A method for predicting contamination distribution in a robotic machining process

Abstract

The application discloses a kind of robot machining process pollutant distribution prediction methods, collect the small particle dust mass concentration of each collection point at different machining conditions of robot in each tool position, chip mass and chip quantity change with processing time, the short-term memory network model constructed at each collection point is trained;In actual processing, the small particle dust mass concentration of current time, chip mass and chip quantity are predicted based on the original data set data of each collection point and short-term memory network model;With the change of tool center coordinates and processing time, the small particle dust mass concentration of current time, chip mass and chip quantity predicted at each collection point are respectively superimposed with the small particle dust mass concentration, chip mass and chip quantity obtained at each time before current time, so that the application can predict the pollutant distribution at each collection point corresponding to robot machining trajectory.

Description

Technical Field

[0001] This invention belongs to the field of robot machining contaminant detection technology, and particularly relates to a method for predicting contaminant distribution during robot milling or grinding processes. Background Technology

[0002] In actual production processes, robotic polishing and grinding generate a large amount of dust pollutants. Small particles tend to remain suspended in the air, while larger particles settle rapidly due to gravity. Different degrees of pollutant dispersion lead to different levels of harm to the human body and require different dust removal processes. Understanding the dispersion of dust pollutants generated during processing is a fundamental condition for evaluating the degree of hazard and selecting dust collectors. The dispersion of dust pollutants is analyzed by studying the particle size distribution of the dust. Higher particle dispersion means longer airborne time and greater chance of absorption by the human body. Higher concentrations of small dust particles in the working environment and longer exposure times result in more severe harm to the human body. Most of the dust pollutants generated during robotic polishing and grinding are large particles that accumulate in the work area, polluting the environment and requiring timely cleanup. The study of dust pollutant concentration distribution characteristics in metal processing can provide guidance for the research of local dust removal methods. The method of establishing a dust pollutant concentration prediction model by linear regression of the data relationship between processing parameters and dust concentration can only predict the total amount of dust pollutants in the processing process or the average amount of dust in the area divided by direction. This prediction model is not applicable to non-linearly distributed dust pollutants in robotic processing. In the field of robotic processing, robots are used to process large and complex workpieces. During grinding and polishing or grinding, the robot has a large working space and a wide processing trajectory range. The large range of trajectory changes affects the characteristic distribution of dust pollutants in the processing process. Summary of the Invention

[0003] This invention addresses the current lack of effective detection and prediction methods for contaminant generation and distribution during robotic milling and grinding processes by providing a rapid prediction method for contaminant distribution in robotic machining. The invention includes: a method for collecting chips by setting up a chip collection box in the machining area; a statistical prediction method for the microscopic distribution of small particulate contaminants floating above the chip collection box using a laser dust analyzer, showing how the distribution of these small particles varies with process parameters such as machining depth, machining trajectory, and machining time; a statistical prediction method for the weight of large particulate contaminants (chips) by recording the weight of these large particles, showing how the weight varies with process parameters such as machining depth, machining trajectory, and machining time; and a method for extrapolating the distribution of contaminants moving along the machining trajectory from the contaminant distribution characteristics at various fixed positions during robotic machining.

[0004] This invention continuously changes the tool position and collects the particle dust mass concentration, chip mass, and chip quantity at each collection point under different robot machining conditions (spindle speed, machining depth, and robot end effector movement speed) over machining time. These data are recorded in a raw dataset corresponding to each collection point as a data source for sample information. The sample information is used to train a short-term memory network model constructed at each collection point. In actual machining, the particle dust mass concentration, chip mass, and chip quantity at the current moment are predicted based on the raw dataset data of each collection point and the short-term memory network model. On this basis, a machining trajectory is introduced. As the tool center coordinates and machining time change, the particle dust mass concentration, chip mass, and chip quantity predicted at each collection point (the short-term memory network model at each collection point changes with the tool center coordinates and machining time, thus outputting different prediction data) are superimposed with the particle dust mass concentration, chip mass, and chip quantity obtained at previous moments, thereby predicting the distribution of pollutants (including particle dust and chips) at each collection point corresponding to the robot machining trajectory. The specific steps of this invention are as follows:

[0005] S1 mounts the spindle onto the robot's end effector and fixes the tool on the spindle; sets multiple dust collection points; and sets the spindle speed, machining depth, and the movement speed of the robot's end effector.

