting the degradation increment in each time period through the working condition data of the cutting process, and predicting the degree of degradation based on the accumulation of the degradation increment. A multi-model fusion dynamic prediction method for cutting tool wear is proposed. By combining the convolutional neural network and the BP neural network, and integrating the multi-source dynamic data information obtained by the machine tool CNC system with a small amount of manually measured tool wear data, the dynamic prediction of the tool wear of the cutting tool in the future period of time is realized. Summary of the Invention

[0004] In view of the above analysis, the embodiment of the present invention aims to provide a dynamic prediction method for cutting tool wear using multi-model fusion to solve the current technical problems of poor cutting tool wear prediction accuracy and poor practicality.

[0005] A multi-model fusion method for dynamic prediction of cutting tool wear includes the following steps:

[0006] S1. Determine the cutting process machine tool working condition data to be collected and determine the corresponding data collection method;

[0007] S2. Conduct tool wear processing test:

[0008] Design cutting tests; perform cutting operations; collect tool wear and degradation data during the cutting process, as well as collect machine tool operating condition data during the same cutting process;

[0009] S3, preprocessing the tool wear degradation data obtained in S2, extracting the tool wear degradation data during the stable wear period in S2 and the corresponding machine tool working condition data; establishing multiple data sample sets based on the time series; and generating a tool wear variation curve;

[0010] S4. Establish a wear increment prediction model:

[0011] Use the data sample set to train the convolutional neural network model and build a convolutional neural network wear increment prediction model;

[0012] S5. Establish tool wear prediction model:

[0013] According to the data samples in S3, a tool wear prediction model based on BP neural network is established;

[0014] S6. Establish a hybrid prediction model:

[0015] The wear increment prediction model of S4 and the tool wear prediction model of S5 are combined under the constraints of the prediction result matching test criteria to form a hybrid prediction model; it is used to predict the tool wear of the tool under the same process conditions in the future.

[0016] Furthermore, the cutting process machine tool working condition data related to the tool wear process in S1 is data related to the cutting force, including spindle load, spindle torque and spindle power.

[0017] Furthermore, the data acquisition mode in S1 includes one or any combination of the serial port mode of the RS-232 protocol, the Ethernet mode of the TCP / IP protocol, and the data acquisition card mode.

[0018] Furthermore, the process method in the key step of S2 includes the following steps:

[0019] S21, selecting a key cutting step for which tool wear status monitoring is required;

[0020] S22, selecting process parameters for key cutting steps;

[0021] S23. Perform a cutting test to obtain tool wear degradation data during the cutting process and cutting process working condition data during the same period of time.

[0022] Furthermore, the “preprocessing” described in S3 refers to eliminating abnormal data in the cutting process machine tool working condition data obtained in S2, filling the abnormal data with the mean value of normal data adjacent to the abnormal data, and smoothing the measurement data.

[0023] Furthermore, the data sample set described in S4 is a data sample set established by machine tool operating data collected during the stable wear period of the cutting test.

[0024] Furthermore, the "wear increment prediction model" in S4 is a performance degradation increment prediction model constructed using a convolutional neural network method and using preprocessed data signals as input. The steps of establishing the wear increment prediction model include:

[0025] S41, mapping the working condition data input within a certain period of time into a degradation increment;

[0026] S42. Add BN normalization layers to each layer of the model;

[0027] S43, downsampling the output result after the first pooling layer and adding it to the output result after the second pooling layer;

[0028] S44, repeat S43 until the model converges;

[0029] S45, the output is connected to two fully connected layers;

[0030] S46, output degradation increment S91. Use convolutional neural network to construct performance degradation increment prediction model for preprocessed data signal;

[0031] Furthermore, the tool wear prediction model in S5 is a regression model constructed through a BP neural network; the input parameters of the regression model include the tool wear degradation amounts at multiple previous moments, the cutting time from the previous moment to the current moment, and the degradation increment in the current cutting time period obtained by the wear increment prediction model; the regression model is used to predict the tool wear degradation amount at the next moment.

[0032] Furthermore, the "hybrid prediction model" described in S6 refers to the integration of the wear increment prediction model of the convolutional neural network in S4 with the tool wear amount prediction model based on the BP neural network to construct a hybrid prediction model. S91. Use the convolutional neural network to construct a performance degradation increment prediction model for the preprocessed data signal;

[0033] S92. Obtaining previous degradation data, including tool wear degradation amounts at multiple previous moments, cutting time from the previous moment to the current moment, and degradation increment tool wear degradation amount in the current cutting time period obtained by the wear increment prediction model;

[0034] S93. Through the convolutional neural network model, a regression model is constructed to output the tool wear degradation amount at the prediction moment.

