Electric arc stability online detection method based on target optimal neural network

Abstract

The invention discloses an arc stability online detection method based on a target optimal neural network. The method comprises the following steps: establishing a general model of an input and output mapping relation of a rotating arc stability sensor and determining important parameters of the general model; a parameter adaptive model establishment method based on a target optimal neural network comprises three parts of contents: identifying the variation of important parameters of a general model of a rotating arc stability sensor based on the neural network, automatically adjusting the weight of the neural network based on a target optimal criterion, and establishing a parameter adaptive model based on the target optimal criterion. Establishing a parameter adaptive model of the rotating arc stability sensor based on the target optimal neural network; the invention discloses an arc stability online detection method based on a self-adaptive model. According to the method, the arc stability can be accurately detected, the accuracy of arc stability detection and control can be improved, the welding quality can be improved, and certain economic benefits can be generated.

Description

Technical Field

[0001] The invention relates to the field of electronic information technology, and in particular to an arc stability online detection method based on a target optimal neural network. Background Art

[0002] The actual welding system often works in an uncertain and changeable working environment, which causes the model parameters to change and makes the general model unable to adapt well to the uncertain or changeable working conditions. The output of the general model of the system is an approximate value of the arc stability evaluation index. In order to improve the accuracy of online detection of arc stability, it is necessary to invent a method for online establishment of an adaptive model of the arc stability detection system that adapts to uncertain working conditions, so as to realize online self-tuning of the model parameters according to the working conditions, improve the accuracy of online detection of arc stability, and help improve the rotating arc sensor weld tracking accuracy and weld formation quality, which has practical application value. Summary of the invention

[0003] In order to solve the problems existing in the prior art, the present invention provides an online detection method for arc stability based on a target optimal neural network, a method for online establishing an adaptive model of an arc stability detection system that adapts to uncertain working conditions, to achieve online self-tuning of model parameters according to the working conditions, and further to achieve accurate detection of arc stability based on the adaptive model, which helps to improve the quality of automatic welding of machines.

[0004] To achieve this object, the present invention adopts the following technical solutions:

[0005] The invention discloses an arc stability online detection method based on a target optimal neural network, comprising: establishing a general model of the input-output mapping relationship of a rotating arc stability sensor and determining its important parameters, a parameter adaptive model establishment method based on the target optimal neural network, and an arc stability online detection method based on the adaptive model.

[0006] Furthermore, the general model of the mapping relationship between the input and output of the rotating arc stability sensor is established and its important parameters are determined. The specific contents are as follows: the general model Z() of the mapping relationship between the input and output of the rotating arc stability sensor satisfies the following formula:

[0007] w=Z(η(i),β(i),α(i),γ(i),α max ,γ max ,a,b,c,d,σ) (1)

[0008] In the formula, η(i), β(i), α(i) and γ(i) are four characteristic quantities of the rotating arc sensor signal. The characteristic quantity η(i) can describe the number of times the arc is broken or close to being broken, the characteristic quantity β(i) can describe the number of changes in the direction of the sampled welding current in the same circle of rotating arc, the characteristic quantity α(i) can describe the changes in the magnitude of adjacent sampled welding currents in the same circle of rotating arc, and the characteristic quantity γ(i) can describe the changes in the magnitude of sampled welding currents at the same position of two adjacent circles of rotating arc. These four characteristic quantities can be calculated from the collected rotating arc sensor signal. Therefore, the parameter α max , γ max The values ​​of , a, b, c, d, and σ determine the accuracy of the general model of the input-output mapping relationship of the rotating arc stability sensor. max is the value of α(i) when the arc is unstable, γ max is the value of γ(i) when the arc is unstable, σ is the standard deviation of the Gaussian function, a is the arc stability point, b is the arc under-stability point, c is the point corresponding to the beginning of arc instability, and d is the arc instability point. After theoretical analysis and experimental verification, it is found that the values ​​of parameters c, d and σ satisfy c = 1, d = 2 and σ = 0.4. w is the identified comprehensive evaluation index of arc stability.

[0009] Since the parameter α max and γ max The accuracy of arc stability recognition is very sensitive to the arc stability point a and arc instability point b. Therefore, parameter α max , γ max , a and b are four important variable parameters that determine the precise model of the input-output mapping relationship of the rotating arc stability sensor, and affect the accuracy of identifying arc stability.

[0010] Furthermore, the arc stability online detection scheme based on the target optimal neural network is specifically as follows: first, the target optimal criterion is designed, and the parameter adaptive model establishment method based on the target optimal online self-learning neural network is studied. Then, the parameters of the general model of the rotating arc stability sensor established previously are adjusted online, and an adaptive model of arc stability and rotating arc sensor signal based on the target optimal neural network is established, and the sensitivity of the arc stability evaluation index to the characteristic quantity of the rotating arc sensor signal is calculated. Finally, the arc stability online detection method based on the adaptive model is studied, and the arc stability online detection experiment is carried out to identify the comprehensive evaluation index of arc stability.

[0011] Furthermore, the method for establishing a parameter adaptive model based on the target optimal neural network is specifically as follows: the output of the general model of the system is an approximate value of the arc stability evaluation index, and the adaptive model based on the target optimal neural network can make the identified arc stability evaluation index more accurate. If the working conditions and welding parameters change, the parameters of the arc stability and rotating arc sensor signal mapping relationship model will change. Taking the change in the characteristic quantity of the arc sensor signal as input, the target optimal criterion is designed, and based on the target optimal neural network, the parameters of the general model of the system are adjusted, and finally the online adaptive detection of arc stability is realized through the adaptive model.

