CN105046322A - Method for diagnosing leading screw faults - Google Patents

Abstract

The invention discloses a method for diagnosing leading screw faults, belongs to the technical field of machinery fault diagnosis, and especially relates to ball screw fault diagnosis. The problem that a conventional intelligent leading screw fault diagnosis system is difficult for manual extraction of characteristics and limited in applying shallow-layer network non-linear expression capability. A diluted self-coding deep neural network structure is employed, and a network identification model employs a Softmax regression classifier to determine the hidden layer quantity of a network structure. The quantity of the input terminals and the quantity of the output terminals of a fault diagnosis model are determined, a training sample set, pre-training, fine tuned training and a fault diagnosis model test sample set are prepared, and the fault diagnosis performance of the fault diagnosis model is tested. Data sections in the test sample are integrally input continuously and sequentially, and the output quantity of the model is recorded to obtain an actual output list of the model. The ideal output list of the model is compared with the designed output list to obtain the fault performance test and evaluation result of the fault diagnosis model.

Description

A kind of leading screw method for diagnosing faults

Technical field

The invention belongs to technology for mechanical fault diagnosis field, particularly ball-screw fault diagnosis.

Background technology

Along with the progress of science and technology and the development of industrial requirement, ball-screw on the one hand continuous to complicated, at a high speed, efficient, light-duty, miniature or large-scale future development, but face harsher work and running environment on the other hand.While meeting equipment requirement, the Potential feasibility that leading screw breaks down and mode are also in corresponding increase, and leading screw is once break down, just may destroy whole equipment and even affect whole production run, cause huge economic loss, also may cause catastrophic casualties and form serious social influence.Therefore, to ball-screw carry out on-line monitoring, carry out equipment failure study mechanism, set up effectively, accurately fault diagnosis system seem very important.

Current leading screw fault diagnosis and state monitoring method less, " Southwest Jiaotong University's journal " 2010,45 (5):. disclose the leading screw forecasting technique in life span based on artificial intelligence, adopt B-spline fuzzy neural network to set up leading screw Life Prediction Model.Each rank power spectrum after " machine science and technology " 2013,5:003. discloses and extract WAVELET PACKET DECOMPOSITION, as characteristic parameter, utilizes BP neural network leading screw performance degradation assessment model.These methods are predicted mainly for the performance degradation of leading screw, and the method for its application does not have feature learning ability.

The feature learning of rising in recent years can well solve the problem of feature extraction.Feature learning is the Machine Learning Problems that original input data can be converted to the character representation of the study that can effectively exercise supervision by a class, this is a kind of feature extracting method of data-driven, learning process does not need the participation of the priori of the mankind, and therefore the method for this type of feature learning has the features such as self-adaptation is strong, strong robustness.The method of feature learning is applied to mechanical fault diagnosis system, a large amount of raw data that scene produces can be converted into valuable character representation, for recognition system provides valuable eigenwert.

Summary of the invention

The object of this invention is to provide a kind of leading screw method for diagnosing faults, it effectively can solve existing leading screw Intelligent Fault Diagnose Systems and manually extract feature difficulty and apply shallow-layer network non-linear expression problem limited in one's ability.

The object of the invention is to be achieved through the following technical solutions: design a kind of feature learning ability that can realize having, have the ball-screw method for diagnosing faults of multitiered network structure, its step is as follows:

One, the network structure with feature learning ability is set up

1, the structure of diagnostic model is determined

The present invention adopts dilution own coding deep neural network structure, and the model of cognition of network selects Softmax to return sorter, determines network structure hidden layer quantity.

2: the input end quantity determining fault diagnosis model

If fault diagnosis model has m input end node, m node input signal constitutes an input end vector x, is expressed as follows:

x＝(x ₁,x ₂,…,x _m)

In formula, x _ifor i-th input end node of model, i=1,2 ..., m.

3, the output terminal quantity of fault diagnosis model is determined

The quantity of fault diagnosis output terminal determined by leading screw fault type quantity, if the fault that leading screw has n kind dissimilar, the output terminal vector y of model is:

y＝(y ₁,y ₂,…,y _n)

In formula, y _jfor a jth output signal of leading screw fault diagnosis model, j=1,2 ..., n.

