CN107782857A - Flexible measurement method based on the edible fungus fermented process key parameter for improving CS BPNN - Google Patents
Flexible measurement method based on the edible fungus fermented process key parameter for improving CS BPNN Download PDFInfo
- Publication number
- CN107782857A CN107782857A CN201710928140.9A CN201710928140A CN107782857A CN 107782857 A CN107782857 A CN 107782857A CN 201710928140 A CN201710928140 A CN 201710928140A CN 107782857 A CN107782857 A CN 107782857A
- Authority
- CN
- China
- Prior art keywords
- bird
- variable
- value
- particle
- edible fungus
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
- G01N33/02—Food
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Food Science & Technology (AREA)
- Chemical & Material Sciences (AREA)
- Medicinal Chemistry (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Measuring Or Testing Involving Enzymes Or Micro-Organisms (AREA)
Abstract
The invention discloses the flexible measurement method based on the edible fungus fermented process key parameter for improving CS BPNN, this method is analyzed edible fungus fermented process mechanism first, is chosen suitable external variable and is used based on discrete PSO choosing auxiliary variables algorithm to choose auxiliary variable of the most suitable external variable as soft-sensing model;Training sample database is established according to history tank batch data, it is normalized;Then a BP neural network model is designed, utilizes the weights and threshold value of improved cuckoo algorithm optimization neutral net;Neutral net after optimization is trained, hypha biomass is predicted with the model optimized.The present invention is using improved cuckoo algorithm come Optimized BP Neural Network, compared with the CS BP neural network methods of standard, convergence rate and local search ability greatly promote, the real-time online for realizing edible fungus fermented process key variable is accurately predicted, significant to the parameter detecting and optimal control of edible fungus fermented process.
Description
Technical field
The present invention be it is a kind of be used to solving it is edible fungus fermented during be difficult to bacterium with physical sensors On-line sampling system
The On-line Estimation problem of this critical biochemical variable of silk biomass, belong to the technical field of hard measurement and soft instrument construction.
Background technology
Fermentation is a kind of important industrial production engineering, and it is typically characterised by, and inherent mechanism is complicated, repeatability is poor, production
Fluctuation is poor, has the non-linear and time-varying characteristics of height;On-line measurement is carried out to it, is to carry out the dynamic controls such as feed supplement, oxygen supply
Important evidence, while be also to optimize premise and the basis of scheduling.In general, measuring for significant process variable is to pass through
Laboratory periodically samples, off-line operation, analysis measure.There is large effect in the cycle of sampling to the variable value measured,
And often lagged on the time, cause variable information to lag so that fermentation process is difficult to real-time closed-loop optimal control.Therefore,
Research how to obtain in time it is edible fungus fermented during key variables status information, to fermentation process build optimal growth ring
Border, and then implement optimal control, improve product yield and be respectively provided with significance with quality.
BP neural network hard measurement is a kind of data-driven flexible measurement method widely used in recent years, because its is excellent
None-linear approximation ability, the soft sensor modeling of nonlinear system process it has been widely used to.In the present invention, for food
With the problems such as bacterium fermentation process is non-linear, large time delay, easy microbiological contamination and critical biochemical parameter are difficult to real-time online measuring, one is proposed
Kind improves the soft-measuring modeling method of CS-BP neutral nets.This method enables CS algorithms to ensure global optimizing early stage in search
Power;In the search later stage, local message is made full use of, search precision is improved, shortens learning time.
The content of the invention
In order to solve it is edible fungus fermented during it is extremely important, but be difficult to physical sensors On-line sampling system or in real time
The deficiency of the measuring method of the very high quantity of state of cost is measured, the present invention proposes a kind of based on the edible mushroom for improving CS-BPNN
Fermentation process key variables flexible measurement method, the measurement signal of the input variable provided by conventional on-line measurement instrument, is provided
The predicted value of current key state variable, parameter detecting and optimization operation for edible fungus fermented process provide critical process and referred to
Mark.
Technical scheme is as follows:
Based on the edible fungus fermented process key state variable flexible measurement methods of improved CS-BPNN, comprise the following steps:
Step 1:Choosing auxiliary variables, choose can direct measurement and the external variable closely related with process, with based on from
The choosing auxiliary variables algorithm for dissipating PSO chooses auxiliary variable of the most suitable external variable as soft-sensing model;
Step 2:Tranining database is established, gathers the auxiliary variable and key state of some history tank batches under same process
Variable data, the set of construction input and output vector pair, generates muscle-setting exercise sample database;And it is normalized place
Reason;Wherein input vector is auxiliary variable, and output vector is key stato variable;
Step 3:BP neural network model is designed, the neuron number of its input, hidden layer and output layer is set;
Step 4:Using the weights and threshold value of improved cuckoo algorithm optimization BP neural network, it is best to find prediction effect
Weights and threshold value;Neutral net after training optimization obtains soft-sensing model.
Step 5:Using the soft-sensing model trained, vector is newly entered according to current tank to be predicted batch, obtained
Obtain the predicted value of key stato variable.