[0006] The S2 spindle rotates, and the robot's end effector moves the tool a preset distance (a very small distance set to be related to the movement speed) to process the workpiece. Chips are collected by chip collection boxes placed at each dust collection point, and the concentration of small dust particles is monitored in real time by a laser dust meter placed at each dust collection point. The chip collection boxes at all dust collection points are replaced every preset time. The mass of collected chips is obtained by weighing each replaced chip collection box. Then, the chips are dispersed by a vacuum disperser, photographed by microscopy, image processed, contour recognized, and counted to obtain the number of chips. After the number of replacements reaches the set value, the next step is executed.

[0007] S3 continuously changes the movement speed of the end effector within a preset range and repeats step S2;

[0008] S4 continuously changes the processing depth within the preset processing depth range and repeats steps S2 and S3.

[0009] S5 continuously changes the spindle speed within the preset range and repeats steps S2, S3 and S4.

[0010] S6 continuously changes the tool center coordinates within the preset range and repeats steps S2, S3, S4 and S5.

[0011] S7 records the monitoring data of small particulate dust mass concentration, chip mass and chip quantity at each collection point as a function of spindle speed, machining depth, robot end effector movement speed, machining time and tool center coordinates into a raw dataset;

[0012] S8 transforms the data in each original dataset into the data format required by the network model, and then divides it into training and test sets in an 8:2 ratio.

[0013] S9 normalizes each training set X.

[0014] For each collection point, S10 constructs a long short-term memory network model to predict the mass concentration of small particulate dust, chip mass, and chip quantity during robot processing, and trains the long short-term memory network model using the normalized training set X at each collection point.

[0015] S11 performs inverse normalization on the dataset predicted by the Long Short-Term Memory Network model at each collection point, and calculates the error using the predicted dataset and the test set.

[0016] During the actual machining process of the S12 robot, for each tool center location, when the machining time does not exceed the time recorded in the original dataset, the particle dust mass concentration, chip mass, and chip quantity at each collection point directly call the data from the original dataset. When the machining time exceeds the time recorded in the original dataset, the particle dust mass concentration, chip mass, and chip quantity at each collection point are predicted based on a short-term memory network model. As the tool center coordinates and machining time change, the predicted particle dust mass concentration, chip mass, and chip quantity at each collection point are superimposed with the particle dust mass concentration, chip mass, and chip quantity obtained at previous times, thereby predicting the distribution of pollutants (including particle dust and chips) at each collection point corresponding to the robot's machining trajectory.

[0017] Preferably, the material of the chip collection box is ethylene.

[0018] Preferably, the training set In the training set X = 8, the elements in the first row represent the spindle speed at each acquisition time, with a total of m acquisition times; the elements in the second row represent the machining depth at each acquisition time; the elements in the third row represent the movement speed of the robot end effector at each acquisition time; the elements in the fourth row represent the number of chips at each acquisition time; the elements in the fifth row represent the chip mass at each acquisition time; the elements in the sixth row represent the concentration of small particulate dust at each acquisition time; the elements in the seventh row represent the x-coordinate of the tool center at each acquisition time; and the elements in the eighth row represent the y-coordinate of the tool center at each acquisition time. Both the x-axis and y-axis are on a horizontal plane. The training set X is then normalized to obtain... in, x i，min Let x be the minimum value in the i-th row of the training set X. i，max Let X be the maximum value in the i-th row of the training set X. After transformation, X... R All element values ​​are between 0 and 1, thus completing the normalization process.

[0019] Preferably, the Long Short-Term Memory (LSTM) network model includes an input layer, an LSTM layer 1, a Dropout layer 1, an LSTM layer 2, a Dropout layer 2, a Dense layer, and an output layer; the activation function is the ReLU function, which transforms the data transmitted by the LSTM layer 1 into [samples, timesteps, features], where samples represents the number of samples in the data transmitted by the LSTM layer 1, and its value is equal to m; timesteps represents the length of the time window; and features represents the dimension of the data transmitted by the LSTM layer 1, and its value is equal to n.

[0020] Input layer: Input training set X after normalization;

[0021] LSTM Layer 1: Training data on spindle speed, machining depth, robot end effector movement speed, small particle dust mass concentration, chip mass and chip quantity from the input layer; setting the output memory for each time step, with a quantity of 100.