[0012] The method for establishing a parameter adaptive model based on a target optimal neural network mainly includes three parts: identifying the change in important parameters of a general model of a rotating arc stability sensor based on a neural network, automatically adjusting the weights of the neural network based on a target optimal criterion, and establishing a parameter adaptive model of a rotating arc stability sensor based on the target optimal neural network.

[0013] Furthermore, the neural network-based identification of the changes in the important parameters of the general model of the rotating arc stability sensor is as follows: the changes in the characteristic quantities of the arc sensor signal (Δη(i), Δβ(i), Δα(i), Δγ(i)) are used as input, and the changes in the important parameters of the general model of the rotating arc stability sensor (Δα max ,Δγ max ,Δa,Δb) is the output.

[0014] The total input Y of the mth neuron in the hidden layer is m (i) satisfies the following equation,

[0015]

[0016] In the formula, w k,m (i) is the connection weight from the kth neuron in the input layer to the mth neuron in the hidden layer, X k (i) is the input of the kth neuron in the input layer, m is the serial number of the hidden layer neuron, and m = 1, 2, ..., 6. Δη(i) is the change of η(i), Δβ(i) is the change of β(i), Δα(i) is the change of α(i), and Δγ(i) is the change of γ(i).

[0017] The Sigmoid function is used as the activation function h() of the hidden layer neurons, and the output Y of the i-th neuron of the m-th hidden layer is m '(i) satisfies the following formula,

[0018]

[0019] Where Ym (i) is the total input of the i-th neuron in the hidden layer.

[0020] The actual output Z of the nth neuron in the output layer is n (i) satisfies the following equation,

[0021]

[0022] We can get four outputs of the neural network.

[0023]

[0024] The output error e of the i-th output of the n-th neuron in the output layer n (i) satisfies the following equation,

[0025]

[0026] In formula (4-6), w m,n (i) is the connection weight from the mth neuron in the hidden layer to the nth neuron in the output layer, Y m '(i) is the output of the i-th neuron of the m-th hidden layer, and n is the serial number of the neuron in the output layer. Δα max (i) is the general model parameter α max The change in (i), Δγ max (i) is the parameter γ max (i), Δa(i) is the i-th change in parameter a, and Δb(i) is the i-th change in parameter b. n,0 (i) is the ideal output of the nth neuron in the output layer, Z n,0 (i) = (Δα max , Δγ max , Δa, Δb). n (i) is the actual output of the i-th neuron in the output layer.

[0027] Furthermore, the weights of the neural network are automatically adjusted based on the target optimal criterion. Specifically, the target optimal criterion needs to be designed to realize the automatic adjustment of the parameters of the general model of the sensor system based on the target optimal neural network. After analyzing the relationship between the change in the characteristic quantity of the arc sensor signal and the change in the important parameters of the general model, the target optimal criterion can be determined, which is divided into three categories: the first type of rule indicates that there is a broken arc or a situation close to a broken arc; the second type of rule indicates that although there is no broken arc or a situation close to a broken arc, the arc stability has changed in general intensity; the third type of rule indicates that the arc stability has not changed within the error range. The following will explain one case of the first type of rule, the second type of rule, and the third type of rule, respectively, and the meanings of other rules can be inferred by analogy.

[0028] A case of the first type of rule: when the number of arc breaking or near arc breaking increases by 64 times, the number of changes in the direction of the sampled welding current in the same circle of rotating arc increases by 20 times, the change in the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc increases by 1, and the change in the characteristic quantity describing the change in the magnitude of the sampled welding current at the same position of two adjacent circles of rotating arc increases by 1, indicating that the arc stability becomes very poor at this time. At this time, the maximum value α of the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc max The change Δα of (i) max (i) To add 3 levels, the maximum value γ of the characteristic quantity describing the change of the sampling welding current at the same position of two adjacent rotating arcs max The change Δγ of (i) max (i) should also be increased by 3 levels. As the number of arc breaking increases, the arc instability increases. Therefore, the i-th change Δa(i) of the arc stable point a and the i-th change Δb(i) of the arc instability point b should be reduced by 3 levels to make the current arc stability state farther away from the arc stable point and the arc instability point.

[0029] A case of the second type of rule: when the number of arc breaking or near arc breaking increases by 0 times, the number of changes in the direction of the sampled welding current in the same circle of rotating arc decreases by 1 time, the change in the characteristic quantity describing the change in the magnitude of adjacent sampled welding currents in the same circle of rotating arc decreases by 0.25, and the change in the characteristic quantity describing the change in the magnitude of the sampled welding currents at the same position of two adjacent circles of rotating arcs decreases by 0.25, indicating that there is no arc breaking or near arc breaking at this time. At this time, the maximum value α of the characteristic quantity describing the change in the magnitude of adjacent sampled welding currents in the same circle of rotating arc max The change Δα of (i) max (i) To reduce the level by 1, the maximum value γ of the characteristic quantity describing the change of the sampling welding current at the same position of two adjacent rotating arcs max The change Δγ of (i) max (i) should also be reduced by 1 level. As the arc stability is developing in the direction of improvement, the i-th change amount Δa(i) of the arc stable point a and the i-th change amount Δb(i) of the arc instability point b should both be increased by 1 level to make the current arc stability state closer to the arc stable point and the arc instability point.

[0030] A case of the third type of rule: when the number of arc breaking or near arc breaking increases by 0 times, the number of changes in the direction of the sampled welding current in the same circle of rotating arc increases by 0 times, the change in the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc changes within ±0.1, and the change in the characteristic quantity describing the change in the magnitude of the sampled welding current at the same position of two adjacent circles of rotating arc changes within ±0.1. At this time, the arc stability does not change within the error range. Therefore, the maximum value α of the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc max The change Δα of (i) max (i) unchanged, the maximum value γ of the characteristic quantity describing the change of the sampling welding current at the same position of two adjacent rotating arcs max The change Δγ of (i) max (i) remains unchanged, the i-th change Δa(i) of the arc stable point a and the i-th change Δb(i) of the arc instable point b remain unchanged.