Further, step 1 includes:
Step 1.1, fermentation process external variable data are gathered:Pass through gas flow sensor, speed probe, CO2Gas
Quick electrode, liquid-level probe, thermal resistance, dissolved oxygen electrode, pH electrodes, peristaltic pump collection air mass flow q, motor speed of agitator r,
CO2Burst size u, fermentating liquid volume v, fermentation jar temperature t, dissolved oxygen DO, zymotic fluid acidity-basicity ph, glucose feeding speed ρ,
Ammonia aqua stream rate of acceleration η;Offline inspection obtains key stato variable, including mycelia biology after zymotic fluid is centrifuged device separation
Measure S;
Step 1.2, the choosing auxiliary variables algorithm based on discrete PSO, finds optimal location and the particle of speed, by its from
Dispersion, obtain most suitable auxiliary variable combination.Specific implementation includes:
The detailed process of the step 1.2 includes:
Step 1.2.1:Initialize discrete PSO population and parameter, initialization population number M=22, inertia weight w=0.8,
Aceleration pulse c1=c2=0.7, maximum speed limit Vmax=4;Randomly generate M particle x of element 0 and 1i=(xi1,xi2,...,
xiD), i=1 ..., M, form initial population;Randomly generate initial velocity v of the random number between (0,1) as each particleij,
Form initial velocity matrix VM×D.The history desired positions p of each particleiWith population desired positions pgInitial value is all set to D dimensions 0
Vector;
Step 1.2.2:Evaluate the adaptive value of each particle;According to PLS regression algorithms and formula
The fitness function of each particle is calculated, seeks its adaptive value;Wherein, yi,actFor sample i actual measured value, y-i,predFor with
The model of the data foundation of i-th of sample is removed to yi,actPredicted value, yavgFor yi,actAverage value, h for obtain maximum
When pivot number,Predictive ability closer to 1 explanation model is better;
Step 1.2.3:More new particle optimal location.To each particle xi, by its adaptive value f (xi) lived through most with it
Good position piCorresponding adaptive value f (pi) make comparisons, if f (xi) it is better than f (pi), then by xiReplace with current desired positions;
Step 1.2.4:Population Regeneration optimal location.To desired positions p in current population (M particle)best, corresponded to
Adaptive value f (pbest) the desired positions p that is lived through with populationgCorresponding adaptive value f (pg) make comparisons, if f (pbest) excellent
In f (pg), then by pbestReplace with population optimal location;To prevent result to be absorbed in local optimum, by population optimal location pgNote
After lower, then that particle corresponding to it reinitialized;
Step 1.2.5:At the t+1 moment, particle i is in d (d=1, the 2 ..., D) speed of dimension space and the expression of position
Formula is: vid(t+1)=w × vid(t)+c1r1×(pid(t)-xid(t))+c2r2×(pgd(t)-xid) and x (t)id(t+1)=xid
(t)+vid(t+1);Position and the speed of current particle are adjusted according to the two formula, and according to formula
By the particle discretization after renewal;
Step 1.2.6:Judge whether to reach end condition, if above-mentioned condition meets, terminate iteration, otherwise return to step
1.2.2。
Further, step 3 realize it is specific as follows:
The network structure of selection is three-layer neural network structure, determines that input layer is arranged to 5 according to the number of auxiliary variable
Individual neuron, determine that output layer is arranged to 1 neuron according to the number of key stato variable, according to structured approach empirical equation
Midn=2inn+1 sets the nodes of hidden layer, and wherein midn is node in hidden layer, and inn is input layer number.
Further, the realization of step 4 comprises the following steps:
Step 4.1, BP neural network is regarded to the fitness function for improving CS algorithms as, then using algorithm itself
The strong advantage of global optimizing ability goes to seek optimal weight threshold combination.BP network structures are primarily determined that according to sample dimension,
Weights and threshold total number are determined successively, and then determine to improve the code length of cuckoo individual in CS algorithms.Randomly generate n
Bird's Nest xi(i=1,2 ..., n), each Bird's Nest represents the weights and threshold value of one group of neutral net that will optimize training, sets
Population scale n, maximum iteration Nmax, maximum probability of detection pamax, minimum probability of detection pamin, probability of detection pa;
Step 4.2, according to fitness functionCurrent optimal Bird's Nest is selected in calculatingWherein n is
Total sample number, y'(i) i-th sample hands-on output valve, y (i) is the desired output of i-th of sample.WillProtect
The next generation is left to, other Bird's Nests are according to formulaCarry out location updating, wherein i=1,2 ..., n;Represent position of i-th of Bird's Nest in the t times iteration, a is step-length scale factor, n be Bird's Nest quantity,Represent dot product;L
(λ) obeys L é vy distributions, is random flight step-length.I.e.:Wherein,For t
The optimum position of generation storage, μ and the equal Normal Distributions of ξ.The fitness value of these Bird's Nest positions is calculated simultaneously, is done with the previous generation
Compare one group chosen and remain into the next generation, obtain one group of preferably Bird's Nest position;
Step 4.3, according to formulaTo paEnter
Row renewal, wherein pamaxAnd paminThe maximum and minimum value of probability of detection is represented respectively;NmaxFor maximum iteration, NiTo work as
Preceding iterations (1≤Ni≤Nmax).Produce and obey equally distributed random number r ∈ (0,1), with paContrast, if r > pa, then it is right
Bird's nest position is changed at random, on the contrary then constant, i.e., changes the larger bird's nest position of probability of detection at random, retain probability of detection
Less bird's nest position;
Step 4.4, the Bird's Nest position after re-test changes, and contrasted with previous generation Bird's Nests.It is finally more excellent in test result
One group of Bird's NestIn select the present age global optimum positionAnd judge its fminWhether essence is reached
Degree requires.If meeting the requirements,For global optimum, return to step 4.2 if not meeting;
Step 4.5, the optimal solution vector of acquisition is subjected to Gray code operation, extracts the weights and threshold value of BP neural network
Establish afterwards and improve CS-BPNN models, neutral net is trained using sample data, when training error no longer reduces, training then may be used
Terminate.
