Fig。 8 Membership functions of the control oil pressure

    Fuzzification process of characteristic signal parameters is to transfer the precise  input values  of characteristic  signal parameter to fuzzy membership value。  Firstly,  the input values of characteristic signal parameter to each range, the range of slants small, slants big, normal are transferred。 Secondly,  fuzzification  process  is  conducted  to  the characteristic signal parameters that have been transferred to certain domain range。 The process would turn the original

precise input into fuzzy membership value between 0~1。

2。3 Training and learning of fuzzy neural network model of the hydraulic system

We make the characteristic signal of the hydraulic system of outriggers  as  the  fuzzy  neural  network's  input and  the corresponding  failure  causes  of system  as  fuzzy  neural network's output。 Then we set up network model of fault diagnosis respectively as figure 1 。The process of training and learning are shown in fig。 9。

Each operation loop consists the following:

     Firstly, it reads training sample data and the range of each feature parameter from the knowledge databases, and operates fuzzification process with the sample data of input fault。Second,  it writes the fuzzification process data and the expected output fault samples data into the neural network。Then it calculates the output of each layer node by the neural network。 Third, it calculates the error between actual output of output layer node and expected output, and determines if the training results meet the requirement of accuracy。

    If precision requirement or to the maximum number of training are met, it stores this training's network weights and threshold into knowledge database, before ends this training process。

   If the precision requirement or to the maximum number of trainings are not met, it implements backward transmission calculation on the direction of reducing the error, and adjusts the weights and threshold of output layer and hidden layer。 After  that,  it  implements  a  new  forward  transmission calculation  to  calculate  the  output  error  before  next comparison to the precision requirements  and maximum number of trainings。 Repeat the steps the above steps until the two criteria are met。

3 Implementation of the monitoring and fault diagnosis

Fig。  10  shows  the  software  architecture  of  condition monitoring and fault diagnosis。  We adopt the top-down approach for software developing。 The software is pided into separate modules, which is convenient for debugging, code maintaining and extensions。

Fig。l l shows the hardware architecture of the monitoring and fault diagnosis。 The hardware set consists of monitoring sensor, PLC controller system, data acquisition boards and vehicle-mounted computer or pc, etc。

  Fig。 10 Software architecture

Fig。 11 Hardware architecture

Fig。 12 shows the user interface of the hardware system of the fuzzy neural network fault diagnosis。

    User interface is pided into 3 areas visually, namely the real-time parameters monitoring area, diagnosis report and maintenance suggestion display area and tools button area。

   Diagnosis  report  consists  of  diagnosis  time,  diagnosis algorithm ID, fault code, fault phenomena, fault location and  fault cause。Maintenance  suggestion consists   of recommendations for machine operation in words or graphs。

   Tool buttons includes initial diagnosis button, radar view button, stop diagnosis button and exit button。 Initial diagnosis button and stop diagnosis button designed by the way of interlock。 The radar view button leads to the radar view of failure probability of the diagnosis system。