[0031] The target optimal criterion is used as the objective function of the neural network, and the gradient descent method is used to adjust the connection weights of neurons in each layer of the neural network.

[0032] Furthermore, the parameter adaptive model of the rotating arc stability sensor is established based on the target optimal neural network. The specific content is as follows: the target optimal neural network is trained using the target optimal criterion. When the error corresponding to each optimal relationship criterion is very small, the connection weight w from the input layer neuron to the hidden layer neuron is 1 , the connection weights w between the hidden layer neurons and the output layer neurons 2 Satisfy the following formulas respectively,

[0033]

[0034]

[0035] The identified connection weights w 1 and 2 Substituting into the target optimal neural network, the change of important parameters of the general model of the rotating arc stability sensor can be identified, and considering the influence of the previous model parameters on the current adaptive model parameter identification, it can be identified that the adaptive parameters of the rotating arc stability sensor model satisfy the following formula:

[0036]

[0037] In the formula, (α max (i),γ max (i), a(i), b(i)) are the adaptive parameters of the rotating arc stability sensor model identified for the i-th time, (Δα max(i),Δγ max (i), Δa(i), Δb(i)) are the important parameters of the general model of the rotating arc stability sensor (α max ,γ max ,a,b),λ 3 , 4 , 5 , 6 and 7 is the weighting coefficient.

[0038] The identified adaptive parameters are used to replace the parameters of the general model of the rotating arc stability sensor, and a parameter adaptive model of the rotating arc stability sensor based on the target optimal neural network can be established. The parameter adaptive model establishment method based on the target optimal neural network studied in the present invention can realize online self-tuning of model parameters according to working conditions, thereby online establishing an adaptive model of the rotating arc stability sensor that adapts to uncertain working conditions.

[0039] Furthermore, the online detection method of arc stability based on the adaptive model is specifically as follows: taking the four characteristic quantities (η(i), β(i), α(i), γ(i)) of the i-th rotating arc sensor signal as input and utilizing the adaptive model of the rotating arc stability sensor, the i-th identified arc stability comprehensive evaluation index w(i) can be output.

[0040] By using the relationship between the arc length change coefficient K(i) and the four characteristic quantities of the rotating arc sensor signal, it is possible to calculate that the sensitivity of the arc stability comprehensive evaluation index to the four characteristic quantities of the rotating arc sensor signal satisfies the following formula:

[0041]

[0042] In the formula, k η (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity η(i), k β (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity β(i), k α (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity α(i), k γ (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity γ(i). w(i) is the arc stability comprehensive evaluation index identified for the i-th time, K(i) is the arc length change coefficient identified for the i-th time, α max (i) and γ max (i) is the adaptive parameter of the rotating arc stability sensor model identified for the i-th time using equation (9). 1 , w 2 , w 3 and 4are the weights of the four characteristic quantities of the rotating arc sensor signal. n represents the number of filter currents used to determine the number of times the sampling welding current direction changes in the same circle of rotating arc, and its value is 20.

[0043] When the welding parameters are determined, the arc stability generally does not change suddenly, or when the welding parameters are continuously adjusted, the welding arc stability also changes gradually. Therefore, in order to reduce the interference of noise, the median filtering method is used to filter out the arc stability comprehensive evaluation index w 0 (i) satisfies the following equation,

[0044] w 0 (i)=median[w(i),w(i-1),w(i-2),w(i-3),w(i-4)] (11)

[0045] Where w(i) is the comprehensive evaluation index of arc stability identified for the i-th time, and median[] is the median function.

[0046] The beneficial effects of the present invention are:

[0047] The advantage of the present invention is that an online adaptive detection method for the arc stability of a rotating arc sensor corner welding is invented based on a target optimal neural network. Through the content of the present invention, the model parameters can be self-tuned online according to the working conditions, the accuracy of the online detection of the arc stability can be improved, the accuracy of the arc stability detection and control can be improved, and it is helpful to improve the rotating arc sensor weld tracking accuracy and the weld forming quality, thereby helping to improve the quality of welding, having practical application value, and can generate certain economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A general model of the input-output mapping relationship of the rotating arc stability sensor of the present invention;

[0049] Figure 2 The invention provides a rotating arc sensing corner welding arc stability online detection solution based on a target optimal neural network;

[0050] Figure 3 A method for establishing a parameter adaptive model based on a target optimal neural network of the present invention;

[0051] Figure 4 The optimal neural network for identifying the change of important parameters of the general model of the rotating arc stability sensor of the present invention;

[0052] Figure 5 The robot of the present invention tracks the straight fillet weld after welding;

[0053] Figure 6 The present invention identifies Figure 5 Corresponding arc stability and welding deviation change curves;

[0054] Figure 7 The robot of the present invention tracks the fillet weld that bends outward after welding;

[0055] Figure 8 The present invention identifies Figure 7 The corresponding arc stability and welding deviation change curves. DETAILED DESCRIPTION

[0056] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0057] The invention discloses an arc stability online detection method based on a target optimal neural network, comprising: establishing a general model of the input-output mapping relationship of a rotating arc stability sensor and determining its important parameters, a parameter adaptive model establishment method based on the target optimal neural network, and an arc stability online detection method based on the adaptive model.