[0022] Dropout layer 1: Randomly discard a portion of the memory data from LSTM layer 1, with the ratio set to 0.2;

[0023] LSTM Layer 2: Training data from memory cells filtered out by Dropout Layer 1, with only the last time step outputting memory cells, and the number of memory cells is 80;

[0024] Dropout layer 2: Randomly discard a portion of the memory data from LSTM layer 2, with the ratio set to 0.2;

[0025] Dense layer: The operation rule is output = activation(dot(input, weight) + bias), where output is the final predicted output, activation is the activation function, dot is the matrix multiplication function, input is the input of the Dense layer, weight is the weight matrix, and bias is the bias value.

[0026] Output layer: Set up 1 neuron, which is directly connected to the Dense layer, and outputs data on the mass concentration of small dust particles, chip mass, and chip quantity.

[0027] Compared with existing technologies, this invention uses a long short-term memory network model to predict the generation and distribution of small particulate dust and chips during robot machining. The method is simple and easy to implement. This invention uses a chip collection box and a laser dust meter to collect dust distribution information in the production area. Based on the collected data, it can predict the pollutant concentration and obtain the predicted pollutant concentration distribution during robot machining under different processing trajectories and processing conditions. This provides guidance for the research of dust removal methods. This invention has high reliability, strong practicality, good effect, and is easy to promote and use. Attached Figure Description

[0028] Figure 1 This is a flowchart of the present invention;

[0029] Figure 2 This is a schematic diagram of the long short-term memory network model architecture in this invention;

[0030] Figure 3 This is a schematic diagram of an LSTM gate structure. Detailed Implementation

[0031] The invention will now be further described with reference to the accompanying drawings.

[0032] like Figure 1 As shown, a method for predicting contaminant distribution during robot processing includes the following steps:

[0033] S1 mounts the spindle onto the robot's end effector and fixes the tool on the spindle; sets multiple dust collection points; and sets the spindle speed, machining depth, and the movement speed of the robot's end effector.

[0034] The S2 spindle rotates, and the robot's end effector moves the tool a preset distance to process the workpiece. Chips are collected by chip collection boxes placed at each dust collection point, and the concentration of small dust particles is monitored in real time by a laser dust meter placed at each dust collection point. The chip collection boxes at all dust collection points are replaced every preset time. The mass of collected chips is obtained by weighing each replaced chip collection box. Then, the chips are dispersed by a vacuum disperser, photographed by microscopy, image processed, contour recognized, and counted to obtain the number of chips. After the number of replacements reaches the set value, the next step is executed.

[0035] S3 continuously changes the movement speed of the end effector within a preset range and repeats step S2;

[0036] S4 continuously changes the processing depth within the preset processing depth range and repeats steps S2 and S3.

[0037] S5 continuously changes the spindle speed within the preset range and repeats steps S2, S3 and S4.

[0038] S6 continuously changes the tool center coordinates within the preset range and repeats steps S2, S3, S4 and S5.

[0039] S7 records the monitoring data of small particulate dust mass concentration, chip mass and chip quantity at each collection point as a function of spindle speed, machining depth, robot end effector movement speed, machining time and tool center coordinates into a raw dataset;

[0040] S8 transforms the data in each original dataset into the data format required by the network model, and then divides it into training and test sets in an 8:2 ratio.

[0041] S9 normalizes each training set X.

[0042] For each collection point, S10 constructs a long short-term memory network model to predict the mass concentration of small particulate dust, chip mass, and chip quantity during robot processing, and trains the long short-term memory network model using the normalized training set X at each collection point.

[0043] The Long Short-Term Memory (LSTM) network model includes an input layer, LSTM layer 1, Dropout layer 1, LSTM layer 2, Dropout layer 2, a Dense layer, and an output layer, as follows: Figure 2As shown; the activation function is the ReLU function, which transforms the data transmitted from the LSTM layer into [samples, timesteps, features]. The samples represent the number of samples in the data transmitted from the LSTM layer, and its value is equal to m. The timesteps represent the length of the time window, and the features represent the dimension of the data transmitted from the LSTM layer, and its value is equal to n.

[0044] Input layer: Input training set X after normalization;

[0045] LSTM Layer 1: Training data on spindle speed, machining depth, robot end effector movement speed, small particle dust mass concentration, chip mass and chip quantity from the input layer; setting the output memory for each time step, with a quantity of 100.