Two, fault diagnosis model is trained

1, training sample set is prepared

To each leading screw fault type, the data being not less than 100 groups are adopted to form respective training sample set.

2, pre-training

By original signal as ground floor network input signal I ₁=x, the middle layer of ground floor network is as the input layer I of second layer network ₂=m ₁, I ₂refer to the input layer of second layer network, m ₁refer to the middle layer of ground floor network.Every layer network is first trained respectively, when each layer parameter of training, fix other each layer parameter and remain unchanged, then superposition becomes multitiered network successively.Every layer network training is divided into cataloged procedure and decode procedure, and every layer network coding step is:

a ¹＝x

z ²＝W ⁽¹⁾a ¹+b ⁽¹⁾

a ²＝f(z ²)

In formula, a ¹for the value of ground floor network input is input signal values, W ⁽¹⁾the weighted value of ground floor, b ⁽¹⁾the bias of ground floor, z ²the input value of the second layer, a ²be the activation value of the second layer, f () is activation function.Decoding step is the inverse process of coding step:

z ³＝W ^(1，1)′a ²+b ^(1，1)′

a ³＝f(z ³)

In formula, W ^(1,1) ', b ^(1,1)' be W ^(1,1), b ^(1,1)inverse vector; z ³for the input value of third layer, a ³for the activation value of third layer (output layer).

3, fine setting training

Pre-training network structure top layer adds output layer I _o, after above-mentioned pre-training process completes, the parameter simultaneously being adjusted all layers by back-propagation algorithm, to improve result, is defined as fine setting training; Fine setting is trained for supervised learning process, input amendment set T=(X _t, Y _t), wherein X _tfor sample input signal, Y _tthe label data corresponding for input signal and fault type, determine the error value epsilon of finely tuning training, and when model meets following formula, when namely error range is less than or equal to error amount, training stops.

Three, test failure diagnostic model

1, fault diagnosis model test sample book collection is prepared

Test sample book collection is formed with the data segment comprising all fault types being different from training sample set, according to putting in order of test sample book collection, a corresponding pull spring thick stick State-output table, the ideal referred to herein as fault diagnosis model exports table; The corresponding test sample book of every a line in table concentrates the malfunction represented by corresponding line to export, and namely the ideal of fault diagnosis model exports.

2, the performance of fault diagnosis of test failure diagnostic model

Data segment successively in continuous input test sample set, the output quantity of record cast, obtains the actual output table of model, the ideal output table of model and design output table is contrasted, obtains the faulty behavior test and evaluation result of fault diagnosis model.

The present invention compared with prior art has following beneficial effect:

1, this method has feature learning ability

The degree of depth that the present invention proposes the study of a kind of feature based learns the leading screw fault diagnosis model of neural learning network.Model have employed degree of depth learning network, has feature learning ability.Degree of depth learning network forms more abstract high level by combination low-level feature and represents attribute classification or feature, to find that the distributed nature of data represents.Such a construction ensures study and the extraction of leading screw vibration signal essential characteristic, for ensureing to differentiate that the fault type of different leading screw provides condition.

2, this method multitiered network structure has good convergence

The model training of fault diagnosis of the present invention have employed without this training method of the supervision every layer network of pre-training, for effective convergence of multitiered network structure provides condition.Present invention employs the degree of depth learning network with feature learning ability, the core thinking of the training of degree of depth learning neural network is: unsupervised learning is used for the pre-training of each layer network; Only train one deck with unsupervised learning, using the input of its training result as its high one deck at every turn; All layers are finely tuned with the supervise algorithm under from top.Without the supervision every layer network of pre-training respectively, making the parameter of every layer network have an initial value, again by there being the fine setting of supervision to train on this initial value basis, in the hope of the optimum solution of whole network, convergence can be reached.

3, this method has better Fault Identification ability.

The present invention adopts model can extract the essential characteristic of leading screw vibration data, and non-linear expression's ability of degree of depth learning network is strong, makes its discriminating power stronger.Which enhance the Fault Identification ability of fault diagnosis model.

Accompanying drawing illustrates:

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is feature learning network diagram of the present invention;

Embodiment:

Below in conjunction with diagram, the present invention is described further.A leading screw fault diagnosis model for feature based study, comprises the steps:

One, the network structure with feature learning ability is set up

1, the structure of diagnostic model is determined

Adopt dilution own coding deep neural network structure, the model of cognition of network selects Softmax to return sorter, and the hidden layer of network selects two layers.

