A device and method for predicting the life of a machine tool

Abstract

The present application relates to a kind of machine tool tool life prediction device and life prediction method, belong to numerical control machine tool tool life prediction field, solve the machine tool protocol inconsistency in prior art, data is not unified, tool life prediction is not accurate etc., the machine tool tool life prediction device of the present application includes edge unit and monitoring control unit;Edge unit includes AI microprocessor, 4G DTU, ESP WIFI and A / D input port;Monitoring control unit is used to process the machine tool data uploaded by edge unit.It is suitable for the real-time prediction tool life device and method of multiple machine tools, under the premise that it does not affect normal machining of machine tool, obtains the running data when it is processed, using the method of deep learning, establishes life prediction model, and is deployed in device, real-time prediction of tool life on edge side, improve its utilization and accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of numerical control machine tool tool life prediction, and relates to a machine tool tool life prediction device and a life prediction method, in particular to an edge side attention mechanism based LSTM network life prediction device and a life prediction method. BACKGROUND

[0002] In the process of numerical control machine tool machining, failures often occur, among which accidents caused by tool damage account for 80% of the total accidents in the factory, machine tool downtime caused by tool failure exceeds 20%, and maintenance cost accounts for 15% to 40% of the production cost. Therefore, it is particularly important to monitor the tool state in real time and predict its remaining life. Different equipment manufacturers have different degrees of openness to equipment data and communication protocols, which also makes the interoperability between various types of equipment a challenge.

[0003] At present, most enterprises rely on experience or regular replacement strategies when replacing tools, which leads to insufficient use of tools, increases costs, and cannot timely handle problems that occur during tool use. In view of the existing problems, some researchers have also carried out related research.

[0004] Xi'an University of Electronic Science and Technology in its application for a public number CN113560955A "A numerical control machine tool tool remaining service life prediction method, system and application" proposes a tool life prediction method. The method collects controller signals and sensor signals in the working process of the numerical control machine tool, and uses a long short-term memory network and an attention mechanism to establish a tool remaining service life prediction model to realize the prediction of the remaining service life of the numerical control machine tool tool, and improves the generalization ability of the tool remaining service life prediction model. However, it does not solve the problem of inconsistent machine tool protocols and non-uniform data, and the actual application range is relatively limited.

[0005] Keda Xunfei South China Artificial Intelligence Research Institute (Guangzhou) Co., Ltd. in its application for a public number CN114378639A "A tool life prediction method and related device". The method extracts various image signals and sensor signals of the tool to be tested, combines image and sensor feature data, and uses a preset attention mechanism to realize accurate tool remaining life prediction. By comprehensively utilizing the characteristics of image and sensor signals, the invention can significantly improve the prediction accuracy, thereby improving the processing efficiency, quality, and reducing the production cost. However, the implementation process is complex, requires high computing resources, and is highly dependent on data quality, which may cause problems in device adaptation, resulting in limited universality, and may also bring high cost and potential real-time problems.

[0006] Shenzhen Advanced Technology Research Institute in its application for the publication number CN115600512A "a kind of tool life prediction method based on distributed learning", by the tool feature data in each terminal local model training is carried out, and the training loss is sent to the central terminal for model parameter updating, to realize the accurate prediction of tool life. This method can protect data privacy while improving the accuracy of the prediction model using data from multiple manufacturers. However, this method also has some shortcomings, including complex implementation process, high computing resources, and may face data synchronization and communication delay problems. In addition, the system configuration and maintenance of distributed learning requires higher, which may lead to increased deployment and operation costs.

[0007] To solve the problems of complex installation, high cost and poor adaptability in existing tool monitoring and life prediction schemes, an improved method is proposed, which optimizes the attention mechanism based on the LSTM architecture. The adaptive attention mechanism is introduced in the input layer, which can dynamically adjust the weight and optimize the information flow processing. At the same time, the improved LSTM unit improves the modeling ability of long and short-term dependencies. In addition, intelligent data fusion technology is introduced in the data preprocessing stage, which automatically corrects and standardizes different sources of data at the edge layer to ensure the consistency and reliability of the data. The invention aims to provide a real-time tool life prediction device and method that is easy to install, cost-effective and widely applicable, to reduce the bandwidth pressure of local area network and improve the utilization rate of tools. SUMMARY

[0008] To solve the problems of inconsistent machine tool protocols, non-uniform data, and inaccurate tool life prediction, the invention provides a machine tool life prediction device and method, which is suitable for real-time prediction of tool life of various machine tools. Without affecting the normal processing of machine tools, the running data during processing is obtained, and a life prediction model is established using deep learning method and deployed in the device to predict the tool life in real time on the edge side, improving the utilization rate and accuracy.

