CN108921233A - A kind of Raman spectrum data classification method based on autoencoder network - Google Patents
A kind of Raman spectrum data classification method based on autoencoder network Download PDFInfo
- Publication number
- CN108921233A CN108921233A CN201810857856.9A CN201810857856A CN108921233A CN 108921233 A CN108921233 A CN 108921233A CN 201810857856 A CN201810857856 A CN 201810857856A CN 108921233 A CN108921233 A CN 108921233A
- Authority
- CN
- China
- Prior art keywords
- layer
- characteristic
- autocoder
- network
- classification
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Abstract
The Raman spectrum data classification method based on autoencoder network that the invention proposes a kind of.The present invention extracts the Raman spectrogram of the Alzheimer's blood platelet in different sick ages by optical tweezer Raman system, and pretreated Raman spectrogram is randomly divided into training set and test set；Using training set sample as input data, the sparse autoencoder network of stack of two layers of characteristic layer is constructed by layer-by-layer greedy coaching method；The further feature of second layer autocoder is trained into Softmax classifier as input data, using Softmax classifier after training as Softmax classification layer with the output layer of the sparse autoencoder network of the stack for replacing two layers of characteristic layer；Preliminary classification network, sorter network after being optimized by back-propagation algorithm are constructed according to the sparse autoencoder network of the stack of two layers of characteristic layer；Using test set as sorter network input data after optimization, the sorter network of neuronal quantity optimization is obtained by optimizing neuronal quantity.The present invention improves the accuracy and stability of classification.
Description
Technical field
The invention belongs to artificial intelligence application technical fields, and in particular to a kind of Raman spectrum number based on autoencoder network
According to classification method.
Background technique
Alzheimer disease (Alzheimer's Disease) is one of the most common type disease type in senile dementia.It is one
Kind central nervous system degeneration, and a kind of chronic degenerative disease, can cause to damage, clinical symptoms show as recognizing to brain
Know dysfunction, memory disorders and aphasis etc., directly affects the normal life of patient.It is shown according to data in 2006,
The AD patient in the whole world existing more than 200,000,000.Also, social senilization's degree is also higher and higher, and the afflicted patient of AD is also more and more,
Alzheimer disease has become a kind of disease that entire society has to face.Each AD patient can to one family or
Several families bring heavy burden and affect indirectly the every aspect of society by relevant medical plan etc., because
This, the treatment of this disease has caused extensive public attention.But up to the present, due to the understanding to AD pathogenesis
Deficiency, diagnosing and treating also rely primarily on the clinical experience of attending physician, psychological test and medical imaging, such as computerized tomography
Scan (CT) and Magnetic resonance imaging (NMR) etc..However, these diagnostic methods have certain subjectivity, mistaken diagnosis and leakage mostly
It examines phenomenon to happen occasionally, and haves the defects that early stage AD patient can not be identified.Therefore, there is an urgent need to one in clinical practice
Objective, Accurate Diagnosis AD the foundation of kind, so as to identify and diagnose AD in the different phase of disease.
Single celled Raman spectrum detects and is taken the lead in reporting by Pupples in nineteen ninety, and subsequent Raman spectroscopy is in disease
Diagnosis aspect using more and more, Raman spectroscopy has the advantages of non-destructive and Noninvasive testing so that
It can ideally detect biological sample, obtain valid data.And Raman spectroscopy combines some sorting algorithms such as partially most
The statistical analysis techniques such as small square law, linear discriminant analysis, support vector machines can be used for analyzing Alzheimer's not
The different characteristic of same disease age blood platelet is simultaneously identified.
In terms of biology, Raman spectrum is applied to be often some biologists or chemistry to the research of cytopathy
What family carried out, they generally use the tools such as SPSS software to realize analysis and processing to data.Traditional Raman spectrum classification side
The methods of method, such as Principal Component Analysis, Partial Least Squares, these methods are all by initial data by a way
Transformation is that the mapping of feature space finds suitable character representation method in new feature space, to realize classification.This
Class method often calculates complexity, and classifying quality is bad.The learning method of some shallow-layers, such as support vector machines, linear regression point
Analysis etc., then be to find out decision boundary according to the tape label data of input.But usually will appear such a case, that is, it inputs
When certain samples and other differences between samples are larger in training sample, the decision boundary of classification will receive large effect.