2: the input end quantity determining fault diagnosis model

Fault diagnosis model has m input end node, and m node input signal constitutes an input end vector x, is expressed as follows:

x＝(x ₁,x ₂,…,x _m)

In formula, x _ifor i-th input end node of model, i=1,2 ..., m;

In this example, the terminal input section m that counts is 12.Be respectively temporal signatures value: root-mean-square value x ₁, peak factor x ₂, pulse factor x ₃, shape factor x ₄, nargin factor x ₅, peak-to-peak value x ₆, continuous wavelet transform energy eigenvalue x ₇, continuous wavelet transform energy eigenvalue x ₈, continuous wavelet transform energy eigenvalue x ₉, continuous wavelet transform energy eigenvalue x ₁₀, continuous wavelet transform energy eigenvalue x ₁₁, continuous wavelet transform energy eigenvalue x ₁₂.

3, the output terminal quantity of fault diagnosis model is determined

The quantity of fault diagnosis output terminal determined by leading screw fault type quantity, if leading screw has fault dissimilar in n, the output terminal vector y of model is:

y＝(y ₁,y ₂,…,y _n)

In formula, y _jfor a jth output signal of leading screw fault diagnosis model, j=1,2 ..., n.

In this example, fault type number n=4.Be respectively: normal condition y ₁, early stage degenerate state y ₂, performance degradation state y ₃with malfunction y ₄.

Two, fault diagnosis model is trained

1, training sample set is prepared

To leading screw in normal condition, early stage degenerate state, respectively under performance degradation state and malfunction gathers 250 groups of sample datas and forms respective training sample set;

2, pre-training

By original signal as ground floor network input signal I ₁=x.The middle layer of ground floor network is as the input layer I of second layer network ₂=m ₁, in formula, I ₂refer to the input layer of second layer network, m ₁refer to the middle layer of ground floor network.Every layer network is first trained respectively, when each layer parameter of training, fixes other each layer parameter and remains unchanged.Superposition becomes multitiered network successively again.Every layer network training as shown in Figure 2, is divided into cataloged procedure and decode procedure.Every layer network coding step is:

a ¹＝x

z ²＝W ⁽¹⁾a ¹+b ⁽¹⁾

a ²＝f(z ²)

In formula, a ¹for the value of ground floor (input layer) is input signal values, W ⁽¹⁾the weighted value of the 1st layer, b ⁽¹⁾the bias of the 1st layer, z ²the input value of the second layer (hidden layer), a ²be the activation value of the second layer (hidden layer), f () is activation function.Decoding step is the inverse process of coding step:

z ³＝W ^(1,1)′a ²+b ^(1,1)′

a ³＝f(z ³)

In formula, W ^(1,1)', b ^(1,1)' be W ^(1,1), b ^(1,1)inverse vector.Z ³for the input value of third layer (output layer), a ³for the activation value of third layer (output layer).

In this example, the middle layer neuron number of ground floor network is 10, and the neuron number of second layer network is 8, and activation function is sigmoid function:

$f\left(x\right)=\frac{1}{1+e^{-x}}$

3, fine setting training

Pre-training network structure top layer adds output layer I _o.After above-mentioned pre-training process completes, adjust the parameter of all layers to improve result by back-propagation algorithm colleague, be defined as fine setting training.Fine setting is trained for supervised learning process, input amendment set T=(X _t, Y _t), wherein X _tfor sample input signal, Y _tthe label data corresponding for input signal and fault type, determine the error value epsilon of finely tuning training, and when model meets following formula, when namely error range is less than or equal to error amount, training stops.

$\mathrm{\θ}=||Y_t-{\hat{Y}}_t||\≤\mathrm{\ϵ}$

In formula, for the output valve of model.

In this example, the sample set quantity of fine setting training is 1000 groups, is 250 groups of normal condition samples, 250 groups of early stage degenerate state samples, 250 groups of performance degradation state sample and 250 groups of malfunction samples.