[0009] The invention provides a machine tool life prediction device, which includes an edge unit and a monitoring and control unit.

[0010] The edge unit includes an AI microprocessor, a 4G DTU, an ESP WIFI, and an A / D input port.

[0011] The monitoring and control unit is used to process the machine tool data uploaded by the edge unit.

[0012] Optionally, the AI microprocessor is used to collect and process machine tool data in real time, perform processing tasks, extract various data features to determine tool status, and run a life prediction model to input feature data to obtain tool remaining life.

[0013] Optionally, the 4G DTU is used for communication with the monitoring control unit, and the collected data is transmitted to the cloud server for life prediction model training, and the trained life prediction model is deployed to the edge unit.

[0014] Optionally, the ESP WIFI is used for communication with the monitoring control unit, and machine tool data, state and prediction information are displayed in real time.

[0015] Optionally, the A / D input port is used to collect machine tool data during processing.

[0016] Another aspect of the present application also provides a machine tool tool life prediction method, using the machine tool tool life prediction device described above, the specific steps are as follows:

[0017] Step 1, collect machine tool data; divide the machine tool data into training set and test set after processing;

[0018] Step 2, build a life prediction model, and use the training set to train the life prediction model;

[0019] Step 3, after training the life prediction model, use the test set to set different starting points, and use the starting point and the machine tool data before the starting point to make prediction;

[0020] Step 4, predict whether the value at the next time point is greater than the maximum life threshold of the tool, if not, add the predicted data to the tail of the test set and return to step 3, if yes, go to step 5;

[0021] Step 5, make multi-step prediction until the predicted maximum spindle load is greater than the threshold, and the prediction is finished, and the prediction result is obtained.

[0022] Optionally, the specific steps of collecting machine tool data in step 1 and dividing machine tool data into training set and test set after processing are as follows:

[0023] Step 11, divide the machine tool data according to the processing period; calculate a plurality of characteristic values of the machine tool data in each processing period, and select a characteristic value positively correlated with tool wear from the plurality of characteristic values;

[0024] Step 12, dimension reduction is performed on the characteristic value positively correlated with tool wear obtained in step 11, and the least redundancy and maximum correlation method is used to select the most important characteristic value of tool wear;

[0025] Step 13, define the training data set and the test set based on the five most important characteristic values of tool wear.

[0026] Optionally, the plurality of characteristic values include mean, variance, peak factor, margin factor, skewness, kurtosis, waveform factor, maximum value, standard deviation, and root amplitude.

[0027] Optionally, the specific steps of training the life prediction model in step 2 are as follows:

[0028] Step 21, using the sequence segmentation function to segment the main shaft load signal of the continuous time points in the training data set into input sequences of input time point samples and output time point samples;

[0029] Step 22, normalizing the input sequences of the input time point samples and the output time point samples to the interval [0, 1];

[0030] Step 23, adjusting the input sequences of the input time point samples and the output time point samples normalized to the interval [0, 1] to the input format required by the long short-term memory network (LSTM) layer;

[0031] Step 24, based on the input format obtained in step 23, constructing a four-layer long short-term memory network (LSTM) based on attention mechanism, using grid search method to find the optimal hyperparameters, and obtaining the trained life prediction model.

[0032] Optionally, the four-layer long short-term memory network (LSTM) based on attention mechanism includes an attention layer, an LSTM layer, a Dropout layer, and a fully connected layer.

[0033] Compared with the prior art, the present application has at least the following beneficial effects:

[0034] 1. The machine tool tool life prediction device and method disclosed in the present application realizes the isomerization of heterogeneous devices by acquiring internal data and collecting spindle current data through direct communication with the machine tool, defining a general node model of the machine tool to collect data, and pre-processing the machine tool data to eliminate invalid data, thereby reducing bandwidth pressure and improving real-time performance.