Beneficial effects of the present invention:
1. the online process data that the present invention is provided using computer system and conventional instrumentation, only by a small amount of
Artificial sample, the hard measurement for realizing the improvement CS-BP neutral nets of the key stato variable of edible fungus fermented process builds
Mould, solves the problem for not having state variable on-line checking instrument to be difficult to on-line checking in fermentation process.
2. compared with manual sampling is chemically examined, reduce the workload of site operation personnel, reduce artificial in fermentation process
The uncertainty of the measurement introduced is operated, improves the ageing of measurement, reduces asking of sampling that the data brought lag offline
Topic.
3. the present invention using improved cuckoo algorithm come Optimized BP Neural Network, the CS-BP neutral net sides with standard
Method is compared, and being overcome in the present invention in flexible measurement method in standard CS-BP neutral nets convergence rate and local search ability has
The problem of be short of.
4. the present invention considers the factor for influenceing edible fungus fermented process key state variable comprehensively, largely using existing
Conventional detection signal realizes the on-line prediction of key stato variable, using it is simple, easy, cost is relatively low, hard measurement result compared with
Accurately.This method contributes to the optimal control of edible fungus fermented process and optimization to run.
Brief description of the drawings
Flow, measuring instrumentss and the allocation of computer figure of the edible fungus fermented processes of Fig. 1;
Fig. 2 is based on the edible fungus fermented key variables soft-sensing model Establishing process figure for improving CS-BP neutral nets;
In Fig. 1:1 biological fermentation tank, 2 steam generators, 3 air compressors, 4 air cleaners, 5 gas flows sensing
Device, 6 speed probes, 7CO2Gas sensing electrode, 8 liquid-level probes, 9 thermal resistances, 10 dissolved oxygen electrodes, 11pH electrodes, 12 peristaltic pumps,
13 peristaltic pumps, 14 whizzers, 15 intelligent controllers, 16 host computers.
Label symbol used is as follows in Fig. 1:
Fermentation jar temperature-t, motor speed of agitator-r, fermentating liquid volume-v
Air mass flow-q, CO2Release rate-u, glucose feeding speed-ρ,
Ammonia aqua stream rate of acceleration-η, dissolved oxygen DO, zymotic fluid acid-base value-pH.
In Fig. 1, solid arrow represents logistics (zymotic fluid, water, air and steam) direction, and dotted line represents signal stream.
Embodiment
Below in conjunction with the implementing procedure shown in the examples of implementation and Fig. 2 of edible fungus fermented process key state variable prediction
Figure, is described in detail to of the embodiment of the present invention.
Select auxiliary vector:Choosing auxiliary variables, energy direct measurement and the external variable closely related with process are chosen, used
Auxiliary variable of the most suitable external variable as soft-sensing model is chosen based on discrete PSO choosing auxiliary variables algorithm;
1. the acquisition of fermentation process data
It is edible fungus fermented that fermented and cultured is carried out using SF-500L types fermentation monitoring system.Monitoring system as shown in Figure 1 is by sending out
Fermentation tank 1, steam generator 2, air compressor 3, air cleaner and 4 water systems composition.Pass through during edible fungus fermented
Gas flow sensor 5, speed probe 6, CO2 gas sensing electrodes 7, liquid-level probe 8, thermal resistance 9, dissolved oxygen electrode 10, pH electricity
Pole 11, peristaltic pump 12 and peristaltic pump 13 gather air mass flow q, motor speed of agitator r, CO2Release rate u, fermentating liquid volume v, hair
Fermentation tank temperature t, dissolved oxygen DO, zymotic fluid acidity-basicity ph, glucose feeding speed ρ and ammonia aqua stream rate of acceleration η.Zymotic fluid pass through from
Offline inspection obtains hypha biomass after centrifugal separator 14 separates.
The process of choosing auxiliary variables based on discrete PSO is as follows:
For obtained external variable data (fermentation jar temperature t, motor speed of agitator r, fermentating liquid volume v, air mass flow
q、 CO2Release rate u, glucose feeding speed ρ, ammonia aqua stream rate of acceleration η, dissolved oxygen DO, zymotic fluid acidity-basicity ph) be based on from
The choosing auxiliary variables algorithm for dissipating PSO searches out auxiliary variable of the most suitable external variable as soft-sensing model.