Three, test failure diagnostic model

1, fault diagnosis model test sample book collection is prepared

Test sample book collection is formed with the data segment comprising all fault types being different from training sample set, according to putting in order of test sample book collection, a corresponding pull spring thick stick State-output table, the ideal referred to herein as fault diagnosis model exports table; The corresponding test sample book of every a line in table concentrates the malfunction represented by corresponding line to export, and namely the ideal of fault diagnosis model exports;

In this example, get test sample book number totally 400 groups, be 100 groups of normal condition samples, 100 groups of early stage degenerate state samples, 100 groups of performance degradation state sample and 100 groups of malfunction samples.

2, the performance of fault diagnosis of test failure diagnostic model

Data segment successively in continuous input test sample set, the output quantity of record cast, obtains the actual output table of model, the ideal output table of model and design output table is contrasted, obtains the faulty behavior test and evaluation result of fault diagnosis model.

In this example, the test accuracy rate of test sample book set is 92.5%, meets actual requirement of engineering.

Claims (1)

1. a leading screw method for diagnosing faults, its step is as follows:

One, the network structure with feature learning ability is set up

(1) structure of diagnostic model, is determined

The present invention adopts dilution own coding deep neural network structure, and the model of cognition of network selects Softmax to return sorter, determines network structure hidden layer quantity;

(2): the input end quantity determining fault diagnosis model

If fault diagnosis model has m input end node, m node input signal constitutes an input end vector x, is expressed as follows:

x＝(x ₁,x ₂,…,x _m)

In formula, x _ifor i-th input end node of model, i=1,2 ..., m;

(3) the output terminal quantity of fault diagnosis model, is determined

The quantity of fault diagnosis output terminal determined by leading screw fault type quantity, if the fault that leading screw has n kind dissimilar, the output terminal vector y of model is:

y＝(y ₁,y ₂,…,y _n)

In formula, y _jfor a jth output signal of leading screw fault diagnosis model, j=1,2 ..., n;

Two, fault diagnosis model is trained

(1), training sample set is prepared

To each leading screw fault type, the data being not less than 100 groups are adopted to form respective training sample set;

(2), pre-training

By original signal as ground floor network input signal I ₁=x, the middle layer of ground floor network is as the input layer I of second layer network ₂=m ₁, I ₂refer to the input layer of second layer network, m ₁refer to the middle layer of ground floor network; Every layer network is first trained respectively, when each layer parameter of training, fix other each layer parameter and remain unchanged, then superposition becomes multitiered network successively; Every layer network training is divided into cataloged procedure and decode procedure, and every layer network coding step is:

a ¹＝x

z ²＝W ⁽¹⁾a ¹+b ⁽¹⁾

a ²＝f(z ²)

In formula, a ¹for the value of ground floor network input is input signal values, W ⁽¹⁾the weighted value of ground floor, b ⁽¹⁾the bias of ground floor, z ²the input value of the second layer, a ²be the activation value of the second layer, f () is activation function, and decoding step is the inverse process of coding step:

z ³＝W ^(1,1)′a ²+b ^(1,1)′a ³＝f(z ³)

In formula, W ^(1,1)', b ^(1,1)' be W ^(1,1), b ^(1,1)inverse vector; z ³for the input value of third layer, a ³for the activation value of third layer (output layer);

(3), fine setting training

Pre-training network structure top layer adds output layer I _o, after above-mentioned pre-training process completes, the parameter simultaneously being adjusted all layers by back-propagation algorithm, to improve result, is defined as fine setting training; Fine setting is trained for supervised learning process, input amendment set T=(X _t, Y _t), wherein X _tfor sample input signal, Y _tthe label data corresponding for input signal and fault type, determine the error value epsilon of finely tuning training, and when model meets following formula, when namely error range is less than or equal to error amount, training stops;

Test failure diagnostic model

(1) fault diagnosis model test sample book collection, is prepared

Test sample book collection is formed with the data segment comprising all fault types being different from training sample set, according to putting in order of test sample book collection, a corresponding pull spring thick stick State-output table, the ideal referred to herein as fault diagnosis model exports table; The corresponding test sample book of every a line in table concentrates the malfunction represented by corresponding line to export, and namely the ideal of fault diagnosis model exports;

(2), the performance of fault diagnosis of test failure diagnostic model

Data segment successively in continuous input test sample set, the output quantity of record cast, obtains the actual output table of model, the ideal output table of model and design output table is contrasted, obtains the faulty behavior test and evaluation result of fault diagnosis model.