[0038] The present invention will be further explained below in conjunction with the drawings.
[0039] Such as figure 1 , 2 As shown in and 3, the mechanical grinding condition detection system based on neural network of the present invention includes a sensor information acquisition module, a neural network learning module and a neural network operation output module.
[0040] The sensor information collection module is an acoustic emission sensor, which mainly collects various acoustic emission signals during processing according to a preset sampling frequency.
[0041] The neural network learning module is used to determine the weight threshold parameter of the neural network and pass it to the neural network operation output module for use. In the learning and training stage, the sample library is formed by collecting the acoustic emission signal when the tool is in contact with the processed workpiece, the acoustic emission signal when the tool (grinding wheel) is passivated, the acoustic emission signal when the workpiece is burned, and the crack, etc., and the neural network learning module After learning and training, the weight threshold parameters of the neural network are obtained, and the parameters are passed to the neural network operation output module.
[0042] The neural network operation output module is used for judging the working status of the workpiece and the tool in the current mechanical grinding process. After the parameters of the neural network are determined, this module receives the real-time data from the sensor information acquisition module, uses formula (1), formula (2) and formula (3) to calculate whether the tool is in contact with the workpiece, and the tool (grinding wheel) Information on working conditions such as whether passivation, whether the workpiece is burnt, or whether there is a crack.
[0043] (1) Collect the acoustic emission information of the processing site under standard working conditions: first set the system in the training and learning mode. During grinding, the frequency of the acoustic emission signal of the structure-borne sound is generally in the range of 50K~300K. Therefore, the sampling frequency f of the acoustic emission sensor is set to 1MHz; because the sampling frequency f of the acoustic emission sensor is greater than or equal to twice the grinding Only by processing the frequency of the structure-borne acoustic emission signal can the on-site acoustic emission signal be collected effectively; the acoustic emission sensor collects the on-site acoustic emission signal in real time according to the preset sampling frequency f.
[0044] The stated standard operating conditions are: the contact parameter Y(1) is 1 when the workpiece and the tool are in contact, and 0 when the workpiece and the tool are not in contact, Y(1) is a Boolean variable type; the tool passivation coefficient Y(2) is When the tool has just been trimmed, it is 0. During the use of the tool, the passivation coefficient gradually increases until it reaches 100. At this time, the tool needs to be trimmed. Y(2) is an integer variable type; the burn parameter Y(3) is when the workpiece is burned. 1. When the workpiece is not burned, it is 0, Y(3) Boolean variable type; the crack parameter Y(4) is 1 when the workpiece has cracks, and 0 when the workpiece has no cracks, Y(4) Boolean variable type.
[0045] The acoustic emission signal sequence is marked with X vector, X 1 Means t 0 The first sampling point starts at time, Xi is the i-th sampling point, i=1 to N, and N is the length of the sampling sequence. In the case of a given sampling frequency, the sampling sequence length N is closely related to the response time of the system to the judgment of working conditions. The larger the N, the longer the response time of the system; the smaller the N, the shorter the response time of the system to working condition judgment. However, choosing a smaller length N may affect the accuracy of the neural network's judgment of working conditions. When the sampling frequency is 1 MHz, in order to make the minimum accurate response time of this method in the order of milliseconds, the value of N is 1000. The sampling time of the acoustic emission signal vector X is 1ms, plus the neural network operation time (microsecond level), to ensure that the system can respond correctly in milliseconds after the working conditions change.
[0046] (2) Start the grinder to perform grinding processing, repeatedly contact and disengage the tool and the workpiece; start the system to collect the acoustic emission signal X when the tool is in contact with the workpiece, and collect 100 sets of data for use.
[0047] Grind the tool to a sharp state and set the passivation coefficient of this system to 0. Start the grinding process of the grinder to process the workpiece. Collect 100 sets of acoustic emission signals X during the machining process, covering the data from the sharp to the dull of the tool.
[0048] The operator selects the flawed and burned workpieces for grinding, and collects 100 sets of acoustic emission signals X for standby during the processing.
[0049] The operator selects the cracked workpiece for grinding, and collects 100 sets of acoustic emission signals X for standby during the processing.
[0050] In order to ensure that the neural network can stably express the functional relationship between the input variable and the output variable, it is necessary to collect multiple sample groups to perform the above-mentioned operations to calculate and revise the weight threshold of the neural network. According to experiments, when the number of samples is greater than 50, after the above training and learning, the neural network established by this method can accurately calculate the output conditions.
