CN105244038A - Ore dressing equipment fault abnormity audio analyzing and identifying method based on HMM - Google Patents

Abstract

The invention provides an ore dressing equipment fault abnormity audio analyzing and identifying method based on an HMM (Hidden Markov Model), and relates to the digital audio processing technical field. The method includes the steps: inputting an ore dressing equipment audio signal in a WAV format, pre-processing a collected audio sample, extracting a linear prediction cepstrum coefficient (LPCC), a Mel frequency cepstrum coefficient (MFCC) and other characteristics and taking the characteristics as characteristic parameters, carry out training by a Baum-Welch algorithm, obtaining a state transfer probability matrix through training, conducting identification by means of a Viterbi algorithm, and realizing identification by calculating the maximum probability of an unknown audio signal during a transfer process as well as by a model corresponding to the maximum probability. The method can effectively detect abnormal sound in an audio signal so as to effectively identify fault abnormities of ore dressing equipment.

Description

A kind of preparation equipment failure exception audio analysis based on HMM and recognition methods

Technical field

The present invention relates to a kind of preparation equipment failure exception audio analysis based on hidden Markov model (HiddenMarkovModel, HMM) and recognition methods, belong to Digital Audio-Frequency Processing Techniques field.

Background technology

Preparation equipment is widely used in the various field such as ore dressing, resource reclaim of mineral products, and purposes is very extensive.A large amount of exploitation of mineral resources, all can propose more and more higher requirement to preparation equipment for the environmental consciousness that stock number constantly reduces and the mankind strengthen day by day, impel preparation equipment constantly to more greatly, more excellent and energy-efficient future development.

Along with the develop rapidly of science and technology and improving constantly of factory automation degree, daily servicing and the trimming of preparation equipment just seem particularly important.Both at home and abroad the research of this respect is all paid much attention to always, how always to be very important problem in fault early detection tendency.

Under normal circumstances, equipment failure and component defect all can in some specific modes, and the change etc. as the fluctuation of electric current, the vibration change of surface of shell, the overheating of parts diverse location or audio amplitude and frequency shows.If collection record can be carried out to these information, adopt Intelligentized Information means, in conjunction with expertise and all kinds of analytical approach, just effectively can realize the fault distinguishing of preparation equipment, thus the personnel that cause of reduction equipment failure and property loss.

Traditional maintenance means are all be manually main, make regular check on equipment operation situation and maintain equipment by maintainer, or the sound sent when listening to all parts work when patrolling and examining, thus judge whether operating equipment exists fault.

But in actual production workshop, fault produce low-frequency noise may cover by the HFS in man-made noise (man-made noise, be commonly referred to as during factory normally produces, due to the noise that the related causes such as plant equipment frictional impact, vibrations produce).Meanwhile, man-made noise is very large on the impact of the hearing of staff, works in such a case for a long time, to hearing and healthyly can cause irreversible damage.

Therefore, the Incipient Fault Diagnosis of preparation equipment has become the focus of research both at home and abroad, carry out the fault diagnosis research of preparation equipment for production practices, avoid major accident generation to have important realistic meaning.

Summary of the invention

Technical matters to be solved by this invention is: carry out fault diagnosis and identification to the audio frequency that preparation equipment produces.

The technical scheme adopted is:

Step 1: set up preparation equipment failure exception audio sample storehouse;

Step 2: pre-treatment step is carried out to the audio sample gathered;

Step 3: audio frequency characteristics parameter extraction and dimensionality reduction;

Step 4: adopt Baum-Welch algorithm to train; Set up hidden Markov parameter training model, obtain state transition probability matrix by training;

Step 5: adopt Viterbi algorithm by calculating the maximum probability of unknown sound signal in transfer process, and the model corresponding according to maximum probability identifies.

Advantage of the present invention: the invention provides a kind of preparation equipment failure exception audio analysis based on hidden Markov model and recognition methods, the time dependent characteristic of this sound signal, the concept of employing state is more appropriate, the change of audio signal characteristic shows as the transfer from a state to another state, and feature just shifts with certain probability from a state to another state.Therefore, this transfer process can well be represented with hidden Markov model.The method, based on the audio analysis of HMM and multiple features fusion and recognition technology, has higher fault recognition rate.