Because hard measurement choosing auxiliary variables are the optimum organization problems of each independent variable, i.e., each particle is 0,1 composition
A vector, 0, which represents certain variable, is not selected, 1 represent it is selected.This just needs to expand continuous PSO to binary space
Exhibition, construct a kind of binary system particle Optimized model of discrete form.
If candidate's auxiliary variable collection is combined into m × D dimension spaces.The then position x of each particleiIt is D's to be expressed as one group of long degree
0th, 1 coded combination, it represents a kind of selection situation of auxiliary variable.The initial velocity v of particleiCan take between (0,1) with
Machine number.Its iterative process is continuous, is no longer 0,1 binary number after each renewal of particle position.Therefore, in order to maintain grain
The binary form of son, needed after particle position renewal by xiChange into 0 or 1.
It is comprised the following steps that:
Step 1:Discrete PSO population and parameter is initialized, initialization population number M=22, inertia weight w=0.8, is added
Fast constant c1=c2=0.7, maximum speed limit Vmax=4.Randomly generate M particle x of element 0 and 1i=(xi1,xi2,...,
xiD), i=1 ..., M, form initial population.Randomly generate initial velocity v of the random number between (0,1) as each particleij,
Form initial velocity matrix VM×D.The history desired positions p of each particleiWith population desired positions pgInitial value is all set to D dimensions 0
Vector;
Step 2:Evaluate the adaptive value of each particle.According to PLS regression algorithms and formula
The fitness function of each particle is calculated, seeks its adaptive value.Wherein, yi,actFor sample i actual measured value, y-i,predFor with
The model of the data foundation of i-th of sample is removed to yi,actPredicted value, yavgFor yi,actAverage value, h for obtain maximum
When pivot number,Predictive ability closer to 1 explanation model is better;
Step 3:More new particle optimal location.To each particle xi, by its adaptive value f (xi) the best position that is lived through with it
Put pi=(pi1,pi2,…,pid) corresponding to adaptive value f (pi) make comparisons, if f (xi) it is better than f (pi), then by xiReplace with and work as
Preceding desired positions;
Step 4:Population Regeneration optimal location.To desired positions p in current population (M particle)best, by corresponding to it
Adaptive value f (pbest) the desired positions p that is lived through with populationg=(pg1,pg2,…,pgd) corresponding to adaptive value f (pg) make ratio
Compared with if f (pbest) it is better than f (pg), then by pbestReplace with population optimal location., will to prevent result to be absorbed in local optimum
Population optimal location pgAfter writing down, then that particle corresponding to it reinitialized;
Step 5:In the speed of d (d=1,2 ..., D) dimension space and position it is formula in t+1 moment particle i: vid(t+
1)=w × vid(t)+c1r1×(pid(t)-xid(t))+c2r2×(pgd(t)-xid) and x (t)id(t+1)=xid(t)+vid(t+
1).Wherein, r1,r2Random number respectively between (0,1),;Position and the speed of current particle are adjusted according to this two formula, and
According to formulaBy the particle discretization after renewal;
Step 6:Judge whether to reach end condition, if above-mentioned condition meets, terminate iteration, otherwise return to step two.
According to above step from air mass flow q, broth temperature t, fermentating liquid volume v, CO2Release rate u, ammonia aqua stream accelerate
CO is obtained in 9 variables such as rate η, glucose feeding speed ρ, dissolved oxygen DO, motor mixing speed r, zymotic fluid acidity-basicity ph2
5 release rate u, air mass flow q, dissolved oxygen DO, fermentating liquid volume v, zymotic fluid acidity-basicity ph variables are to hypha biomass shadow
Ring maximum.
2. establish training sample database:Gather the auxiliary variable and key state of some history tank batches under same process
Variable data, the set of construction input and output vector pair, generates muscle-setting exercise sample database, and it is normalized place
Reason;Wherein input vector is auxiliary variable, and output vector is key stato variable.It is implemented as follows:
Edible fungus fermented process forms sample according to following structure, and collects some history tank batch instructions under same process
Practice sample data, the data of 10 batches are collected in the embodiment of the present invention, wherein 90% data are made as sample data set, 10%
For validation data set.Sample is expressed as { xk,yk, wherein xkFor the input of sample, that is, the auxiliary vector-air mass flow chosen
Q, fermentating liquid volume v, CO2Release rate u, dissolved oxygen DO, zymotic fluid acidity-basicity ph.The output y of samplekDominated for be predicted
Variable-hypha biomass S.
Training sample acquisition and recording structure such as table 2, the time is the sampling period in fermentation process, offline to reduce leading variable
Error and Assay, sample choice is carried out using result of laboratory test three times according to same sample leading variable, finally averaged:
The sample data structure of table 2
Should be representative in view of sample data, and coverage is wider as far as possible, should at least include fermentation
Process normal range of operation, manually regulate and control fermentation pressure tank, fermentation jar temperature and motor speed of agitator, in production technology
Change the operating point of fermentation process, sample examination after each operating condition changes system steadily in the range of permission as far as possible.