[0035] Furthermore, the “prediction result matching test criterion” described in S6 means that the degradation amount calculated by the trained hybrid prediction model complies with the following formula:

[0036]

[0037] in,

[0038] y(t)-the predicted value of tool wear degradation at the current moment t;

[0039] y(t-1) - the predicted value of tool wear degradation at the previous moment t-1 or the actual tool wear degradation value obtained by manual measurement because the predicted result does not match the actual tool wear degradation value;

[0040] x(t) - the degradation increment during this period predicted using the operating data from time (t-1) to time (t);

[0041] q-degradation increment deviation threshold; where 0 <q<1。

[0042] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0043] 1. The multi-model fusion dynamic prediction method for cutting tool wear proposed in this invention can dynamically predict the tool wear of cutting tools in the future. By integrating a small amount of manual measurement with model prediction results, the prediction results can be corrected in real time, making the results more credible.

[0044] 2. The present invention considers the cumulative effect of time, uses a convolutional neural network to evaluate the degradation increment over a period of time, and uses this as input to predict the tool wear degradation amount over a period of time in the future using a BP neural network-based regression model. This makes the prediction result monotonically increase, and the model is closer to the actual situation.

[0045] 3. The present invention uses the working condition data obtained from the machine tool to predict the wear status, without adding external sensors or wiring, making the method of the present invention more economical, practical and feasible.

[0046] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following content, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained as particularly pointed out in the text and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0048] Figure 1 This is a flowchart of the dynamic prediction of cutting tool wear based on multi-model fusion of the present invention;

[0049] Figure 2 This is a structural diagram of the convolutional neural network model of the present invention;

[0050] Figure 3 It is a flowchart of the algorithm of the hybrid model of the present invention;

[0051] Figure 4 Schematic diagram of a tool wear measurement curve of a tool in an embodiment of the present invention;

[0052] Figure 5 The predicted value and actual measured value of the tool wear prediction model trained in the embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram comparing the discrete points of tool wear values indirectly measured by an artificial intelligence model and the discrete points of tool wear values directly measured by humans in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The following combination Figures 1-6 The preferred embodiments of the present invention will be described in detail, wherein the accompanying drawings constitute a part of the present invention and are used to illustrate the principles of the present invention together with the embodiments of the present invention, but are not used to limit the scope of the present invention.

[0055] During the cutting process, the wear of the cutting tool has a direct impact on the quality of the machined parts. Effectively predicting the tool wear status in the future before cutting is of great value in ensuring the quality of part processing. Based on this, the present invention provides a multi-model fusion cutting tool wear dynamic prediction method. By combining convolutional neural networks and BP neural networks, and integrating multi-source dynamic data information obtained by the machine tool CNC system with a small amount of manually measured tool wear data, the tool wear in the future is dynamically predicted. Based on the prediction results, the operator can judge whether the current equipment can complete the corresponding processing task in the future, so that the judgment of whether to replace the tool can be made in advance to ensure the processing quality.

[0056] like Figure 1 As shown, the specific implementation steps of the multi-model fusion cutting tool wear dynamic prediction method of the present invention are as follows:

[0057] S1: Determine the cutting process machine tool working condition data related to the tool wear process and select the corresponding data collection method.

[0058] Specifically, the "cutting process machine tool operating condition data related to tool wear" described in S1 primarily refers to the blunting of the tool during cutting, which causes changes in tool geometry and abnormal fluctuations in physical properties related to the machining state, including cutting forces, vibration, and noise. Most CNC machine tools currently used by domestic manufacturers require external sensors to acquire this data. However, considering that adding external sensors can interfere with the normal cutting process, it is difficult for manufacturers to modify the machine tools to accommodate these sensors. Therefore, the ideal method for monitoring tool conditions during actual cutting operations is to utilize data generated internally by the machine tool to monitor the current tool wear state without any external sensors. Therefore, the collection and analysis of internal machine tool operating condition data related to the tool is selected, primarily including data related to cutting forces, such as spindle load, spindle torque, and spindle power.