Accompanying drawing explanation

Fig. 1: based on preparation equipment failure exception audio analysis and the identification block diagram of HMM

Embodiment

As shown in Figure 1, input preparation equipment sound signal, pre-service is carried out to the audio sample gathered, extract the characteristic parameter of sound signal, apparatus for establishing failure exception audio sample storehouse, for audio sample signal sets up hidden Markov parameter training model, and the reference template storehouse of forming device audio frequency fault type.Separately audio frequency to be measured is carried out inputting, pre-service, extraction sound signal characteristic parameter after, adopt Viterbi algorithm, calculate the maximum probability of unknown sound signal in transfer process, and carry out template matches according to model corresponding to maximum probability and reference template storehouse, thus identify the fault type of sample to be tested.

The present invention adopts following technical scheme:

Based on preparation equipment failure exception audio analysis and a recognition methods of HMM, its step is as follows:

Step 1: set up preparation equipment failure exception audio sample storehouse;

This audio sample storehouse comprises the sound signal of multiple preparation equipment series under different conditions: equipment comprise a series of to five series once, little watt, bearing and machine top conical gear in secondary tube, pinion stand, synchronous motor, equipment state comprises five kinds: normal, leakage of oil, swing, vibration and shake; All from enterprises production locale collection; The content stored in Sample Storehouse is sound signal;

What store in Sample Storehouse is sound signal entirely, and its voice attribute is set to monophony (8KHz, 8,705Kb/s), and form is " .wav ", and the sound signal duration of recording is not at 1 ~ 15 minute etc.

Step 2: pre-treatment step;

Pre-emphasis, framing and windowing process are carried out, so that subsequent treatment to the audio sample gathered.Pre-emphasis is promoted by the HFS of voice signal, its objective is the high frequency resolution increasing voice; Voice signal is divided into one by one by framing (being generally 10ms-30ms), and its object is to possess short-term stationarity characteristic; Windowing process, its objective is the discontinuity problem in order to reduce frame starting and ending signal.Utilize overlapping discrete method [overlapping discrete method is prior art] to carry out framing, make transitions smooth between frame and frame, ensure its continuity.It is 256 (frame length is 32ms) that the present invention arranges frame length, and it is 128 (frame moves as 16ms) that frame moves.

Step 3: audio frequency characteristics parameter extraction and dimensionality reduction;

After pre-service, extract the characteristic parameter of audio sample, and dimension-reduction treatment is carried out to this characteristic parameter; The characteristic parameter that the present invention extracts comprises: the average of short-time energy and short-time magnitude, standard deviation, first order difference average, first order difference standard deviation, totally 8 dimensions; The average of short-time zero-crossing rate, standard deviation, first order difference average, first order difference standard deviation, totally 4 dimensions; Average, standard deviation, first order difference average, the first order difference standard deviation of 12 dimension LPCC and first order difference thereof, totally 96 dimensions; Average, standard deviation, first order difference average, the first order difference standard deviation of 12 dimension MFCC and first order difference thereof, totally 96 dimensions; Amount to 204 dimensional feature parameters;

Dimension-reduction treatment of the present invention refers to and realizes dimensionality reduction by principal component analysis (PCA), obtains sample characteristics vector sequence.[principal component analysis (PCA) is prior art].The object of dimensionality reduction is to delete redundant information on the one hand, thus reduces the calculated amount of algorithm for pattern recognition; To improve the validity of feature to classification, avoiding information to disturb on the other hand.

Step 4: based on the training process of Baum-Welch algorithm;

Obtain the audio frequency characteristics parameter after dimensionality reduction, adopt Baum-Welch algorithm to train, and in conjunction with sample labeling, for audio sample signal sets up hidden Markov parameter training model, obtain state transition probability matrix by training;

HMM is a kind of statistical recognition method based on parameter, and usually its model is defined as QUOTE λ=(A, B, π) λ=(A, B, π), determined by model parameter N, M and probability distribution parameters A, B, its characteristic parameter is defined as follows:

State transition probability distribution QUOTEA=[a _ij] A=[a _ij], wherein

a _ij＝P[q _t+1＝j|q _t＝i]1≤i≤N，1≤j≤N

Observe the probability distribution QUOTEB=[b of symbol _i(k)] B=[b _j(k)], wherein

b _j(k)＝P[o _t＝V _k|q _t＝j]1≤k≤M，1≤j≤N

Initial state probabilities distribution QUOTE π=[π _i] π=[π _i], wherein

π _i＝P[q ₁＝i]1≤i≤N

Wherein: N is the state number in hidden Markov model, and M is the symbolic number can observed in each state.Each observes symbol is QUOTEV={v ₁, v ₂..., v _mv={v ₁, v ₂..., v _m, observation sequence is QUOTEO={o ₁, o ₂..., o _to={o ₁, o ₂..., o _t.