[0058] A general model of the input-output mapping relationship of the rotating arc stability sensor is established and its important parameters are determined. The specific contents are as follows:

[0059] The working principle of the rotating arc stability sensor is as follows Figure 1 As shown, the arc length self-regulation system is used as a sensitive element to convert the arc stability into the arc length self-regulation capability; the rotating arc is used as a conversion element to modulate the arc length self-regulation capability into an arc length recovery value; the welding circuit and the data acquisition and processing system are used as a conversion circuit and an output unit respectively, so that the arc length recovery value can be converted into a rotating arc sensor signal.

[0060] like Figure 1 As shown in the figure, the arc length change coefficient K(i) is used as the evaluation index of the arc length recovery value, and the general model of welding circuit, data acquisition and processing system K(i)=R(η(i),β(i),α(i),γ(i)) is established. 1 , w 2 , w 3 and 4 are the weights of the four features, and their values ​​satisfy w 1 =w 2 =1 and w 3 =1-w 4=0.8. n represents the number of filter currents used to determine the number of changes in the direction of the sampled welding current in the same rotating arc. In the present invention, its value is 20. In the model, η(i), β(i), α(i) and γ(i) can be calculated from the collected rotating arc sensor signal. In addition, in the general model, α max is the value of the characteristic quantity α(i) when the arc is unstable, γ max is the value of the characteristic quantity γ(i) when the arc is unstable. However, α max With γ max The value of is uncertain and will change with the change of welding parameters and working conditions. max With γ max These are two important parameters that determine the general model of the conversion circuit and the output unit.

[0061] The arc length self-regulation ability linguistic variable T contains three fuzzy subsets: STRONG means the arc length self-regulation ability is strong, COMMON means the arc length self-regulation ability is average, and POOR means the arc length self-regulation ability is poor. Figure 1 As shown in the figure, the relationship model between the arc length change coefficient and the arc length self-regulation ability satisfies T = g (K (i) ), a, b, c, d and σ are the five important parameters that determine the relationship model.

[0062] like Figure 1 As shown, the arc stability linguistic variable W satisfies W = (WELL, GENERAL, BAD), where WELL indicates arc stability, GENERAL indicates arc instability, and BAD indicates arc instability. Taking the current arc length self-adjustment capability T(i) and the previous arc length self-adjustment capability T(i-1) as input, the current arc stability can be inferred by using the fuzzy relationship between the arc length self-adjustment capability and arc stability, and then the centroid method is used for defuzzification to identify the arc stability comprehensive evaluation index w that matches the current welding state.

[0063] In summary, the general model Z() of the input-output mapping relationship of the rotating arc stability sensor satisfies the following formula:

[0064] w=Z(η(i),β(i),α(i),γ(i),α max ,γ max ,a,b,c,d,σ) (1)

[0065] In the formula, η(i), β(i), α(i) and γ(i) are four characteristic quantities of the rotating arc sensor signal. The characteristic quantity η(i) can describe the number of times the arc is broken or close to being broken, the characteristic quantity β(i) can describe the number of changes in the direction of the sampled welding current in the same circle of rotating arc, the characteristic quantity α(i) can describe the changes in the magnitude of adjacent sampled welding currents in the same circle of rotating arc, and the characteristic quantity γ(i) can describe the changes in the magnitude of sampled welding currents at the same position of two adjacent circles of rotating arc. These four characteristic quantities can be calculated from the collected rotating arc sensor signal. Therefore, the parameter α max , γ max The values ​​of , a, b, c, d, and σ determine the accuracy of the general model of the input-output mapping relationship of the rotating arc stability sensor. max is the value of α(i) when the arc is unstable, γ max is the value of γ(i) when the arc is unstable, σ is the standard deviation of the Gaussian function, a is the arc stability point, b is the arc under-stability point, c is the point corresponding to the beginning of arc instability, and d is the arc instability point. w is the identified comprehensive evaluation index of arc stability.

[0066] Since the parameter α max and γ max It will change with the change of welding parameters and welding conditions, thus affecting the accuracy of subsequent estimation of other parameters. In addition, a = 0.35, b = 0.8 and d = 2, the distance from the arc under-stability point b to the arc unstable point d is 2.7 times the distance from the arc under-stability point b to the arc stable point a, so the accuracy of arc stability identification is very sensitive to parameters a and b.

[0067] Therefore, only by increasing the parameter α max , γ max Only by improving the accuracy of the estimated values ​​of a, a and b can we establish an accurate model of the input and output mapping relationship of the rotating arc stability sensor and further improve the recognition accuracy of arc stability. The following will further study the method of automatically adjusting these four key parameters online according to the welding status.

[0068] The arc stability online detection scheme based on the target optimal neural network is as follows:

[0069] like Figure 2As shown in the figure, first, the target optimal criterion is designed, and the parameter adaptive model establishment method based on the target optimal online self-learning neural network is studied. Then, the parameters of the general model of the rotating arc stability sensor established previously are adjusted online, and an adaptive model of arc stability and rotating arc sensor signal based on the target optimal neural network is established, and the sensitivity of the arc stability evaluation index to the characteristic quantity of the rotating arc sensor signal is calculated. Finally, the online detection method of arc stability based on the adaptive model is studied, and the online detection experiment of arc stability is carried out to identify the comprehensive evaluation index of arc stability.

[0070] The parameter adaptive model establishment method based on the target optimal neural network is as follows:

[0071] The output of the general model of the system is an approximate value of the arc stability evaluation index, while the adaptive model based on the target optimal neural network can make the identified arc stability evaluation index more accurate. If the working conditions and welding parameters change, the parameters of the arc stability and rotating arc sensor signal mapping relationship model will change. Figure 3 As shown, the change in the characteristic quantity of the arc sensor signal is taken as input, the target optimal criterion is designed, and the parameters of the general model of the system are adjusted based on the target optimal neural network. Finally, the online adaptive detection of arc stability is realized through the adaptive model.