[0059] The "data acquisition method" described in S1 primarily refers to the fact that different CNC systems have different data development interface protocols. To obtain working condition data within CNC machine tools, different communication protocols must be selected based on the system. These typically include RS-232 serial port mode, TCP / IP Ethernet mode, and data acquisition card modes. The Ethernet method provides richer information collection, capable of acquiring data related to cutting process conditions. Currently, most major CNC system manufacturers are equipped with LAN ports for data acquisition. For example, Siemens CNC systems collect machine tool data through OPC, while FANUC systems collect machine tool data through FOCAS2.

[0060] Since cutting tool wear is closely related to cutting force, information about cutting force can be characterized by spindle load and spindle torque.

[0061] Preferably, this embodiment uses the OPC DA protocol to collect the spindle load data and spindle torque data in the Siemens 840Dsl CNC system.

[0062] S2: Conduct tool wear processing tests. According to the actual processing technology, select a process method in a key step and perform cutting processing. Collect tool wear degradation data and cutting process working condition data in the same time interval from cutting with a new tool to cutting with an old tool.

[0063] Specifically, the “process method in a key step” described in S2 is:

[0064] Since there are many factors that affect tool wear, different cutting parameters and different processing objects will lead to different wear rates; therefore, the "process method in a key step" in the method of the present invention specifically includes the following steps:

[0065] a1. Select a fixed cutting object, such as an aluminum alloy blank;

[0066] a2. Select a key process step, such as the process of entering a stable surface cutting process after removing the surface layer of the aluminum alloy blank;

[0067] a3. Select the cutting parameters for the key steps, that is, select the same cutting parameters (spindle speed, feed rate, cutting depth, etc.) used throughout the entire cutting process of S2;

[0068] a4. Obtaining corresponding data, that is, obtaining corresponding tool wear degradation data and cutting process working condition data within the same time period at different times and at the same time intervals.

[0069] Preferably, when conducting the tool wear processing test in this embodiment, the process method in a key step is selected according to the actual processing technology, and cutting processing is performed to collect the tool wear degradation data in the process of cutting from a new tool to an old tool and the cutting process working condition data within the same time.

[0070] The following are examples:

[0071] 1. In this embodiment, the aluminum alloy blank is used;

[0072] 2. Milling the upper surface is the "key process"; first, remove the surface layer of the aluminum alloy blank, and then enter the key step of stable surface cutting;

[0073] 3. Select the corresponding test process parameters as "set the spindle speed to 6000r / min, the feed rate to 2000mm / min, and the cutting depth to 1mm"; this step corresponds to step S3 above;

[0074] 4. Set the data set: Use a new milling cutter to cut the workpiece. After each cut, measure tool wear using a microscope and record the wear of one cutting edge. Simultaneously, collect cutting condition data, such as spindle load and torque, from the machine tool's CNC system throughout the entire cutting process. This cutting condition data is then divided into separate trials, each involving cutting a surface, and mapped to the corresponding tool wear to form the set data set.

[0075] S3: Preprocess the acquired wear degradation data and draw a tool wear variation curve.

[0076] Specifically, the "preprocessing" described in S3 refers to the appropriate manual processing of the measured performance degradation data without affecting the evaluation results. Due to equipment and human error, measurement results may contain certain errors. For example, the amount of tool wear degradation measured in the last measurement may be smaller than that in the previous measurements. This is inconsistent with actual conditions. Therefore, this situation needs to be addressed by removing such abnormal data and filling the gaps with the mean of the normal data before and after to smooth the measured data and calculate the degradation increment for each time period.

[0077] Preferably, in the process of measuring tool wear in this embodiment, due to equipment and human measurement errors, the measurement results may have certain errors. For example, the tool wear measured in the last measurement is smaller than the tool wear measured in the previous measurements, which is not in line with the actual situation. Therefore, this situation needs to be processed, such abnormal data is eliminated, and the average of the normal data before and after is taken for filling, and the measurement data is smoothed. Since the evaluation results of this example are calculated within a relatively large tool wear fluctuation range, and the data preprocessing method has little effect on the adjustment of the measurement data, within the fluctuation range of 0.001mm, the measurement data preprocessing method will not have a significant impact on the final evaluation results, but only facilitates subsequent modeling.

[0078] like Figure 4 The figure shows the tool wear curve obtained from the first test measurement of this example. Because the wear rate is high in the early stages of wear and the tool cannot function properly in the later stages, this example considers using data from the stable wear period for modeling. Furthermore, considering that the limited amount of data in the early stages could affect the model's prediction performance, calculations were considered starting after 30 measurement data points were acquired.