[0046] Dropout layer 1: Randomly discard a portion of the memory data from LSTM layer 1, with the ratio set to 0.2;

[0047] LSTM Layer 2: Training data from memory cells filtered out by Dropout Layer 1, with only the last time step outputting memory cells, and the number of memory cells is 80;

[0048] Dropout layer 2: Randomly discard a portion of the memory data from LSTM layer 2, with the ratio set to 0.2;

[0049] Dense layer: The operation rule is output = activation(dot(input, weight) + bias), where output is the final predicted output, activation is the activation function, dot is the matrix multiplication function, input is the input of the Dense layer, weight is the weight matrix, and bias is the bias value.

[0050] Output layer: Set up 1 neuron, which is directly connected to the Dense layer, and outputs data on the mass concentration of small dust particles, chip mass, and chip quantity.

[0051] The schematic diagrams of the LSTM gate structures in LSTM layer 1 and LSTM layer 2 are as follows: Figure 3 As shown, Xt=[x1,x2,…,x t [x] is the input dataset that includes input data at each time point. t The input data at the current moment, h = [h1, h2, ..., h t [h] represents the output dataset. t The output data at the current moment, C = [C1, C2, ..., C t ] represents the cell state dataset, C t This represents the current state of the cell.

[0052] LSTM layer 1 and LSTM layer 2 use a linear, cumulative approach to process sequence data. The calculation process is as follows:

[0053] f t =σ(w f ·[h t-1 ,x t ]+b f )

[0054] i t =σ(w i ·[h t-1 ,x t ]+b i )

[0055]

[0056]

[0057] o t =ρ(w o ·[h t-1 ,x t ]+b o )

[0058] h t =o t ·tanh(c t )

[0059] Among them, f t This represents the forget gate, σ represents the sigmoid function, and w f The weight matrix h represents the weight of the forget gate. t-1 x represents the output at the previous time step. t b represents the input data at the current moment. f Indicates the bias value of the forget gate, i t Indicates the input gate, w i Let b represent the weight matrix of the input gate. i This represents the deviation value of the input gate. w represents the current input cell state. c The weight matrix representing the current input cell state, b c c represents the deviation value of the current input cell state. t This represents the current cell state, determined by the forget gate f. t Multiply by the cell state c at the previous moment t-1 In addition to the input gate i t Multiply by the current input cell state o t Indicates the output gate, w oLet b represent the weight matrix of the output gate. o h represents the bias value of the output gate. t This represents the final output, which is determined by the output gate O. t Multiply by the current cell state c t The value of the tanh function.

[0060] S11 performs inverse normalization on the dataset predicted by the Long Short-Term Memory Network model at each collection point, and calculates the error using the predicted dataset and the test set.

[0061] During the actual machining process of the S12 robot, for each tool center location, when the machining time does not exceed the time recorded in the original dataset, the particle dust mass concentration, chip mass, and chip quantity at each collection point directly call the data from the original dataset. When the machining time exceeds the time recorded in the original dataset, the particle dust mass concentration, chip mass, and chip quantity at each collection point are predicted based on a short-term memory network model. As the tool center coordinates and machining time change, the predicted particle dust mass concentration, chip mass, and chip quantity at each collection point are superimposed with the particle dust mass concentration, chip mass, and chip quantity obtained at previous times, thereby predicting the distribution of pollutants (including particle dust and chips) at each collection point corresponding to the robot's machining trajectory.