[0072] (1) Identify the changes in important parameters of the general model of the rotating arc stability sensor based on the neural network

[0073] like Figure 4 As shown in the figure, the change of the characteristic value of the arc sensor signal (Δη(i), Δβ(i), Δα(i), Δγ(i)) is used as input, and the change of the important parameters of the general model of the rotating arc stability sensor (Δα max ,Δγ max ,Δa,Δb) is the output.

[0074] The total input Y of the mth neuron in the hidden layer is m (i) satisfies the following equation,

[0075]

[0076] In the formula, w k,m (i) is the connection weight from the kth neuron in the input layer to the mth neuron in the hidden layer, X k (i) is the input of the kth neuron in the input layer, m is the serial number of the hidden layer neuron, and m = 1, 2, ..., 6. Δη(i) is the change of η(i), Δβ(i) is the change of β(i), Δα(i) is the change of α(i), and Δγ(i) is the change of γ(i).

[0077] The Sigmoid function is used as the activation function h() of the hidden layer neurons, and the output Y of the i-th neuron of the m-th hidden layer is m '(i) satisfies the following formula,

[0078]

[0079] Where Y m (i) is the total input of the i-th neuron in the hidden layer. The meanings of other physical quantities have been introduced in formula (2).

[0080] The actual output Z of the nth neuron in the output layer is n (i) satisfies the following equation,

[0081]

[0082] We can get four outputs of the neural network.

[0083]

[0084] The output error e of the i-th output of the n-th neuron in the output layer n (i) satisfies the following equation,

[0085]

[0086] In formula (4-6), w m,n (i) is the connection weight from the mth neuron in the hidden layer to the nth neuron in the output layer, Y m '(i) is the output of the i-th neuron of the m-th hidden layer, and n is the serial number of the neuron in the output layer. Δα max (i) is the general model parameter α max The change in (i), Δγ max (i) is the parameter γ max (i), Δa(i) is the i-th change in parameter a, and Δb(i) is the i-th change in parameter b. n,0 (i) is the ideal output of the nth neuron in the output layer, Z n,0 (i) = (Δα max , Δγ max , Δa, Δb). n (i) is the actual output of the i-th neuron in the output layer.

[0087] (2) Automatically adjust the weights of the neural network based on the target optimal criterion

[0088] like Figure 3As shown in Figure 1, to achieve automatic adjustment of the parameters of the general model of the sensor system based on the target optimal neural network, it is necessary to design the target optimal criterion. By analyzing the relationship between the change in the characteristic quantity of the arc sensor signal and the change in the important parameters of the general model, the target optimal criterion can be determined, as shown in Table 1. There are 11 rules in Table 1, which are divided into three categories: the first category of rules includes rules 1-4, indicating that there is a situation of arc breaking or near arc breaking; the second category of rules includes rules 5-8, indicating that although there is no arc breaking or near arc breaking, the arc stability has changed in general intensity; the third category of rules includes rules 9-11, indicating that the arc stability has not changed within the error range. The following will explain rule 1 (belonging to the first category of rules), rule 7 (belonging to the second category of rules), and rules 9-11 (the third category of rules), and the meanings of other rules can be inferred by analogy.

[0089] Table 1 Optimal relationship criteria between the change of arc sensor signal characteristic quantity and the change of important parameters of general model

[0090]

[0091] Rule 1: When the number of arc breaking or arc near breaking increases by 64 times, the number of changes in the direction of the sampled welding current in the same rotating arc increases by 20 times, the change in the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same rotating arc increases by 1, and the change in the characteristic quantity describing the change in the magnitude of the sampled welding current at the same position of two adjacent rotating arcs increases by 1, indicating that the arc stability becomes very poor at this time. At this time, the maximum value α of the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same rotating arc max The change Δα of (i) max (i) To add 3 levels, the maximum value γ of the characteristic quantity describing the change of the sampling welding current at the same position of two adjacent rotating arcs max The change Δγ of (i) max (i) should also be increased by 3 levels. As the number of arc breaking increases, the arc instability increases. Therefore, the i-th change Δa(i) of the arc stable point a and the i-th change Δb(i) of the arc instability point b should be reduced by 3 levels to make the current arc stability state farther away from the arc stable point and the arc instability point.

[0092] Rule 7: When the number of arc breaking or arc near breaking increases by 0, the number of changes in the direction of the sampled welding current in the same circle of rotating arc decreases by 1, the change in the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc decreases by 0.25, and the change in the characteristic quantity describing the change in the magnitude of the sampled welding current at the same position of two adjacent circles of rotating arc decreases by 0.25, indicating that there is no arc breaking or arc near breaking at this time. At this time, the maximum value α of the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc max The change Δα of (i) max (i) To reduce the level by 1, the maximum value γ of the characteristic quantity describing the change of the sampling welding current at the same position of two adjacent rotating arcs max The change Δγ of (i) max (i) should also be reduced by 1 level. As the arc stability is developing in the direction of improvement, the i-th change amount Δa(i) of the arc stable point a and the i-th change amount Δb(i) of the arc instability point b should both be increased by 1 level to make the current arc stability state closer to the arc stable point and the arc instability point.

[0093] Rules 9-11: When the number of arc breaking or arc near breaking increases by 0, the number of changes in the direction of the sampled welding current in the same circle of rotating arc increases by 0, the change in the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc changes within ±0.1, and the change in the characteristic quantity describing the change in the magnitude of the sampled welding current at the same position of two adjacent circles of rotating arc changes within ±0.1. At this time, the arc stability has not changed within the error range. Therefore, the maximum value α of the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc max The change Δα of (i) max (i) unchanged, the maximum value γ of the characteristic quantity describing the change of the sampling welding current at the same position of two adjacent rotating arcs max The change Δγ of (i) max (i) remains unchanged, the i-th change Δa(i) of the arc stable point a and the i-th change Δb(i) of the arc instable point b remain unchanged.