[0079] S4: Extract tool wear degradation data and corresponding working condition data during the stable wear period, establish multiple data sample sets based on the time series, and use convolutional neural networks to establish a wear increment prediction model for these samples.

[0080] like Figure 2 As shown in Figure 4, the "stable wear period" described in S4 refers to an intermediate stage in the tool wear process. Generally, the tool wear process is divided into three stages: initial wear, stable wear, and severe wear. When a new tool begins cutting, the cutting edge is relatively sharp and the tool wear rate is rapid, resulting in significant vibration in the cutting system. The workpiece surface roughness fluctuates within a certain range as the tool wear on the flank increases. The amplitude of the roughness variation increases as tool wear increases. After the initial wear period, tool wear increases, the contact area between the workpiece and the flank increases, and the pressure per unit area gradually decreases. The tool enters a stable wear period, where the wear rate slows and the cutting process is relatively smooth. This stage also represents the effective operation of the tool. During this period, workpiece machining accuracy gradually improves as tool wear increases. When tool wear reaches a certain level, the cutting edge becomes blunt, cutting forces and cutting temperatures increase dramatically, and the tool enters a severe wear period, with a sharp increase in the wear rate and a tendency towards system instability. Continuing to use a tool during the intense wear phase is not worth the effort. It accelerates the consumption of tool material and compromises machining quality. If edge chipping occurs, the losses are even greater. Therefore, to ensure machining quality, the tool must be replaced when it reaches the blunting threshold before entering the intense wear phase. Given the high wear rate during the early stages of wear and the inability to function properly in the later stages, the proposed method utilizes data from the stable wear phase for modeling and prediction.

[0081] Specifically, the "wear increment prediction model" described in S4 refers to predicting the degradation increment in each time period through the working condition data of the cutting process, and predicting the degree of degradation based on the accumulation of degradation increments. The method proposed in the present invention uses the convolutional neural network method to directly construct a performance degradation increment prediction model for the preprocessed data signal. The convolutional neural network established in this study is to map the working condition data input within a certain period of time into a degradation increment. The convolutional neural network model structure used adopts the design idea of the residual network. The specific structure of the residual network is as follows. Figure 2Specifically, considering the advantages of batch normalization (BN) layers, such as preventing overfitting and accelerating model convergence, BN layers are incorporated into each layer of the model. The "ReLU" after the BN layer represents a nonlinear transformation of the feature matrix, increasing the sparsity of the network, making the extracted features more representative and enhancing generalization. Each value in the pooling layer's feature matrix corresponds to the larger of the two adjacent elements in the BN layer's feature matrix. When operating on the feature matrix in the BN layer, regions do not overlap, resulting in a halved size of the feature matrix after pooling. A residual network structure appears after the first pooling layer. The output of the pooling layer is downsampled and directly added to the output of the second pooling layer, accelerating model convergence and preventing vanishing gradients. The subsequent convolution and pooling processes are similar to those described above. After the convolution and pooling operations, the output is connected to two fully connected layers, resulting in the final output being a degenerate increment.

[0082] Preferably, this embodiment uses the tool wear data of the stable period and the corresponding spindle load and torque as training samples to train the convolutional neural network model, and sets the input vector of each sample to be a 32×2 matrix, which means that the acquisition period is 0.1s, and a total of 32 sets of spindle load and torque signals (32 in the matrix) are collected within a time of 3.2s; the output of the model is the corresponding wear increment under the corresponding 32×2 matrix input. Although the speed is slower than calculating by using the vibration signal collected by the acceleration sensor, for engineering applications, predicting the current tool wear level once every 3.2s of data is of certain guiding significance to the operator. During the training process, 30 samples are randomly selected for training each time, and the remaining samples are used for verification. The training is terminated when the model meets the accuracy requirements, and the optimal wear increment prediction model is obtained.

[0083] S5: Based on the data samples established in S3, a tool wear prediction model is established using the BP neural network.

[0084] Specifically, the "tool wear prediction model" described in S5 uses the tool wear degradation at several previous moments (multiple previous moments), the cutting time from the previous moment to the current moment, and the degradation increment within the current cutting time period predicted by the wear increment prediction model as input. A regression model constructed using a BP neural network is used to predict the tool wear degradation at the next moment. BP neural network is short for error back propagation neural network and is generally composed of an input layer, one or more hidden layers, and an output layer. Preferably, the present invention uses a three-layer BP neural network to construct the tool wear prediction model.