3. build BP neural network model:Setting input layer is 5 neurons, and output layer is 1 neuron, by BP
It is 11 neurons that hidden layer is chosen in the adaptive number experiment of neutral net.Wherein BP neural network input layer is 5 neurons,
Fermentating liquid volume v, air mass flow q, CO are corresponded to respectively2Release rate u, dissolved oxygen DO, zymotic fluid acidity-basicity ph, 1 god of output layer
Through member, to the key parameters in requisition for prediction --- hypha biomass.
4. using the BP neural network of improved cuckoo algorithm optimization step 3, optimal weights and threshold value are found, is obtained
To the soft-sensing model of optimization.Shown in comprising the following steps that:
(1) first according to the dimension of sample data, i.e. the number of auxiliary variable determines the structure of neutral net, including power
The sum of value and threshold value, and then determine to improve the dimension of search space in CS algorithms;Secondly, Bird's Nest coordinate and nerve net are constructed
Mapping relations between network weight threshold.Weights and threshold value to be determined can be considered as to improve in CS algorithms and need what is found
Bird's nest coordinate, all Bird's Nest monomers can be all regarded as one of combination of weights and threshold value, then this also represents one accordingly
The structure of kind BP neutral nets, and the quality of any one Bird's Nest is all determined by fitness function.According to neural network weight
With threshold value feature, cuckoo individual UVR exposure uses real number coding method, and each individual is represented by a real number string, the real number string
It is made up of the weights between input layer, hidden layer, output layer and threshold value, and each cuckoo individual may pass through Gray code
Extract all weights and threshold value.The search target of algorithm, exactly find coordinate (the i.e. one group of power for most adapting to Bird's Nest
Value and threshold value), the fitness function of selection is worth to minimum value;Finally, BP neural network is assigned the Bird's Nest coordinate components
It is used for building the CSBP neutral nets after improving for use as its weights and threshold value;
(2) network structure chosen is to apply three-layer neural network structure the most typical, and wherein input layer sets 5 god
Through member, output layer sets 1 neuron, according to structured approach empirical equation:Midn=2inn+1, wherein midn are hidden layer section
Points, inn is input layer number.When with improved CS Algorithm for Training BP neural network, it is net to define Bird's Nest location components
The weights and threshold value of network.Connection weight number is 5*11+11*1=66, threshold value 11+1=12, so Bird's Nest dimension is set to 66
+ 12=78.Bird's Nest population quantity is n=50, the maximum probability p that bird egg is foundamax=0.9, minimum probability of detection pamin=
0.1, iterations is set to Nmax=2000;
(3) according to fitness functionOptimal Bird's Nest is selected in calculatingWherein n is total sample number,
Y'(i) the hands-on output valve of i-th of sample, y (i) are the desired output of i-th of sample.
Other Bird's Nest positions are according to formulaCarry out location updating, wherein i=1,2 ..., n;Represent position of i-th of Bird's Nest in the t times iteration;A is step-length scale factor;N is Bird's Nest quantity;Represent dot product;L
(λ) obeys L é vy distributions, is random flight step-length.I.e.:WhereinFor t generations
The optimum position of storage, μ and the equal Normal Distributions of ξ, random numbers of the β between (0,2).Calculate these Bird's Nest positions simultaneously
Fitness value, one group chosen is compared with the previous generation and remains into the next generation, obtains one group of preferable Bird's Nest position.
(4) according to formulaTo paCarry out more
Newly, wherein pamaxAnd paminThe maximum and minimum value of probability of detection is represented respectively;NmaxFor maximum iteration, NiChanged to be current
Generation number (1≤Ni≤Nmax).Produce and obey equally distributed random number r ∈ (0,1), with paContrast, if r > pa, then to bird's nest
Position is changed at random, on the contrary then constant, i.e., changes the larger bird's nest position of probability of detection at random, retain probability of detection compared with
Small bird's nest position.
(5) the Bird's Nest position after re-test changes, and contrasted with previous generation Bird's Nests.Finally in preferably one group of test result
Bird's NestIn select the present age global optimum positionAnd judge its fminWhether required precision is reached.
If meeting the requirements,For global optimum, (2) are returned if not meeting.
(6) optimal solution vector of acquisition is subjected to Gray code operation, extracts the weights and threshold value of all BP neural networks
Establish afterwards and improve CS-BP models, neutral net is trained using sample data, when training error no longer reduces, training can then tie
Beam.
5. predict key stato variable:The improved CS-BP neural network soft sensor models trained using step 4, root
According to the predicted value for being newly entered vector, obtaining key stato variable of current tank to be predicted batch.It is implemented as follows:
After soft-sensing model is established, realized, and be embedded into intelligent controller 15 using embedded type C Programming with Pascal Language,
As the input vector x of tank to be predicted batchi+1, after measuring instrumentss read in intelligent controller 15, intelligent controller 15 utilizes soft survey
The predicted value of key stato variable is calculated in range sequence, and prediction result is sent into host computer 16 through data channel
Upper real-time display.
Above-described embodiment is used for illustrating the present invention, rather than limits the invention, the present invention spirit and
In scope of the claims, to any modifications and changes of the invention made, protection scope of the present invention is both fallen within.