Summary of the invention
To solve the above-mentioned problems, the invention proposes a kind of Raman spectrum data classification side based on autoencoder network
Method, the present invention is sparse from encoding according to stack, and it is small to be directed to different sick age Alzheimer's blood in conjunction with Softmax classifier
Plate Raman spectrum is analyzed and is identified.
The solution of the present invention is a kind of Raman spectrum data classification method based on autoencoder network, which is characterized in that packet
Include following steps：
Step 1：The Raman spectrum of the Alzheimer's blood platelet in different sick ages is extracted by optical tweezer Raman system
Figure, and Raman spectrogram progress background is reduced, baseline correction, smooth and average treatment, by the Raman spectrogram after processing
It is randomly divided into training set and test set；
Step 2：Using training set sample as first layer autocoder input data, pass through layer-by-layer greedy coaching method training
Each layer weight coefficient and each layer bias vector of first layer autocoder are obtained, by the further feature of first layer autocoder
Each layer of second layer autocoder is obtained by successively greedy coaching method training as second layer autocoder input data
Weight coefficient and each layer bias vector construct two layers of characteristic layer by first layer autocoder and second layer autocoder
The sparse autoencoder network of stack；
Step 3：Using the further feature of second layer autocoder described in step 2 as input data with training
Softmax classifier, by gradient descent algorithm Optimization Solution Softmax classifier, using Softmax classifier after training as
Softmax classifies layer with the output layer of the sparse autoencoder network of the stack for replacing two layers of characteristic layer；
Step 4：According to the fisrt feature layer of the sparse autoencoder network of the stack of two layers of characteristic layer, second layer characteristic layer and
Softmax classification layer building preliminary classification network, is optimized after preliminary classification network is optimized by back-propagation algorithm and is classified
Network；
Step 5：Using test set as sorter network input data after optimization, building neuronal quantity optimizes first layer feature
Layer and neuronal quantity optimize second layer characteristic layer, and it is excellent to optimize first layer characteristic layer, neuronal quantity by neuronal quantity
The Softmax classification layer of sorter network obtains the sorter network of neuronal quantity optimization after changing second layer characteristic layer and optimizing；
Preferably, sample size is m in training set described in step 1, sample size is K in test set；
Preferably, using training set sample as first layer autocoder input data being by step 1 described in step 2
Described in sample x in training set(i)(i∈[1,m],x(i)Expression characteristic dimension is s1Vector), as input data composition the
One layer of input layer；
Described in step 2 by successively greedy coaching method training obtain first layer autocoder each layer weight coefficient and
Each layer bias vector be：
To x(i)It performs the encoding operation to obtain the further feature y of first layer autocoder characteristic layer(i)：
y(i)=s (W1x(i)+b1)
Wherein, W1For first layer autocoder characteristic layer weight coefficient, b1For the biasing of first layer autocoder characteristic layer
Vector matrix, s () are activation primitive, and the present invention is sigmoid function；
To the further feature y of first layer autocoder characteristic layer(i)It is decoded operation and obtains first layer autocoder
Output layer reconstructs vector
Wherein, W '1For first layer autocoder output layer weight coefficient, b '1It is inclined for first layer autocoder output layer
Vector matrix is set, s () is activation primitive, and the present invention is sigmoid function；
Constructing first layer autocoder cross entropy function model is:
Wherein, m is the quantity of sample in training set described in step 1, and W is the weight coefficient of autoencoder network model, and b is certainly
The bias vector matrix of coding network model；
In first layer autocoder characteristic layer, sparse constraint is added to each neuron and realizes sparse expression, neuron j
Average activity be：
Wherein, fθ(x(i)) it be input data is sample x(i)When neuron j activity, defining average activity coefficient is
ρ, average activityThe relative entropy for being ρ with average activity coefficient is：
The sparse self-encoding encoder loss function of first layer autocoder is:
Wherein, s2It is the quantity of neuron in first layer autocoder characteristic layer, j is neuron serial number, and β is sparse system
Number；
When self-encoding encoder loss function minimum sparse to first layer autocoder by using layer-by-layer greedy coaching method into
Row Optimization Solution obtains first layer autocoder characteristic layer weight coefficient W1, first layer autocoder characteristic layer is biased towards
Moment matrix b1, first layer autocoder output layer weight coefficient W '1And first layer autocoder output layer bias vector square
Battle array b '1；