[0085] In this embodiment, a tool wear prediction model is established using the wear increment, the tool wear at the first four measurement moments, and the cutting interval experienced by the wear increment (six parameters). The output is the tool wear at the next measurement moment. To ensure the accuracy of the model, the wear increment data used in the modeling here is the actual measurement data. In the actual prediction process, the wear increment data is obtained through the wear increment prediction model. This case uses a BP neural network to predict tool wear. The input vector of each sample is set to a 6×1 matrix, and the output of the model is the corresponding tool wear. During the model training process, 70% of the samples are used for training, 15% of the samples are used for verification, and 15% of the samples are used for testing.

[0086] like Figure 5 The figure shows a comparison between the model's predictions and the actual measured results. The tool wear prediction model is able to accurately predict the tool wear at the next moment. However, it was found that seven of the 136 predictions were inconsistent with actual conditions, meaning that the tool wear at the next moment was equal to or less than the tool wear at the previous moment.

[0087] S6: Under the constraints of the prediction result matching test criteria, the wear increment prediction model based on convolutional neural network and the tool wear prediction model based on BP neural network are combined to form a hybrid prediction model to predict the tool wear of the tool under the same process conditions in the future.

[0088] like Figure 3 As shown in the figure, the “hybrid prediction model” refers to combining the wear increment prediction model of S4 and the tool wear prediction model of S5 under the constraint of the prediction result matching test criteria to form a hybrid prediction model; it is used to predict the tool wear of the tool under the same process conditions in a period of time in the future.

[0089] The convolutional neural network and the BP neural network are integrated to build a hybrid prediction model. The specific construction process is as follows:

[0090] First, a convolutional neural network is used to directly construct a performance degradation increment prediction model for the preprocessed data signal. This is the first model. Then, considering the cumulative effect of time, the tool wear degradation amount of the previous moments, the cutting time from the previous moment to the current moment, and the degradation increment in the current cutting time period predicted by the convolutional neural network model are used as input. On the basis of the convolutional neural network model, a BP neural network prediction regression model is added. This is the second model. The two models are used to form a hybrid prediction model to realize the prediction of tool wear degradation.

[0091] The use of hybrid prediction model belongs to the long-term cumulative prediction method. The premise of its use is that the previous degradation data must have been obtained. When there is no degradation data at the previous moment, it is impossible to predict the performance tool wear degradation amount by directly using the signal of a period of time. The block diagram of the proposed algorithm is shown in the figure below. Figure 3 As shown in Figure 1. In this process, the tool wear degradation data Y0:t (predicted or actually measured) at and before time t is known. The signal v(st) from time t to time s is used to predict the degradation increment x(st) during this period through a convolutional neural network model. The tool wear degradation data at and before time t, i.e., y(t), y(t-1), …, y(td), and the time interval (st), are used as input to construct a regression model, and the output is the tool wear degradation data y(s) at time s.

[0092] Among them, the "prediction result matching test criteria" described in S6 means that during the prediction, by collecting real-time working condition data during the cutting process and using the trained model to calculate the current performance tool wear degradation, it is smaller than the previous prediction result and does not conform to the actual degradation situation. Therefore, it is necessary to test the predicted performance tool wear degradation. When the prediction result is smaller than the tool wear degradation at the previous moment, or the difference between the prediction result and the tool wear degradation at the previous moment deviates greatly from the degradation increment predicted by the convolutional neural network, it means that the prediction result does not match, and it is necessary to stop the machine, manually measure the performance tool wear degradation, update the data and train the model again. This can reduce the number of shutdown measurements on the one hand, and can also continuously supplement and update the data to improve the reliability of the prediction model.

[0093] The following are the matching test criteria for the prediction results:

[0094]

[0095] Among them, y(t) and y(t-1) are the predicted value of tool wear degradation at the current moment t, the predicted value of tool wear degradation at the previous moment t-1, or the actual tool wear degradation value obtained by manual measurement due to the mismatch between the predicted results; x(t) is the degradation increment during this period predicted using the working condition data from t-1 to t, and q is the deviation threshold of the degradation increment, which is between 0 and 1. The closer it is to 0, the stricter the inspection standard and the better the prediction performance of the model. However, it will also lead to an increase in the number of manual measurements. It is necessary to select an appropriate value based on the actual situation.