[0094] The target optimal criterion in Table 1 is used as the objective function of the neural network, and the gradient descent method is used to adjust the connection weights of neurons in each layer of the neural network.

[0095] (3) Establishment of parameter adaptive model of rotating arc stability sensor based on target optimal neural network

[0096] The target optimal neural network is trained using the target optimal criteria in Table 1. When the error corresponding to each optimal relationship criterion is very small, the connection weight w from the input layer neuron to the hidden layer neuron is1 , the connection weights w between the hidden layer neurons and the output layer neurons 2 Satisfy the following formulas respectively,

[0097]

[0098]

[0099] The identified connection weights w 1 and 2 Substitution Figure 4 The target optimal neural network in can identify the changes in the important parameters of the general model of the rotating arc stability sensor, and considering the influence of the previous model parameters on the current adaptive model parameter identification, it can be identified that the adaptive parameters of the rotating arc stability sensor model satisfy the following formula:

[0100]

[0101] In the formula, (α max (i),γ max (i), a(i), b(i)) are the adaptive parameters of the rotating arc stability sensor model identified for the i-th time, (Δα max (i),Δγ max (i), Δa(i), Δb(i)) are the important parameters of the general model of the rotating arc stability sensor (α max ,γ max ,a,b),λ 3 , 4 , 5 , 6 and 7 is the weighting coefficient.

[0102] The identified adaptive parameters are replaced by Figure 1 The parameters of the general model of the rotating arc stability sensor can be used to establish a parameter adaptive model of the rotating arc stability sensor based on the target optimal neural network. The parameter adaptive model establishment method based on the target optimal neural network studied in the present invention can realize online self-tuning of model parameters according to working conditions, thereby online establishing an adaptive model of the rotating arc stability sensor that adapts to uncertain working conditions.

[0103] The online detection method of arc stability based on adaptive model is as follows:

[0104] Taking the four characteristic quantities (η(i), β(i), α(i), γ(i)) of the i-th rotating arc sensor signal as input and utilizing the adaptive model of the rotating arc stability sensor, the i-th identified arc stability comprehensive evaluation index w(i) can be output.

[0105] use Figure 1 The relationship between the arc length change coefficient K(i) and the four characteristic quantities of the rotating arc sensor signal can be calculated. The sensitivity of the arc stability comprehensive evaluation index to the four characteristic quantities of the rotating arc sensor signal satisfies the following formula:

[0106]

[0107] In the formula, k η (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity η(i), k β (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity β(i), k α (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity α(i), k γ (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity γ(i). w(i) is the arc stability comprehensive evaluation index identified for the i-th time, K(i) is the arc length change coefficient identified for the i-th time, α max (i) and γ max (i) is the adaptive parameter of the rotating arc stability sensor model identified for the i-th time using equation (9). 1 , w 2 , w 3 and 4 are the weights of the four characteristic quantities of the rotating arc sensor signal. n represents the number of filter currents used to determine the number of times the sampling welding current direction changes in the same circle of rotating arc, and its value is 20.

[0108] When the welding parameters are determined, the arc stability generally does not change suddenly, or when the welding parameters are continuously adjusted, the welding arc stability also changes gradually. Therefore, in order to reduce the interference of noise, the median filtering method is used to filter out the arc stability comprehensive evaluation index w 0 (i) satisfies the following equation,

[0109] w 0 (i)=median[w(i),w(i-1),w(i-2),w(i-3),w(i-4)] (11)

[0110] Where w(i) is the comprehensive evaluation index of arc stability identified for the i-th time, and median[] is the median function.

[0111] When starting welding, set the welding current to 250A and the welding voltage to 15V. Figure 5 For the straight fillet weld after robot tracking welding, the weld is not well formed. Figure 6It can be found that the identified arc stability comprehensive evaluation index is less than -1. At this time, the arc is unstable and the arc heat used to melt the welding wire is insufficient, which will cause poor weld formation, which is consistent with the welding experiment results. The unstable arc makes the rotating arc sensor signal unreliable, causing the welding deviation identified based on the unreliable rotating arc sensor signal to fluctuate greatly within ±1.5mm. Therefore, poor arc stability not only affects the quality of weld formation, but also affects the rotating arc sensor weld tracking accuracy, and ultimately affects the welding quality.

[0112] Set the welding current to 250A and the welding voltage to 28V. Figure 7 The robot tracks the fillet weld that bends outward after welding. The weld is beautifully formed and the robot tracks the curved weld with high accuracy. Figure 8 It can be found that when the arc is just started, the identified arc stability comprehensive evaluation index is less than -1, indicating that the arc is unstable and the welding deviation is large. Subsequently, the identified arc stability comprehensive evaluation index is greater than 1, and the arc stability is good at this time. In addition, the welding deviation changes slightly within ±0.3mm, indicating that the arc stability is good and the weld tracking accuracy is high during the welding process. Figure 7 The welding results are consistent with those in .

[0113] The arc stability identified by the method of the present invention is consistent with the actual welding experiment result. Therefore, the method of the present invention can effectively identify the stability of the rotating arc, help to improve the working performance of the rotating arc sensor, and improve the quality of robot automatic welding.

[0114] The above is an embodiment of the present invention. The above embodiments and the specific parameters in the embodiments are only for the purpose of clearly describing the invention verification process, and are not intended to limit the patent protection scope of the present invention. The patent protection scope of the present invention is still subject to its claims. Any equivalent structural changes made by using the contents of the description and drawings of the present invention should be included in the protection scope of the present invention.