Claims (9)
1. the flexible measurement method based on the edible fungus fermented process key parameter for improving CS-BPNN, it is characterised in that including as follows
Step:
Step 1:Choosing auxiliary variables, energy direct measurement and the external variable closely related with process are chosen, with based on discrete PSO
Choosing auxiliary variables algorithm choose auxiliary variable of the most suitable external variable as soft-sensing model;
Step 2:Tranining database is established, gathers the auxiliary variable and key stato variable of some history tank batches under same process
Data, the set of construction input and output vector pair, generate muscle-setting exercise sample database;And it is normalized;Its
Middle input vector is auxiliary variable, and output vector is key stato variable;
Step 3:BP neural network model is designed, the neuron number of its input, hidden layer and output layer is set;
Step 4:Using the weights and threshold value of improved cuckoo algorithm optimization BP neural network, the best power of prediction effect is found
Value and threshold value, the soft-sensing model optimized;
Step 5:Using trained obtained optimization soft-sensing model, vector is newly entered according to current tank to be predicted batch,
Obtain the predicted value of key stato variable.
2. the flexible measurement method according to claim 1 based on the edible fungus fermented process key parameter for improving CS-BPNN,
Characterized in that, the detailed process of step 1 includes:
Step 1.1, fermentation process external variable data are gathered:Pass through gas flow sensor, speed probe, CO2Air-sensitive electricity
Pole, liquid-level probe, thermal resistance, dissolved oxygen electrode, pH electrodes, peristaltic pump collection air mass flow q, motor speed of agitator r, CO2Release
Amount u, fermentating liquid volume v, fermentation jar temperature t, dissolved oxygen DO, zymotic fluid acidity-basicity ph, glucose feeding speed ρ, ammonia aqua stream add
Speed η;Offline inspection obtains key stato variable after zymotic fluid is centrifuged device separation, and the key stato variable includes bacterium
Silk biomass S;
Step 1.2, the choosing auxiliary variables algorithm based on discrete PSO, optimal location and the particle of speed are found, its is discrete
Change, obtain most suitable auxiliary variable combination.
3. the flexible measurement method according to claim 2 based on the edible fungus fermented process key parameter for improving CS-BPNN,
Characterized in that, the detailed process of the step 1.2 includes:
Step 1.2.1:Discrete PSO population and parameter is initialized, initialization population number M=22, inertia weight w=0.8, is accelerated
Constant c1=c2=0.7, maximum speed limit Vmax=4;Randomly generate M particle x of element 0 and 1i=(xi1,xi2,...,xiD), i
=1 ..., M, form initial population;Randomly generate initial velocity v of the random number between (0,1) as each particleij, composition is initially
Rate matrices VM×D.The history desired positions p of each particleiWith population desired positions pgInitial value is all set to the vector of D dimensions 0;
Step 1.2.2:Evaluate the adaptive value of each particle;According to PLS regression algorithms and formula
The fitness function of each particle is calculated, seeks its adaptive value;Wherein, yi,actFor sample i actual measured value, y-i,predFor with removing
Fall the model of the data foundation of i-th of sample to yi,actPredicted value, yavgFor yi,actAverage value, h for obtain maximum when
Pivot number,Predictive ability closer to 1 explanation model is better;
Step 1.2.3:More new particle optimal location.To each particle xi, by its adaptive value f (xi) the best position that is lived through with it
Put piCorresponding adaptive value f (pi) make comparisons, if f (xi) it is better than f (pi), then by xiReplace with current desired positions;
Step 1.2.4:Population Regeneration optimal location.To desired positions p in current population (M particle)best, will be suitable corresponding to it
Should value f (pbest) the desired positions p that is lived through with populationgCorresponding adaptive value f (pg) make comparisons, if f (pbest) it is better than f
(pg), then by pbestReplace with population optimal location;To prevent result to be absorbed in local optimum, by population optimal location pgAfter writing down,
That particle corresponding to it is reinitialized again;
Step 1.2.5:At the t+1 moment, particle i is in d (d=1, the 2 ..., D) speed of dimension space and the expression formula of position:
vid(t+1)=w × vid(t)+c1r1×(pid(t)-xid(t))+c2r2×(pgd(t)-xid) and x (t)id(t+1)=xid(t)+
vid(t+1);Position and the speed of current particle are adjusted according to the two formula, and according to formulaWill
Particle discretization after renewal;
Step 1.2.6:Judge whether to reach end condition, if above-mentioned condition meets, terminate iteration, otherwise return to step 1.2.2.
4. the flexible measurement method according to claim 1 based on the edible fungus fermented process key parameter for improving CS-BPNN,
Characterized in that, the detailed process of step 3 includes:The Artificial Neural Network Structures of selection are three-layer neural network structure, input
Layer number is set according to the number of auxiliary variable, and the number of output layer is set according to the number of key stato variable, hidden layer section
Points are according to structured approach empirical equation:Midn=2inn+1 determines that wherein midn is node in hidden layer, and inn is input layer section
Points.
5. the flexible measurement method according to claim 4 based on the edible fungus fermented process key parameter for improving CS-BPNN,
Characterized in that, the input layer is arranged to 5 neurons, output layer is arranged to 1 neuron, and hidden layer is arranged to 11 god
Through member.