[0096] This embodiment preferably combines the wear increment prediction model based on the convolutional neural network with the tool wear prediction model based on the BP neural network, and uses the machine tool data of the first test to predict the tool wear according to the prediction result matching test standard shown in formula (1). When the prediction result of the artificial intelligence model does not match, the actual tool wear at that moment is used to replace the predicted tool wear, and the model is updated and calibrated.

[0097] like Figure 6 The figure shows the tool wear curves generated by combining indirect wear measurement using an AI model with direct wear measurement, as well as the actual wear curve. Of the 136 predicted results, 26 were inconsistent with the actual situation. In these cases, the machine was stopped for measurement, and the actual measured data replaced the incorrect predicted data, and the model was updated.

[0098] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed in the present invention should be covered by the scope of protection of the present invention.

Claims

1. A multi-model fusion method for dynamic prediction of cutting tool wear, characterized in that: The steps include: S1. Determine the cutting process machine tool working condition data to be collected and determine the corresponding data collection method; S2. Conduct tool wear processing test: Design cutting tests; perform cutting operations; collect tool wear and degradation data during the cutting process, as well as collect machine tool operating condition data during the same cutting process; S3, preprocessing the tool wear degradation data obtained in S2, extracting the tool wear degradation data during the stable wear period in S2 and the corresponding machine tool working condition data; establishing multiple data sample sets based on the time series; and generating a tool wear variation curve; S4. Establish a wear increment prediction model: Use the data sample set to train the convolutional neural network model and build a convolutional neural network wear increment prediction model; S5. Establish tool wear prediction model: Based on the data samples in S3, a BP neural network-based tool wear prediction model is established. The tool wear prediction model is a regression model constructed using the BP neural network. The input parameters of the regression model include the tool wear degradation at multiple previous moments, the cutting time from the previous moment to the current moment, and the degradation increment within the current cutting time period obtained by the wear increment prediction model. The regression model is used to predict the tool wear degradation at the next moment. S6. Establish a hybrid prediction model: The wear increment prediction model of S4 and the tool wear prediction model of S5 are combined under the constraint of the prediction result matching test criteria to form a hybrid prediction model; it is used to predict the tool wear of the tool under the same process conditions in the future period of time; Among them, the prediction result matching test criterion means that the tool wear degradation amount calculated by the trained hybrid prediction model conforms to the following formula: ； Where, y(t)-the predicted value of tool wear degradation at the current moment t; y(t-1) - the predicted value of tool wear degradation at the previous moment t-1 or the actual tool wear degradation value obtained by manual measurement because the predicted result does not match the actual tool wear degradation value; x(t) - the degradation increment during this period predicted using the operating data from time t-1 to time t; q-degradation increment deviation threshold; where 0 <q<1。 2. The dynamic prediction method for cutting tool wear based on multi-model fusion according to claim 1 is characterized in that: The cutting process machine tool working condition data related to the tool wear process in S1 are data related to the cutting force, including spindle load, spindle torque and spindle power.

3. The dynamic prediction method for cutting tool wear based on multi-model fusion according to claim 1, characterized in that: The data acquisition mode in S1 includes one or any combination of the serial port mode of RS-232 protocol, the Ethernet mode of TCP / IP protocol and the data acquisition card mode.

4. The method for dynamic prediction of cutting tool wear based on multi-model fusion according to claim 1, characterized in that: The process method in the key step of S2 includes the following steps: S21, selecting a key cutting step for which tool wear status monitoring is required; S22, selecting process parameters for key cutting steps; S23. Perform a cutting test to obtain tool wear degradation data during the cutting process and cutting process working condition data during the same period of time.

5. The method for dynamic prediction of cutting tool wear based on multi-model fusion according to claim 1, characterized in that: The "pre-processing" mentioned in S3 refers to the elimination of the cutting process obtained in S2 Abnormal data in the machine tool working condition data are filled with the mean value of the normal data adjacent to the abnormal data, and the measurement data are smoothed.

6. The method for dynamic prediction of cutting tool wear based on multi-model fusion according to claim 1, characterized in that: The data sample set described in S4 is a data sample set established by machine tool operating data collected during the stable wear period of the cutting test.

7. The method for dynamic prediction of cutting tool wear based on multi-model fusion according to claim 1, characterized in that: The wear increment prediction model in S4 is a performance degradation increment prediction model constructed by using a convolutional neural network method and taking preprocessed data signals as input.

Images