[0035] 2. The machine tool tool life prediction device and method disclosed in the present application separates data collection and calculation tasks by using a dual-core architecture chip on the edge side, completes the prediction of tool life and displays and configures related parameters of machine tool data, deploys the model from the cloud to the edge, completes the remote update of device parameters and firmware, realizes the decoupling of the controller and the machine tool, and the cooperation of the cloud and the edge task.

[0036] 3. The device and method for tool life prediction of a machine tool disclosed by the application, by establishing an LSTM model based on an attention mechanism, higher weights are given to time points with greater influence in input time series data, the accuracy of tool residual life prediction is improved to more than 94%, and the service life of the tool is improved to 1.5 times of the original. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The application is an application system for an embodiment of the application;

[0038] Fig. 2(a) and (b) is a schematic diagram of the device for tool life prediction of a machine tool of the application;

[0039] Figure 3 The application is a flowchart of LSTM model training and prediction based on an attention mechanism;

[0040] Figure 4 The application is a flowchart of the method for tool life prediction of a machine tool;

[0041] Figure 5 The application is a comparison diagram of tool residual life prediction curves for an embodiment of the application. DETAILED DESCRIPTION

[0042] In order to facilitate understanding of the purpose, advantages and application method of the application, the application will be described in detail below in combination with examples. It should be pointed out that the described embodiments are only intended to facilitate understanding of the application and do not limit the application.

[0043] An embodiment of the application, as shown in Figures 1-5 A device for tool life prediction of a machine tool is disclosed, which comprises an edge unit and a monitoring control unit.

[0044] The tool life prediction device is arranged in a machine tool. The machine tool generates a large amount of machine tool data during processing, and communicates with the tool life prediction device through Ethernet or serial communication.

[0045] Further, the edge unit uses a general OPC UA node to collect spindle load data, and can be applied to machine tools of different protocols. The edge unit communicates with a Fanuc machine tool through a FOCAS protocol, or communicates with a Brother machine tool through a message, etc. Then, the machine tool data transmitted by the machine tool is preprocessed, the data is uniformly placed on the data acquisition node, and then sent to the monitoring control unit. The edge unit is responsible for receiving the latest algorithm model sent by the monitoring control unit, and calling the tool life in the device for real-time prediction;

[0046] The edge unit comprises an AI microprocessor, a 4G DTU, an ESP WIFI and an A / D input port.

[0047] Further, a monitoring control unit is used to process the machine tool data uploaded by the edge unit.

[0048] The monitoring control unit includes cloud server monitoring and real-time data display on the edge side.

[0049] The device for predicting the tool life of a machine tool disclosed in the application directly communicates with the machine tool through an Ethernet port or a serial port to obtain internal data of the machine tool.

[0050] The AI microprocessor adopts a dual-core architecture chip to collect and process machine tool data in real time, perform high-computing processing tasks, extract various data features to determine the tool state, and run a deep learning algorithm model to input the feature data to obtain the remaining tool life.

[0051] The machine tool data collected in real time mainly includes spindle machining load obtained by directly communicating with the machine tool through an Ethernet port or a serial port according to a communication protocol.

[0052] The main information related to tool wear in the various data feature retention signals is removed by using Kalman filtering to eliminate irrelevant information, reduce the complexity of subsequent learning and calculation, and improve the prediction accuracy.

[0053] The 4G DTU is used to communicate with the monitoring control unit, transmit the collected data to the cloud server for model training, and deploy the trained model to the edge unit to realize the collaborative work of the cloud and the edge, and ensure the latest and accurate prediction of the model.

[0054] The ESP WIFI is used to communicate with the monitoring control unit and display machine tool data, state and prediction information in real time on the machine tool side.

[0055] The A / D input port is used to collect the current data of the spindle motor during machining.

[0056] Further, the monitoring control unit trains a life prediction model based on an attention mechanism LSTM model and converts it, and deploys the trained life prediction model to the edge unit, and the specific training steps are as follows.