Claims

1. A method of predicting a contamination distribution of a robotic process, the method comprising: The specific steps of the method are as follows: S1 clamps the main shaft on the end effector of the robot, fixes the tool on the main shaft, sets multiple dust collection points, and sets the main shaft speed, machining depth, and movement speed of the end effector of the robot; S2 the main shaft rotates, and the end effector of the robot drives the tool to move a preset distance to process the workpiece; the cuttings are collected through the cuttings collection box placed at each dust collection point, the mass concentration of small particle dust is monitored in real time through the laser dust instrument placed at each dust collection point, and the cuttings collection box at each dust collection point is replaced once every preset time, the mass of the collected cuttings is obtained by weighing each cuttings collection box after replacement, and then the number of cuttings is obtained through the vacuum disperser, microscopic photography, image processing, contour recognition, and quantity statistics; after the number of replacements reaches a set value, the next step is performed; S3 continuously change the movement speed of the end effector within the preset range of the movement speed of the end effector of the robot and repeat step S2; S4 continuously change the machining depth within the preset range of the machining depth and repeat step S2 and step S3; S5 continuously change the main shaft speed within the preset range of the main shaft speed and repeat step S2, step S3, and step S4; S6 continuously change the tool center coordinates within the preset range of the tool center coordinates and repeat step S2, step S3, step S4, and step S5; S7 record the monitoring data of the mass concentration of small particle dust, the mass of cuttings, and the number of cuttings at each collection point with changes in the main shaft speed, the machining depth, the movement speed of the end effector of the robot, the processing time, and the tool center coordinates in an original data set; S8 convert the data in each original data set into a data format required by a network model, and then divide the data into a training set and a test set according to a ratio of 8:2; S9 normalize each training set X; S10 for each collection point, construct a long short-term memory network model for predicting the mass concentration of small particle dust, the mass of cuttings, and the number of cuttings in the robot processing process, and train the long short-term memory network model with the respective normalized training set X of the collection point; S11 perform inverse normalization processing on the data set predicted and output by the long short-term memory network model at each collection point, and calculate the error with the predicted output data set and the test set; S12 in the actual processing process of the robot, for each tool center position, when the processing time does not exceed the time recorded in the original data set, the mass concentration of small particle dust, the mass of cuttings, and the number of cuttings at each collection point directly call the data of the original data set; when the processing time exceeds the time recorded in the original data set, the mass concentration of small particle dust, the mass of cuttings, and the number of cuttings at each collection point are predicted based on the long short-term memory network model at the current time; with the change of the tool center coordinates and the processing time, the predicted mass concentration of small particle dust, the mass of cuttings, and the number of cuttings at the current time at each collection point are respectively superimposed with the mass concentration of small particle dust, the mass of cuttings, and the number of cuttings obtained at each time before the current time, so as to predict the pollutant distribution at each collection point corresponding to the robot processing trajectory.

2. The method of claim 1, wherein: The material of the chip collecting box is ethylene.

3. The method of claim 1, wherein: Training set In n = 8, each element in the first row represents the corresponding value of the main shaft speed at each collection time, and there are m collection times; each element in the second row represents the corresponding value of the machining depth at each collection time, each element in the third row represents the corresponding value of the movement speed of the robot end effector at each collection time, each element in the fourth row represents the corresponding value of the chip quantity at each collection time, each element in the fifth row represents the corresponding value of the chip mass at each collection time, each element in the sixth row represents the corresponding value of the mass concentration of small particle dust at each collection time, each element in the seventh row represents the corresponding value of the horizontal coordinate of the tool center at each collection time, and each element in the eighth row represents the corresponding value of the vertical coordinate of the tool center at each collection time, wherein the horizontal coordinate axis and the vertical coordinate axis are on the horizontal plane; the training set X is normalized to obtain wherein, x i，min is the minimum value of the i-th row in the training set X, and x i，max is the maximum value of the i-th row in the training set X.

4. A method of predicting contamination distribution for a robotic process according to claim 1, 2 or 3, characterized in that: The long short-term memory network model comprises an input layer, an LSTM layer one, a Dropout layer one, an LSTM layer two, a Dropout layer two, a Dense layer and an output layer; an activation function is a ReLu function, data transmitted by the LSTM layer one is converted into [samples, timesteps, features], samples represent a sample quantity in the data transmitted by the LSTM layer one, the value is equal to m, timesteps represent a time window length, and features represent a data dimension of the LSTM layer one, the value is equal to n; The input layer: input normalized data of the training set X; The LSTM layer one: train spindle speed, machining depth, movement speed of a robot end effector, small particle dust mass concentration, chip mass and chip quantity data from the input layer, set each time step output memory, and the quantity is 100; The Dropout layer one: randomly discard part of the memory data from the LSTM layer one, and the ratio is set to 0.2; The LSTM layer two: train the memory data selected from the Dropout layer one, set only the last time step output memory, and the quantity is 80; The Dropout layer two: randomly discard part of the memory data from the LSTM layer two, and the ratio is set to 0.2; The Dense layer: the operation rule is output = activation(dot(input, weight) + bias), wherein, output is a final output prediction result, activation is an activation function, dot is a matrix multiplication operation function, input is the input of the Dense layer, weight is a weight matrix, and bias is a bias value; The output layer: set 1 neuron, which is directly connected with the Dense layer, and output small particle dust mass concentration, chip mass and chip quantity data.

Images