6. the flexible measurement method according to claim 5 based on the edible fungus fermented process key parameter for improving CS-BPNN,
Characterized in that, 5 neurons of the input layer correspond to fermentating liquid volume v, air mass flow q, CO respectively2Release rate u, dissolving
Oxygen DO, zymotic fluid acidity-basicity ph;1 neuron of the output layer is to the key parameters in requisition for prediction.
7. the flexible measurement method according to claim 1 based on the edible fungus fermented process key parameter for improving CS-BPNN,
Characterized in that, the detailed process of the step 4 includes:
Step 4.1, BP neural network is regarded as to the fitness function for improving CS algorithms, utilizes the global optimizing energy of algorithm itself
The strong advantage of power goes to seek optimal weight threshold combination;BP network structures are primarily determined that according to sample dimension, determine power successively
Value and threshold total number, and then determine to improve the code length of cuckoo individual in CS algorithms;
Randomly generate n Bird's Nest xi(i=1,2 ..., n), each Bird's Nest represents the power of one group of neutral net that will optimize training
Value and threshold value, population scale n, maximum iteration N are setmax, maximum probability of detection pamax, minimum probability of detection pamin, find
Probability pa;
Step 4.2, according to fitness functionCurrent optimal Bird's Nest is selected in calculatingWherein n is sample
Sum, y'(i) i-th sample hands-on output valve, y (i) is the desired output of i-th of sample;WillRetain under
A generation, other Bird's Nests are according to formulaCarry out location updating, wherein i=1,2 ..., n;Represent the
Position of the i Bird's Nest in the t times iteration, a is step-length scale factor, n be Bird's Nest quantity,Represent dot product;L (λ) obeys L é
Vy is distributed, and is random flight step-length;I.e.:Wherein,For the optimal of t generation storages
Position, μ and the equal Normal Distributions of ξ;The fitness value of these Bird's Nest positions is calculated simultaneously, and one chosen is compared with the previous generation
Group remains into the next generation, obtains one group of preferably Bird's Nest position;
Step 4.3, according to formulaTo paCarry out more
Newly, wherein pamaxAnd paminThe maximum and minimum value of probability of detection is represented respectively;NmaxFor maximum iteration, NiFor current iteration
Number (1≤Ni≤Nmax).Produce and obey equally distributed random number r ∈ (0,1), with paContrast, if r > pa, then to bird's nest position
Put and changed at random, it is on the contrary then constant, i.e., change the larger bird's nest position of probability of detection at random, it is less to retain probability of detection
Bird's nest position;
Step 4.4, the Bird's Nest position after re-test changes, and contrasted with previous generation Bird's Nests.Finally in test result preferably one
Group Bird's NestIn select the present age global optimum positionAnd judge its fminWhether reaching precision will
Ask.If meeting the requirements,For global optimum, return to step 4.2 if not meeting;
Step 4.5, the optimal solution vector of acquisition is subjected to Gray code operation, extract BP neural network weights and threshold value after build
It is vertical to improve CS-BPNN models, neutral net is trained using sample data, when training error no longer reduces, training can then terminate.
8. the flexible measurement method according to claim 1 based on the edible fungus fermented process key parameter for improving CS-BPNN,
Characterized in that, the detailed process of step 5 includes:The soft-sensing model of optimization is realized using embedded type C Programming with Pascal Language, and it is embedding
Enter in intelligent controller, as the input vector x of tank to be predicted batchi+1After measuring instrumentss read in intelligent controller, intelligent control
The predicted value of key stato variable is calculated using hard measurement program for device.
9. the flexible measurement method according to claim 8 based on the edible fungus fermented process key parameter for improving CS-BPNN,
Characterized in that, step 5 also includes:Prediction result is sent to real-time display on host computer through data channel.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710928140.9A CN107782857A (en) | 2017-10-09 | 2017-10-09 | Flexible measurement method based on the edible fungus fermented process key parameter for improving CS BPNN |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710928140.9A CN107782857A (en) | 2017-10-09 | 2017-10-09 | Flexible measurement method based on the edible fungus fermented process key parameter for improving CS BPNN |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107782857A true CN107782857A (en) | 2018-03-09 |
Family
ID=61434203
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710928140.9A Pending CN107782857A (en) | 2017-10-09 | 2017-10-09 | Flexible measurement method based on the edible fungus fermented process key parameter for improving CS BPNN |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107782857A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109471363A (en) * | 2018-12-28 | 2019-03-15 | 浙江大学 | Industrial melt index soft measurement method based on aftereffect function and cuckoo search |
CN109547431A (en) * | 2018-11-19 | 2019-03-29 | 国网河南省电力公司信息通信公司 | A kind of network security situation evaluating method based on CS and improved BP |
CN110276483A (en) * | 2019-06-04 | 2019-09-24 | 湖北大学 | Sugaring raw material prediction technique neural network based |