[0057] Step 1, collect machine tool data; divide the machine tool data according to the machining period; calculate ten characteristic values of the machine tool data in each machining period, including mean, variance, peak factor, margin factor, skewness, kurtosis, waveform factor, maximum value, standard deviation and square root amplitude; and select characteristic values positively correlated with tool wear from the ten characteristic values.

[0058] Step 2, reduce the dimensionality of the characteristic values positively correlated with tool wear obtained in step 1, and select the five most important characteristic values of tool wear using the least redundancy and maximum correlation method.

[0059] Specifically, the five most important characteristics of tool wear are standard deviation, peak factor, root square magnitude, waveform factor, and maximum value;

[0060] Furthermore, the expression for the minimum redundancy and maximum correlation is:

[0061] mRMR = max s [D(S,y)-R(S)]

[0062] Where S is the currently selected feature set; max s This indicates that the feature that maximizes the objective function is selected from the candidate feature subset S; y is the number of labels corresponding to the currently selected feature set; D(S,y) is the maximum correlation between the feature value and the label; and R(S) is the redundancy between the feature values ​​within the feature set.

[0063] Furthermore, the expressions for the maximum correlation between feature values ​​and label values ​​and the redundancy among feature values ​​within the feature set are:

[0064]

[0065] Where |S| is the size of the feature set; s i ,s j Let X and Y be the i-th and j-th features in the feature set S; I(X,Y) is the mutual information value between X and Y, where I(s) = ... i ,y) and I(s i ,s j In ), s i Corresponding to X, y and s j The expression for the mutual information value between Y and X is:

[0066]

[0067] In the formula, P(x i Let x be the i-th value of the random variable X. i The probability of P(y) i Let y be the j-th value of the random variable Y. i The probability of P(x) i ,y i Let random variables X and Y both take the value x. i and y i The joint probability.

[0068] Step 3: Define the training dataset and test set based on the five most important features related to tool wear;

[0069] Step 4, using the sequence segmentation function to segment the principal axis load signal of the continuous time points in the training data set into input sequences of input time points (X) and output time point samples (Y);

[0070] It can be understood that each sample in the input time point (X) contains the values of 5 consecutive time points, and the output time point sample (Y) is the true value of the next time point as a label;

[0071] Wherein, the value of one time point corresponds to the 5 characteristic values most important to tool wear calculated from the machine tool processing data of one period.

[0072] Step 5, the input sequences of input time points (X) and output time point samples (Y) are normalized to the interval [0, 1] by Min-Max standardization processing;

[0073] The expression of the Min-Max standardization processing is:

[0074]

[0075] Wherein, x * is the standardized data, x min is the minimum value of the data before standardization, and x max is the maximum value of the data before standardization.

[0076] Step 6, the input sequences of input time points (X) and output time point samples (Y) normalized to the interval [0, 1] by Min-Max standardization processing are adjusted to the input format required by the long short-term memory network LSTM layer (i.e. batch size, time step, input feature number) shape;

[0077] Step 7, based on the input format obtained in step 6, a four-layer long short-term memory network LSTM based on attention mechanism is constructed, the grid search method is used to find the optimal hyperparameters, and a trained life prediction model is obtained.

[0078] Specifically, the first layer is an attention layer, the input shape is a time series data (i.e. batch size, time step ， input feature number), the feature vector of each time step is weighted, two transpose layers are used at the beginning and end to give higher weights to time points with greater influence, while ensuring that the feature output shape after weighting is the same as the input (i.e. batch size, time step ， input feature number);

[0079] The second layer is an LSTM layer, which receives the input data from the attention layer with a shape of (batch size, time steps, input feature number), models the time series of the input sequence, uses 50 neurons and ReLU activation function, calculates the hidden state vector, and outputs the hidden state of the last time step with a shape of batch size 50;

[0080] The third layer is a Dropout layer, which receives the hidden state of the last time step from the LSTM layer with a shape of batch size 50, randomly discards a certain proportion (e.g. 50%) of neurons to prevent overfitting, and outputs the hidden state with a shape of batch size 50;

[0081] The fourth layer is a fully connected layer, which receives the hidden state from the Dropout layer with a shape of batch size 50, maps each input to a new output space, is set to 50 neurons, and uses an activation function to output a feature representation with a shape of batch size 50 for subsequent loss calculation tasks.