Claims

1. An arc stability online detection method based on a target optimal neural network is characterized by: include: A general model of the input-output mapping relationship of the rotating arc stability sensor is established and its important parameters are determined. A parameter adaptive model establishment method based on the target optimal neural network and an arc stability online detection method based on the adaptive model are proposed.

2. The arc stability online detection method based on target optimal neural network according to claim 1 is characterized in that: The general model of the mapping relationship between the input and output of the rotating arc stability sensor is established and its important parameters are determined, including: the general model Z() of the mapping relationship between the input and output of the rotating arc stability sensor satisfies the following formula, w=Z(η(i),β(i),α(i),γ(i),α max ,c max ,a,b,c,d,s) (1) Wherein, η(i), β(i), α(i) and γ(i) are four characteristic quantities of the rotating arc sensor signal. The characteristic quantity η(i) can describe the number of times the arc is broken or close to being broken, the characteristic quantity β(i) can describe the number of changes in the direction of the sampled welding current in the same circle of rotating arc, the characteristic quantity α(i) can describe the change in the magnitude of adjacent sampled welding currents in the same circle of rotating arc, and the characteristic quantity γ(i) can describe the change in the magnitude of the sampled welding currents at the same position of two adjacent circles of rotating arc. These four characteristic quantities can be calculated from the collected rotating arc sensor signal; parameter α max , γ max The values ​​of , a, b, c, d and σ determine the accuracy of the general model of the input-output mapping relationship of the rotating arc stability sensor; α max is the value of α(i) when the arc is unstable, γ max is the value of γ(i) when the arc is unstable, σ is the standard deviation of the Gaussian function, a is the arc stability point, b is the arc under-stability point, c is the point corresponding to the beginning of arc instability, and d is the arc instability point; w is the identified comprehensive evaluation index of arc stability; Since the parameter α max and γ max It will change with the changes of welding parameters and welding conditions. The accuracy of arc stability recognition is very sensitive to the arc stability point a and arc instability point b; therefore, parameter α max , γ max , a and b are four important variable parameters that determine the precise model of the input-output mapping relationship of the rotating arc stability sensor, and affect the accuracy of identifying arc stability.

3. The arc stability online detection method based on target optimal neural network according to claim 2 is characterized in that: The method for establishing a parameter adaptive model based on a target optimal neural network includes three parts: identifying the change in important parameters of a general model of a rotating arc stability sensor based on a neural network, automatically adjusting the weights of the neural network based on a target optimal criterion, and establishing a parameter adaptive model of a rotating arc stability sensor based on the target optimal neural network.

4. The arc stability online detection method based on target optimal neural network according to claim 3 is characterized in that: The method for identifying the change of important parameters of the general model of the rotating arc stability sensor based on the neural network includes: taking the change of the characteristic amount of the arc sensor signal (Δη(i), Δβ(i), Δα(i), Δγ(i)) as input, taking the change of important parameters of the general model of the rotating arc stability sensor (Δα max ,Δγ max ,Δa,Δb) is the output; The total input Y of the mth neuron in the hidden layer m (i) satisfies the following equation, In the formula, w k,m (i) is the connection weight from the kth neuron in the input layer to the mth neuron in the hidden layer, X k (i) is the input of the kth neuron in the input layer, m is the serial number of the hidden layer neuron, satisfying m=1, 2, ..., 6; Δη(i) is the change of η(i), Δβ(i) is the change of β(i), Δα(i) is the change of α(i), and Δγ(i) is the change of γ(i); The Sigmoid function is used as the activation function h() of the hidden layer neurons, and the output Y of the i-th neuron of the m-th hidden layer is m '(i) satisfies the following formula, Where Y m (i) is the total input of the mth neuron in the hidden layer at the i-th time; The actual output Z of the nth neuron in the output layer is n (i) satisfies the following equation, Get the four outputs of the neural network, The output error e of the i-th output of the n-th neuron in the output layer n (i) satisfies the following equation, In formula (4-6), w m,n (i) is the connection weight from the mth neuron in the hidden layer to the nth neuron in the output layer, Y m '(i) is the output of the mth neuron in the hidden layer, n is the number of the neuron in the output layer; Δα max (i) is the general model parameter α max The change in (i), Δγ max (i) is the parameter γ max (i), Δa(i) is the i-th change of parameter a, and Δb(i) is the i-th change of parameter b; Z n,0 (i) is the ideal output of the nth neuron in the output layer, Z n,0 (i) = (Δα max , Δγ max , Δa, Δb), Z n (i) is the actual output of the i-th neuron in the output layer.