[0051] (3) Establish a neural network model: the selected neural network model is a three-layer network: input layer, hidden layer and output layer,
[0052] The input excitation function of the neural network input layer is a Sjgmoid type excitation function:
[0053] f ( X ) = 1 1 + e - X / Q ......Formula 1)
[0054] Among them, Q is the Sigmoid parameter and X is t 0 The N-dimensional vector of the sampling sequence at time, the i-th variable of the vector X is represented by Xi, i=1~N;
[0055] The hidden layer parameter h of the neural network:
[0056] h j = f ( X i = 1 N W ij I X i - θ j ) ...Equation (2)
[0057] Where W ij I Is the weight coefficient from the i-th input variable to the j-th hidden layer node, θ j Is the output threshold of each unit of the hidden layer;
[0058] X is t 0 The N-dimensional vector of the sampling sequence at time, the i-th variable of the vector X is represented by Xi, i=1~N;
[0059] i is the number of input variables, i=1~N;
[0060] j is the number of hidden layer nodes, and j=2*N in this embodiment;
[0061] The output Y of this neural network is:
[0062] Y k = f ( X i = 1 k W jk O h j - γ k ) ...Equation (3)
[0063] Where W jk o Is the weight coefficient from the jth hidden layer node to the kth output node; γ k Is the output threshold of each unit of the output layer; k is the number of output variables, in this method, k = 1 to 4; Y is t 0 The working condition vector at time, a 4-dimensional vector, the i-th variable of vector Y is represented by Yi, i=1 to 4; that is, Y contains four variables, and the first variable Y(1) is a contact parameter, a Boolean variable type, It is 1 when the workpiece and the tool are in contact, and 0 when the workpiece and the tool are not in contact; the second element Y(2) is the tool passivation coefficient, which is an integer variable type. When the tool has just been repaired, the passivation coefficient is 0. During the use of the tool, the passivation coefficient gradually increases until it reaches 100, at which time the tool needs to be repaired; the third variable Y(3) is the burn parameter, a Boolean variable type, when the workpiece is burned, it is 1, when the workpiece is not burned When the time is 0; the fourth variable Y(4) is a crack parameter, a Boolean variable type, which is 1 when the workpiece has a crack, and 0 when the workpiece has no crack.
[0064] (4) Learning and training: Collect the above 4*100 sets of data, select the first 4*50 sets of data as training samples, and input the neural network learning module for learning. Using the error back propagation algorithm (BP algorithm) proposed by Rumehart et al. in 1985 to calculate the parameters of the neural network W ij I , W jk o , Θ, γ, the above parameters are called the weight threshold of the neural network. The calculated weight threshold is passed to the neural network operation output module.
[0065] The error function defined in the learning process of the neural network is
[0066] E = 1 2 X k = 1 N ( Y k - Y ^ k ) 2 ...Equation (4)
[0067] Where Y k Is the actual output of the network, that is, the known operating condition Y; It is the working condition calculated by inputting X vector through formula (1), formula (2) and formula (3);
[0068] Set initial W ij I And W jk o Is a unit matrix, θ and γ are random numbers from 0 to 1, and the training samples composed of the collected sample data X and the corresponding Y are input into the neural network learning module in pairs; calculated by formula (4) Error E, and correct the weight threshold; the correction formula is:
[0069] W = W - μ ∂ E ∂ W ...Equation (5)
[0070] Among them, W is the weight threshold parameter matrix, μ is the correction step size, 0
[0071] Substitute the revised weight threshold into equations (1), (2), (3), and (4), calculate the error E, and repeat until E is less than 0.001; when E is less than 0.001, it is called neural The network converges, and the weight threshold at this time can be used by the neural network operation output module.
[0072] Select the last 4*50 sets of data of 4*100 sets of data, input the neural network operation output module to calculate, output the working conditions, and compare with the actual working conditions of the first 4*50 sets of data, the correct rate of working condition judgment is 99 % Above, it proves that the neural network is stable and reliable.
[0073] (5) After the system is trained and verified, the grinding process is started, and the system starts to collect the acoustic emission signal X at time t at the processing site t ,Send into neural network operation, output working condition vector Y t.
[0074] The neural network operation output module is used to judge the working status of the workpiece and the tool in the current mechanical grinding process, that is, after the neural network weight threshold parameter is determined, the neural network operation output module can be transmitted according to the received sensor information collection module Real-time data vector X t , Use formula (1), formula (2) and formula (3) to calculate the output vector Y t.
[0075] (6) According to the vector Y t The definition of the internal variables can determine whether the tool is in contact with the processed workpiece, the passivation of the tool, whether the workpiece is burned, and whether there is a crack. By querying the working condition vector Y t Corresponding variables of, you can know whether the tool is in contact with the workpiece, the degree of passivation of the tool (grinding wheel), whether the workpiece is burned or whether there is a crack in the workpiece during the current grinding process.
[0076] (7) The control system of the grinder or the operator can prompt for further operations according to the judgment of the system. For example, start to calculate the tool feed, change the tool, change the workpiece, etc. Its operation is not within the scope of the present invention.