[0082] Further, the steps of the grid search method include the following steps:

[0083] S1, define the parameter grid: set the hyperparameters to be optimized and their value ranges, for the hyperparameters θ1, θ2,..., θ M , list all possible combinations (similar to "grid"), see Table 1;

[0084] Table 1: Example of grid search method

[0085]

[0086] S2, traverse all combinations: for each combination of hyperparameters θ (m) in the grid, perform model training and calculate the performance indicator (such as accuracy, F1 score, etc.) under the corresponding combination of hyperparameters.

[0087] S3, use cross-validation method to evaluate the model performance of each combination of hyperparameters, and obtain the performance indicator P(θ (m) ).

[0088] S4, find the combination of hyperparameters θ best with the best performance indicator:

[0089]

[0090] where Θ represents all possible combinations of parameters; P represents the model performance indicator; θ (m) represents the combination of parameters θ of the mth grid search, m is the index of the combination; θ best represents the combination of parameters that can maximize the model performance indicator P; P(θ(m) ) is the performance indicator of the model under the parameter combination θ (m) , i.e. accuracy, F1 score, etc.; arg max represents finding the parameters that can maximize P(θ (m) ); θ∈θ represents that the parameter θ belongs to the parameter set Θ.

[0091] S5, use the Adam optimizer and mean square error (MSE) as the loss function to train the model, set the number of iterations to 1000 times, and use the early stopping method to obtain the optimal model, and set the early stopping step to 50 to prevent overfitting;

[0092] The expression of the mean absolute error (MAE) is:

[0093]

[0094] Where y i is the i-th real label, is the i-th model prediction value, and n is the sample size.

[0095] Update the parameters using the Adam optimizer:

[0096]

[0097] Where m t and v t are the first and second moments of the gradient at time t, respectively; ∈ is a smoothing term to prevent division by zero; θ represents the parameters to be optimized in the model; α represents the learning rate, which controls the magnitude of parameter updates.

[0098] S6, in order to verify the effectiveness of the model, convert the model into a C function and deploy it to the edge unit for running, replace a new tool on the numerical control machine tool for machining, calculate the standard deviation, peak factor, root amplitude, waveform factor and maximum value of the spindle load at each machining in the edge unit, then input to the model for multi-step prediction to obtain the current predicted load value, judge whether it is greater than the replacement threshold, and the predicted tool remaining life is obtained by subtracting the two, compare the predicted tool remaining life with the real remaining life (i.e. the output time point sample (Y) obtained in step 4), if it is close to the real remaining life, the model prediction accuracy is high, which can be used to guide tool replacement, reduce waste or overuse; if it is not close to the real remaining life, adjust the data set, optimize feature extraction or retrain the model to improve adaptability and accuracy.

[0099] Exemplarily, the predicted life curve of the tool with different use times is as shown in Figure 5 , the average accuracy of tool remaining life prediction after 1000 times and beyond can reach more than 94%.

[0100] Another embodiment of the present application discloses a machine tool tool life prediction method, which installs the aforementioned prediction device on the machine tool side, defines a node model common to the machine tool, collects machine tool processing operation data in real time to realize the isomerization of heterogeneous devices, extracts characteristic values, reduces dimensions and standardizes the collected data, then constructs an LSTM model based on an attention mechanism, trains the data after standardization, and finally deploys the trained model to an embedded device for operation and calculation to realize real-time prediction of tool life in the edge device.

[0101] The machine tool tool life prediction method disclosed by the present application uses indirect prediction to analyze data indicators representing tool wear from collected data, establishes a model to predict the changes of the data indicators, and until the indicators reach a threshold value representing tool failure, then calculates the time from the prediction time point to the threshold value as the remaining life of the tool.

[0102] Step 1, divide the data after processing into training set and test set;

[0103] Step 2, build a life prediction model and train it using the training set;

[0104] Further, based on the aforementioned monitoring control unit, the method trains the network model by training the life prediction model based on the LSTM model based on the attention mechanism and converting it, and deploys the trained life prediction model to the edge unit.