CN110309983A (en) * | 2019-07-09 | 2019-10-08 | 重庆大学 | A kind of industry cloud service scheduling of resource matching combined method |
CN112307672A (en) * | 2020-10-29 | 2021-02-02 | 上海电机学院 | BP neural network short-term wind power prediction method based on cuckoo algorithm optimization |
CN116013411A (en) * | 2023-03-24 | 2023-04-25 | 北京蓝晶微生物科技有限公司 | Quantitative relation model, application, fermentation monitoring method, device, system and equipment |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101748186A (en) * | 2009-12-29 | 2010-06-23 | 江苏大学 | FSVM-based lysine fermentation process key state variable soft measuring method and system |
CN101929993A (en) * | 2010-08-20 | 2010-12-29 | 江苏大学 | Dynamic fuzzy neural network based penicillin fermentation process soft measuring modeling method |
-
2017
- 2017-10-09 CN CN201710928140.9A patent/CN107782857A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101748186A (en) * | 2009-12-29 | 2010-06-23 | 江苏大学 | FSVM-based lysine fermentation process key state variable soft measuring method and system |
CN101929993A (en) * | 2010-08-20 | 2010-12-29 | 江苏大学 | Dynamic fuzzy neural network based penicillin fermentation process soft measuring modeling method |
Non-Patent Citations (3)
Title |
---|
孙晨 等: "基于布谷鸟算法优化BP神经网络模型的股价预测", 《计算机应用与软件》 * |
李丽娟 等: "基于离散PSO的软测量辅助变量选择算法", 《系统仿真学报》 * |
杨强大 等: "基于改进BP神经网络的菌体浓度软测量", 《控制与决策》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109547431A (en) * | 2018-11-19 | 2019-03-29 | 国网河南省电力公司信息通信公司 | A kind of network security situation evaluating method based on CS and improved BP |
CN109471363A (en) * | 2018-12-28 | 2019-03-15 | 浙江大学 | Industrial melt index soft measurement method based on aftereffect function and cuckoo search |
CN109471363B (en) * | 2018-12-28 | 2020-08-04 | 浙江大学 | Industrial melt index soft measurement method based on post-effect function and rhododendron search |
CN110276483A (en) * | 2019-06-04 | 2019-09-24 | 湖北大学 | Sugaring raw material prediction technique neural network based |
CN110309983A (en) * | 2019-07-09 | 2019-10-08 | 重庆大学 | A kind of industry cloud service scheduling of resource matching combined method |
CN112307672A (en) * | 2020-10-29 | 2021-02-02 | 上海电机学院 | BP neural network short-term wind power prediction method based on cuckoo algorithm optimization |
CN116013411A (en) * | 2023-03-24 | 2023-04-25 | 北京蓝晶微生物科技有限公司 | Quantitative relation model, application, fermentation monitoring method, device, system and equipment |
CN116013411B (en) * | 2023-03-24 | 2023-05-30 | 北京蓝晶微生物科技有限公司 | Quantitative relation model, application, fermentation monitoring method, device, system and equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107782857A (en) | Flexible measurement method based on the edible fungus fermented process key parameter for improving CS BPNN | |
CN104123476B (en) | Forecasting of Gas Concentration method based on extreme learning machine and its device | |
CN108090500B (en) | Soft measurement method for key parameters in fermentation production process of edible fungi | |
CN111460728B (en) | Method and device for predicting residual life of industrial equipment, storage medium and equipment | |
CN109407654B (en) | Industrial data nonlinear causal analysis method based on sparse deep neural network | |
CN102231057B (en) | Method for carrying out soft-sensing on lysine fermenting process on basis of chaos particle swarm optimization (CPSO) | |
CN106156401B (en) | Multi-combination classifier based data driving system state model online identification method | |
CN101799888B (en) | Industrial soft measurement method based on bionic intelligent ant colony algorithm | |
CN110023850A (en) | Method and control device for control technology system | |
CN110738271B (en) | Concentrate grade prediction method in zinc flotation process | |
CN112488235A (en) | Elevator time sequence data abnormity diagnosis method based on deep learning | |
CN106052753A (en) | Straw fermentation fuel ethanol production process key state variable soft measuring method based on fuzzy support vector machine | |
CN110807554A (en) | Generation method and system based on wind power/photovoltaic classical scene set | |
CN110824915A (en) | GA-DBN network-based intelligent monitoring method and system for wastewater treatment | |
CN108664009A (en) | Divided stages based on correlation analysis and fault detection method | |
CN101748186A (en) | FSVM-based lysine fermentation process key state variable soft measuring method and system | |
CN107341359A (en) | Stalk fermentation produces the flexible measurement method of ethanol process key parameters | |
CN110046377B (en) | Selective integration instant learning soft measurement modeling method based on heterogeneous similarity | |
CN101639902A (en) | Modeling method of support vector machine (SVM)-based software measurement instrument in biological fermentation process | |
CN107766880B (en) | Photosynthetic bacteria fermentation process key parameter soft measurement method based on BA-LSSVM | |
CN109408896B (en) | Multi-element intelligent real-time monitoring method for anaerobic sewage treatment gas production | |
CN103279030B (en) | Dynamic soft measuring modeling method and device based on Bayesian frame | |
CN114239397A (en) | Soft measurement modeling method based on dynamic feature extraction and local weighted deep learning | |
CN105930936B (en) | A kind of growth prediction method producing microorganism contained by food in the process of circulation | |
CN113988311B (en) | Quality variable prediction method, device, terminal and storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180309 |