Resistance Spot Welding Quality On-line Monitoring Method

Resistance spot welding quality is instable and difficult to control, which have influenced the wide application of resistance spot welding technology seriously .So it is necessary to develop a kind of online judging system with nondestructive , low cost and high diagnostic reliability. Lots of dynamic information, which can directly and indirectly reflect the quality of weld spot, is included within the dynamic signals of resistance spot welding process. Signal characteristic analysis, parallel treatment and information integration technology are the key to set up quality evaluation model and realize quality inspecting of weld spot online.For online monitoring and controlling quality of weld spot, the electrode voltage, welding current, electrode displacement and dynamic resistance signals were synchronously gathered. The modern analysis methods of signal were adopted to analysis the characteristics of process signals gathered, and the time-domain statistic characteristics were picked up from the four signals to set up a set of data which token the pattern of spot welding process .By means of statistical analysis, pattern recognition and data mining analytic methods, the classification and prediction models of weld spot strength were implemented. The work was done as follows.A data gathering system based on AC6115 AD card, Rogowski current transducer, DA-5 differential voltage transformer displacement sensor of direct current and Hall sensors for current and voltage was developed ,by which electrode voltage ,welding current , electrode displacement and dynamic resistance signals were synchronously gathered. The signals' wave display, edit, signal data display and data file saving were carried out with gathering software developed by Visual Basic 6.0.Time-domain and frequency-domain methods were used to analysis electrode voltage, welding current, electrode displacement and dynamic resistance signals in this thesisdn order to research the relativity between the characters of signals and weld spot process. The preliminary analysis showed that electrode voltage, welding current, electrode displacement and dynamic resistance signals could recognize the affection of craft factor changes such as edge, diffluence and sheet materials thickness. The signals also could monitor explusion phenomenon during welding process. Because the variation of the signals gathered in frequency domain was not obvious, the signal factors were stressly analysised and picked up in time domain .Utilized the resistance and displacement signals' parameter to devide the different phases of the nugget forming, and developed the algorithm for extracting the parameters inspected, by which a factor vector including multi-information merging could be get to express of the resistance spot welding process.Linear, nonlinear regression analysis and RBF neural network were used to predict nugget strength based on the input vector which was constructed by factors inspected of electrode displacement signal. By means of the cross-validation estimate, the linear, nonlinear and RBF neural network models were able to effectively predict the weld strength, and realize evaluating the quality of the joint online. RBF neural network model showed stronger fault-tolerant and cluster abilities, and could be regarded as the follow-up research approach.The data reduction and generalization of factors vector, which is based on resistance and displacement signals, can be accomplished through data discrete technology. When utilized the higher level data concept to replace the low level one, the pattern of different welding course could be constructed. These patterns which corresponding with the different weld joint strength were classified according to different welding current ,and realized different pattern vectors representing the different weld strength. As the states of associative memory of the Hopfield neural network, pattern vectors were stored in the Hopfield network. Depending on the ability of associative memory, the Hopfield neural network classified the pattern vec...