[0105] Step 3, after training the life prediction model, use the test set to set different starting points and use the data before the starting point for prediction;

[0106] Step 4, predict whether the value at the next time is greater than the maximum life threshold of the tool, and if it is not greater than the threshold, add the predicted data to the tail of the test set;

[0107] Specifically, let P(t+1) be the tool life prediction value at the next time t+1, and L be the maximum life threshold of the tool. If P(t+1)<L, then:

[0108] Test set <- Test set U P(t+1).

[0109] Step 5, perform multi-step prediction, and repeat until the predicted spindle load maximum value is greater than the threshold, the prediction is complete, and the prediction result is obtained.

[0110] Further, according to the prediction load trend of the prediction result, the tool replacement decision is obtained; through load feature analysis, the tool wear degree and machining performance are understood; the prediction result is compared with the actual situation to provide a basis for optimizing the model.

[0111] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A machine tool tool life prediction method of predicting a machine tool tool life using a machine tool tool life prediction device, characterized by, The machine tool life prediction device comprises an edge unit and a monitoring control unit; the edge unit comprises an AI microprocessor, a 4G DTU, an ESP WIFI and an A / D input port; the monitoring control unit is used for processing machine tool data uploaded by the edge unit; The specific steps are as follows: Step 1, collect machine tool data; divide the machine tool data into a training set and a test set after processing; Step 2, build a life prediction model, and train the life prediction model using the training set, the specific steps are as follows: Step 21, use the sequence segmentation function to segment the spindle load signal at the continuous time points in the training data set into input sequence of input time point samples and output time point samples; Step 22, normalize the input sequence of input time point samples and output time point samples to the interval [0, 1]; Step 23, adjust the input sequence of input time point samples and output time point samples normalized to the interval [0, 1] to the input format required by the long short-term memory network LSTM layer; Step 24, based on the input format obtained in step 23, construct a four-layer long short-term memory network LSTM based on attention mechanism, find the optimal hyperparameters using grid search method, and obtain the trained life prediction model; after training the life prediction model, use the test set to set multiple prediction starting points, and use the machine tool data before the prediction starting points for prediction; Step 4, predict whether the value at the next time point is greater than the maximum life threshold of the tool, if not, add the predicted data to the tail of the test set and return to step 3, if yes, go to step 5; Step 5, perform multi-step prediction until the predicted maximum spindle load is greater than the threshold, and the prediction is completed to obtain the prediction result.

2. The machine tool tool life prediction method according to claim 1, characterized in that, The AI microprocessor is used for real-time acquisition and processing of machine tool data, and performs processing tasks, extracts various data features to determine the tool state, and runs the life prediction model to input the feature data to obtain the tool remaining life.

3. The machine tool tool life prediction method according to claim 1, characterized in that, The 4G DTU is used for communication with the monitoring control unit, and transmits the collected data to the cloud server for life prediction model training, and deploys the trained life prediction model to the edge unit.

4. The machine tool tool life prediction method according to claim 1, characterized by, The ESP WIFI is used for communication with the monitoring control unit, and real-time display of machine tool data, state and prediction information.

5. The machine tool tool life prediction method according to claim 1, characterized by, The A / D input port is used for collecting machine tool data during processing.

6. The machine tool tool life prediction method according to any one of claims 1 to 5, characterized in that, The specific steps of step 1 of collecting machine tool data; dividing the machine tool data into a training set and a test set after processing are as follows: Step 11, divide the machine tool data according to the processing period; calculate a plurality of characteristic values of the machine tool data of each processing period, and select a characteristic value positively correlated with tool wear from the plurality of characteristic values; Step 12, reduce the dimensionality of the characteristic value positively correlated with tool wear obtained in step 11, and select the most important characteristic value of tool wear using the least redundancy and maximum correlation method; Step 13, define the training data set and the test set based on the most important characteristic value of tool wear.

7. The machine tool tool life prediction method according to claim 6, characterized in that, The plurality of characteristic values include mean, variance, peak factor, margin factor, skewness, kurtosis, waveform factor, maximum value, standard deviation, and root-mean amplitude.

8. The machine tool cutting tool life prediction method according to any one of claims 1 to 5, characterized by, The four-layer attention mechanism-based long short-term memory network (LSTM) includes an attention layer, an LSTM layer, a Dropout layer, and a fully connected layer.

Images