Lead to described in step 2 using the further feature of first layer autocoder as second layer autocoder input data
It crosses successively greedy coaching method training and obtains each layer weight coefficient and each layer bias vector of second layer autocoder：
By the further feature y of first layer autocoder characteristic layer after layer-by-layer greedy coaching method training(i)As input data
As second layer autocoder input layer, the further feature of second layer autocoder characteristic layer is obtained by encoding operation
For：
y'(i)=s (W2y(i)+b2)
Wherein, W2For second layer autocoder characteristic layer weight coefficient, b2For the biasing of second layer autocoder characteristic layer
Vector matrix, s () are activation primitive, and the present invention is sigmoid function；
To the further feature y' of second layer autocoder characteristic layer(i)It is decoded operation, obtains second layer autocoding
Device output layer reconstructs vector：
Wherein, W2' it is second layer autocoder output layer weight coefficient, b'2It is inclined for second layer autocoder output layer
Vector matrix is set, s () is activation primitive, and the present invention is sigmoid function；
The sparse self-encoding encoder loss function of second layer autocoder is constructed according to step 2, certainly in conjunction with the second layer
The quantity s of neuron in dynamic encoder feature layer3, sparse to second layer autocoder certainly by using layer-by-layer greedy coaching method
It is optimized when encoder loss function minimum, obtains second layer autocoder characteristic layer weight coefficient W2, the second layer
Autocoder characteristic layer bias vector matrix b2, second layer autocoder output layer weight coefficient W2' and the second layer it is automatic
Encoder output layer bias vector matrix b'2；
The stack of two layers of characteristic layer is constructed described in step 2 by first layer autocoder and second layer autocoder
The sparse autoencoder network of formula：
It is after successively greedy coaching method training, first layer autocoder input layer is dilute as the stack of two layers of characteristic layer
The input layer for dredging autoencoder network is sparse from coding net using first layer autocoder characteristic layer as the stack of two layers of characteristic layer
The fisrt feature layer of network is sparse from coding net using second layer characteristic layer autocoder characteristic layer as the stack of two layers of characteristic layer
The second feature layer of network, using second layer autocoder output layer as the defeated of the sparse autoencoder network of the stack of two layers of characteristic layer
Layer out constructs the sparse autoencoder network of stack of two layers of characteristic layer；
Preferably, by the further feature y ' of second autocoder described in step 2 described in step 3(i)As input
Data, for a k classification problem, need to export a k dimensional vector and carry out the general of this k estimation to train Softmax classification layer
Rate value, it is assumed that function hθ(x)：
Wherein, θ is the matrixes that whole parameters of model are a k × (n+1) dimension, and subscript refers to a certain specific neuron
Parameter, subscript T represent the transposition of matrix, and n is y '(i)Characteristic dimension；
It is by gradient descent algorithm Optimization Solution Softmax classifier described in step 3：
To input data y '(i)It carries out a Softmax to return, establishing a Softmax recurrence cost function model is：
Wherein, m is the quantity of sample in training set described in step 1, and i is sample serial number, and k is the categorical measure of classification, and j is
Classification sequence number, θ are whole parameters of model, and subscript refers to the parameter of the Softmax classification a certain specific neuron of layer, and n is y '(i)
Characteristic dimension, γ is weight term attenuation coefficient；
Carrying out derivation to cost function is：
Wherein, m is the quantity of sample in training set described in step 1, and i is sample serial number, and k is the categorical measure of classification, and j is
Classification sequence number, θ are whole parameters of model, and subscript refers to the parameter of the Softmax classification a certain specific neuron of layer, and n is y '(i)
Characteristic dimension, γ is weight term attenuation coefficient；
It is minimised as optimization aim with J (θ), by gradient descent algorithm Optimization Solution, thus Softmax after being trained
Classifier；
Preferably, the fisrt feature layer of the sparse autoencoder network of the stack of two layers of characteristic layer described in step 4 is：
y(i)=s (W1x(i)+b1)
Wherein, W1For first layer characteristic layer weight coefficient, b1For first layer characteristic layer bias vector matrix, s () is activation
Function, the present invention are sigmoid function；
The second layer characteristic layer of the sparse autoencoder network of the stack of two layers of characteristic layer is：
y'(i)=s (W2y(i)+b2)
Wherein, W2For second layer autocoder characteristic layer weight coefficient, b2For the biasing of second layer autocoder characteristic layer
Vector matrix, s () are activation primitive, and the present invention is sigmoid function；
The sparse autoencoder network of the stack of two layers of characteristic layer Softmax classification layer be：