Baum_Welch algorithm is used for the parameter estimation of HMM, namely how to train HMM.Its process can be described as: a given observation sequence QUOTEO={o ₁, o ₂..., o _to={o ₁, o ₂..., o _t), determine HMM model QUOTE λ=(A, B, π) λ=(A, B, a π), make QUOTEp (o| λ) p (o| λ) maximum.Its essence is problem is converted into searching make auxiliary function QUOTEQ=(λ, λ ') Q=(λ, λ ') maximized model QUOTE λ '=(A ' B ', π ') λ '=(A ', B ', π '), corresponding parameter estimation formula is:

${a^{\′}}_{ij}=\frac{{\Σ}_{t=1}^{T}p\left(o,q_{t-1}=i,q_t=j|\mathrm{\λ}\right)}{{\Σ}_{t=1}^{T}p\left(o,q_{t-1}=i|\mathrm{\λ}\right)}=\frac{{\Σ}_{t=1}^T{a}_{t-1}\left(i\right)a_{ij}{b}_j\left(o_t\right){\mathrm{\β}}_t\left(j\right)}{{\Σ}_{t=1}^T{a}_{t-1}\left(i\right){\mathrm{\β}}_{t-1}\left(i\right)}$

${b^{\′}}_i\left(k\right)=\frac{{\Σ}_{t=1}^{T}p\left(o,q_t=i|\mathrm{\λ}\right)\mathrm{\δ}\left(o_t,{\mathrm{\υ}}_k\right)}{{\Σ}_{t=1}^{T}p\left(o,q_t=i|\mathrm{\λ}\right)}=\frac{{\Σ}_{t=1}^T{a}_t\left(i\right){\mathrm{\β}}_t\left(j\right)\mathrm{\δ}\left(o_t,{\mathrm{\υ}}_k\right)}{{\Σ}_{t=1}^T{a}_t\left(i\right){\mathrm{\β}}_t\left(i\right)}$

${{\mathrm{\π}}^{\′}}_i=\frac{p(o,q_0=i|\mathrm{\λ})}{p(o,\mathrm{\λ})}=\frac{a_0\left(i\right){\mathrm{\β}}_0\left(i\right)}{{\mathrm{\Σ}}_{j=1}^N{a}_{\mathrm{\τ}}\left(j\right)}$

Step 5: the identifying adopting Viterbi algorithm;

Adopt Viterbi algorithm, by calculating the maximum probability of unknown sound signal in transfer process, and the model corresponding according to maximum probability identifies, thus obtains the fault type of sample to be tested.Identifying can be described as: a given observation sequence QUOTEO={o ₁, o ₂..., o _to={o ₁, o ₂..., o _tand HMM model QUOTE λ=(A, B, π) λ=(A, B, a π), obtain the status switch QUOTEQ making QUOTEp (Q, o| λ) p (Q, o| λ) maximum ^*q ^*.QUOTEQ ^*q ^*the model of corresponding reference library is recognition result.Its mathematical description is as follows:

In t, if: along certain paths QUOTEq ₁, q ₂..., q _t, q _t=iq ₁, q ₂..., q _t, q _t=i produces QUOTEo ₁, o ₂..., o _to ₁, o ₂..., o _tmaximum probability be QUOTE δ _t(i) δ _t(i), that is:

${\mathrm{\δ}}_t\left(i\right)=\underset{q_1,q_2,\mathrm{...},q_{t-1},}{max}(q_1,q_2,...,q_t,q_t=i,o_1,o_2,\mathrm{...},o_t|\mathrm{\λ})$

Then ask for optimum condition sequence QUOTEQ ^*q ^*process can be decomposed into:

Initialization: for QUOTE1≤i≤N1≤i≤N, have

Recursion: for QUOTE2≤t≤T2≤t≤T and QUOTE1≤j≤N1≤j≤N, have

Stop:

$\left\{\begin{array}{c}{P}^{*}=\underset{1\≤i\≤N}{\max }\[{\mathrm{\δ}}_t\left(i\right)\]\\ q^{*}=\arg \underset{1\≤i\≤N}{\max }\[{\mathrm{\δ}}_t\left(i\right)\]\end{array}\right.$

Path is recalled, and determines maximum rating series:

Experimental result and analysis:

Of the present invention based on the preparation equipment failure exception audio analysis of HMM and the system performance of recognition methods in order to verify, acquire the sound signal of multiple preparation equipment series under different conditions as test database.This database has 78 section audio samples, wherein swings and vibrate to acquire 12 sections and 18 sections respectively, and shake and leakage of oil all acquire 6 sections, normal condition 36 sections.