5. The arc stability online detection method based on target optimal neural network according to claim 4 is characterized in that: The method of automatically adjusting the weight of the neural network based on the target optimal criterion includes: to realize the automatic adjustment of the parameters of the general model of the sensor system based on the target optimal neural network, it is necessary to design the target optimal criterion; after analyzing the relationship between the change amount of the characteristic quantity of the arc sensor signal and the change amount of the important parameters of the general model, the target optimal criterion is determined, which is divided into three categories: the first category of rules indicates that there is a situation where the arc is broken or close to being broken; the second category of rules indicates that although there is no situation where the arc is broken or close to being broken, the arc stability has changed in general intensity; the third category of rules indicates that the arc stability has not changed within the error range; A case of the first type of rule: when the number of arc breaking or near arc breaking increases by 64 times, the number of changes in the direction of the sampled welding current in the same circle of rotating arc increases by 20 times, the change in the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc increases by 1, and the change in the characteristic quantity describing the change in the magnitude of the sampled welding current at the same position of two adjacent circles of rotating arc increases by 1, indicating that the arc stability becomes very poor at this time. At this time, the maximum value α of the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc max The change Δα of (i) max (i) To add 3 levels, the maximum value γ of the characteristic quantity describing the change of the sampling welding current at the same position of two adjacent rotating arcs max The change Δγ of (i) max (i) It also needs to be increased by 3 levels; as the number of arc breaking increases, the arc instability increases, therefore, the i-th change Δa(i) of the arc stable point a and the i-th change Δb(i) of the arc instability point b should be reduced by 3 levels, so that the current arc stability state is further away from the arc stable point and the arc instability point; A case of the second type of rule: when the number of arc breaking or near arc breaking increases by 0 times, the number of changes in the direction of the sampled welding current in the same circle of rotating arc decreases by 1 time, the change in the characteristic quantity describing the change in the magnitude of adjacent sampled welding currents in the same circle of rotating arc decreases by 0.25, and the change in the characteristic quantity describing the change in the magnitude of the sampled welding currents at the same position of two adjacent circles of rotating arcs decreases by 0.25, indicating that there is no arc breaking or near arc breaking at this time. At this time, the maximum value α of the characteristic quantity describing the change in the magnitude of adjacent sampled welding currents in the same circle of rotating arc max The change Δα of (i) max (i) To reduce the level by 1, the maximum value γ of the characteristic quantity describing the change of the sampling welding current at the same position of two adjacent rotating arcs max The change Δγ of (i) max (i) It should also be reduced by 1 level; since the arc stability is developing in the direction of improvement, the i-th change amount Δa(i) of the arc stable point a and the i-th change amount Δb(i) of the arc instability point b should be increased by 1 level, so that the current arc stability state is closer to the arc stable point and the arc instability point; A case of the third type of rule: when the number of arc breaking or near arc breaking increases by 0 times, the number of changes in the direction of the sampled welding current in the same circle of rotating arc increases by 0 times, the change in the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc changes within ±0.1, and the change in the characteristic quantity describing the change in the magnitude of the sampled welding current at the same position of two adjacent circles of rotating arc changes within ±0.

1. At this time, the arc stability does not change within the error range. Therefore, the maximum value α of the characteristic quantity describing the change in the magnitude of adjacent sampled welding current in the same circle of rotating arc max The change Δα of (i) max (i) unchanged, the maximum value γ of the characteristic quantity describing the change of the sampling welding current at the same position of two adjacent rotating arcs max The change Δγ of (i) max (i) remains unchanged, the i-th change Δa(i) of the arc stable point a and the i-th change Δb(i) of the arc unstable point b remain unchanged; The target optimal criterion is used as the objective function of the neural network, and the gradient descent method is used to adjust the connection weights of neurons in each layer of the neural network.

6. The arc stability online detection method based on target optimal neural network according to claim 5 is characterized in that: The parameter adaptive model of the rotating arc stability sensor based on the target optimal neural network includes: using the target optimal criterion to train the target optimal neural network, when the error corresponding to each optimal relationship criterion is very small, at this time, the connection weight w1 from the input layer neuron to the hidden layer neuron and the connection weight w2 from the hidden layer neuron to the output layer neuron respectively satisfy the following formulas: Substituting the identified connection weights w1 and w2 into the target optimal neural network, the changes in the important parameters of the general model of the rotating arc stability sensor can be identified. Considering the influence of the previous model parameters on the identification of the adaptive model parameters, it can be identified that the adaptive parameters of the rotating arc stability sensor model satisfy the following formula: In the formula, (α max (i),γ max (i), a(i), b(i)) are the adaptive parameters of the rotating arc stability sensor model identified for the i-th time, (Δα max (i),Δγ max (i), Δa(i), Δb(i)) are the important parameters of the general model of the rotating arc stability sensor (α max ,γ max ,a,b), λ3, λ4, λ5, λ6 and λ7 are weighting coefficients; The identified adaptive parameters are used to replace the parameters of the general model of the rotating arc stability sensor, and a parameter adaptive model of the rotating arc stability sensor based on the target optimal neural network can be established.

7. The arc stability online detection method based on target optimal neural network according to claim 6 is characterized in that: The arc stability online detection method based on the adaptive model includes: taking four characteristic quantities (η(i), β(i), α(i), γ(i)) of the i-th rotating arc sensor signal as input, and using the adaptive model of the rotating arc stability sensor to output the arc stability comprehensive evaluation index w(i) identified for the i-th time; By using the relationship between the arc length change coefficient K(i) and the four characteristic quantities of the rotating arc sensor signal, the sensitivity of the arc stability comprehensive evaluation index to the four characteristic quantities of the rotating arc sensor signal is calculated to satisfy the following formula: In the formula, k η (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity η(i), k β (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity β(i), k α (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity α(i), k γ (i) is the sensitivity of the arc stability comprehensive evaluation index to the characteristic quantity γ(i); w(i) is the arc stability comprehensive evaluation index identified for the i-th time, K(i) is the arc length change coefficient identified for the i-th time, α max (i) and γ max (i) is the adaptive parameter of the rotating arc stability sensor model identified for the i-th time using equation (9); w1, w2, w3 and w4 are the weights of the four characteristic quantities of the rotating arc sensor signal; n represents the number of filter currents used to determine the number of changes in the direction of the sampled welding current in the same circle of rotating arc, and its value is 20; In order to reduce the interference of noise, the comprehensive evaluation index of arc stability w0(i) after filtering out the noise by median filtering satisfies the following formula: w0(i)=median[w(i),w(i-1),w(i-2),w(i-3),w(i-4)] (11) Where w(i) is the comprehensive evaluation index of arc stability identified for the i-th time, and median[] is the median function.

Images