In test experiments, setting HMM model state number N=6, and adopt cross-over experiment verification mode, the training sample in experiment and test sample book separate.By the ratio of adjusting training sample and test sample book, the actual performance of testing algorithm.Table 1 and table 2 are respectively the recognition result under different proportion.

Table 1 is when the ratio of training sample and test sample book is respectively 1:2, that is: the sound signal choosing under often kind of state 1/3 carries out training (swing, vibration, shake, leakage of oil and normal condition be respectively 4,6,2,2 and 12 sections), and test sample book is sample complete or collected works.

Table 2 is when the ratio of training sample and test sample book is respectively 2:1, that is: the sound signal choosing under often kind of state 2/3 carries out training (swing, vibration, shake, leakage of oil and normal condition be respectively 8,12,4,4 and 24 sections), and test sample book is sample complete or collected works.

In above table, the element value except principal diagonal be 0 more, the recognition performance of illustrative system is better.As can be seen from Table 1 and Table 2, by increasing the number of training sample, the discrimination of system can be made to increase.As can be seen from Table 2, adopt audio analysis of the present invention and recognition methods to have very high discrimination, can effectively identify preparation equipment failure exception.

Claims (4)

1., based on preparation equipment failure exception audio analysis and a recognition methods of HMM, its step is as follows:

Step 1: set up preparation equipment failure exception audio sample storehouse;

This audio sample storehouse comprises the sound signal of multiple preparation equipment series under different conditions: equipment comprise a series of to five series once, little watt, bearing and machine top conical gear in secondary tube, pinion stand, synchronous motor, state comprises: normal, leakage of oil, swing, vibration and shake; All from enterprises production locale collection; The content stored in Sample Storehouse is sound signal (WAV form);

Step 2: pre-treatment step;

Pre-emphasis process, framing and windowing process are carried out, so that subsequent treatment to the audio sample gathered;

Step 3: audio frequency characteristics parameter extraction and dimensionality reduction;

After pre-service, extract the characteristic parameter of audio sample, and dimension-reduction treatment is carried out to this characteristic parameter;

Step 4: based on the training process of Baum-Welch algorithm;

Obtain the audio frequency characteristics parameter after dimensionality reduction, adopt Baum-Welch algorithm to train, and in conjunction with sample labeling, for audio sample signal sets up hidden Markov parameter training model, obtain state transition probability matrix by training;

Step 5: the identifying adopting Viterbi algorithm;

Adopt Viterbi algorithm, by calculating the maximum probability of unknown sound signal in transfer process, and the model corresponding according to maximum probability identifies.

2. the preparation equipment failure exception audio analysis based on HMM according to claim 1 and recognition methods, is characterized in that, described step 3 sound intermediate frequency sample characteristics parameter extraction refers to and is extracted following parameter to every frame of every section audio signal:

The average of short-time energy and short-time magnitude, standard deviation, first order difference average, first order difference standard deviation, totally 8 dimensions;

The average of short-time zero-crossing rate, standard deviation, first order difference average, first order difference standard deviation, totally 4 dimensions;

Average, standard deviation, first order difference average, the first order difference standard deviation of 12 dimension LPCC and first order difference thereof, totally 96 dimensions;

Average, standard deviation, first order difference average, the first order difference standard deviation of 12 dimension MFCC and first order difference thereof, totally 96 dimensions;

Amount to 204 dimensional feature parameters;

Described dimension-reduction treatment refers to and realizes dimensionality reduction by principal component analysis (PCA), obtains sample characteristics vector sequence.

3. the preparation equipment failure exception audio analysis based on HMM according to claim 1 and recognition methods, is characterized in that, the preparation equipment failure exception audio frequency training process based on HMM in step 4 is:

In preparation equipment audio sample storehouse, first pre-service is carried out to audio sample, then the characteristic parameter of each audio sample is calculated, and adopt Baum-Welch algorithm to train the proper vector after dimensionality reduction, for each audio sample signal sets up hidden Markov parameter training model, obtain state transition probability matrix by training.

4. the preparation equipment failure exception audio analysis based on HMM according to claim 1 and recognition methods, is characterized in that, the preparation equipment failure exception audio identification process based on HMM in step 5 is:

Adopt Viterbi algorithm to calculate the maximum probability of unknown sound signal in transfer process, and the model corresponding according to maximum probability identify.