For a k classification problem, need to export the probability value that a k dimensional vector carrys out this k estimation, it is assumed that function hθ
(x) as follows：
θ is whole parameters of model, is the matrix of a k × (n+1) dimension, and subscript refers to the ginseng of a certain specific neuron
Number, subscript T represent the transposition of matrix, and n is y '(i)Characteristic dimension；
Preliminary classification network is constructed described in step 4 is：
Using the fisrt feature layer of the sparse autoencoder network of stack as the first layer characteristic layer of preliminary classification network, by stack
Second layer characteristic layer of the second layer characteristic layer of sparse autoencoder network as preliminary classification network is sparse from coding net by stack
Softmax classification layer of the Softmax classification layer of network as preliminary classification network；
Sorter network after preliminary classification network is optimized is optimized by back-propagation algorithm described in step 4：
The first layer characteristic layer of sorter network after optimization：
Wherein,For optimization after first layer characteristic layer weight coefficient,For first layer characteristic layer bias vector square after optimization
Battle array, s () are activation primitive, and the present invention is sigmoid function；
The second layer characteristic layer of sorter network after optimization：
Wherein,For optimization after second layer characteristic layer weight coefficient,For second layer characteristic layer bias vector square after optimization
Battle array, s () are activation primitive, and the present invention is sigmoid function；
The Softmax classification layer of sorter network after optimization：
Wherein,Whole parameters for model after optimization are the matrixes of a k × (n+1) dimension, and subscript refers to a certain specific mind
Parameter through member, subscript T represent the transposition of matrix, and n is y '(i)Characteristic dimension；
Preferably, the optimization first layer characteristic layer of building neuronal quantity described in step 5 is：
Sorter network obtains classification results after test set described in step 1 is input to optimization, net of classifying after traversal optimization
The neuronal quantity of the first layer characteristic layer of sorter network after neuronal quantity, that is, adjusting and optimizing of the first layer characteristic layer of network, Gu
Determine other parameters, sick age classification results and the practical sick age of test set described in step 1 are compared, will be compared with practical sick age
Neuronal quantity of the smallest neuronal quantity of resultant error as the first layer characteristic layer of sorter network after optimization, obtains nerve
First quantity optimization first layer characteristic layer；
Building neuronal quantity described in step 5 optimizes second layer characteristic layer：
Sorter network after the neuronal quantity that test set described in step 1 is input to first layer characteristic layer is optimized, traversal
The second layer characteristic layer of sorter network after neuronal quantity, that is, adjusting and optimizing of the second layer characteristic layer of sorter network after optimization
Neuronal quantity, fixed other parameters compare sick age classification results and the practical sick age of test set described in step 1, will
Nerve with the practical sick age the smallest neuronal quantity of comparing result error as the second layer characteristic layer of sorter network after optimization
First quantity obtains neuronal quantity optimization second layer characteristic layer；
Compared with prior art, the invention has the advantages that improving the accuracy and stability of classification.
Detailed description of the invention
Fig. 1：The method of the present invention flow chart；
Fig. 2：The Raman spectrogram of different disease age alzheimer's disease small white mouse blood platelets；
Fig. 3：The training schematic diagram of two layers of stack autoencoder network first layer；
Fig. 4：The training schematic diagram of two layers of stack autoencoder network second layer；
Fig. 5：The schematic diagram of Softmax classification layer；
Fig. 6：The network structure of two layers of stack autoencoder network combination Softmax classifier.
Specific embodiment
Understand for the ease of those of ordinary skill in the art and implement the present invention, with reference to the accompanying drawing and implements example to this
Invention is described in further detail, it should be understood that and implementation example described herein is merely to illustrate and explain the present invention, and
It is not used in the restriction present invention.
Embodiments of the present invention are introduced below with reference to Fig. 1 to Fig. 6, embodiments of the present invention specifically include following step
Suddenly：
Step 1：The Raman spectrum of the Alzheimer's blood platelet in different sick ages is extracted by optical tweezer Raman system
Figure, and Raman spectrogram progress background is reduced, baseline correction, smooth and average treatment, by the Raman spectrogram after processing
It is randomly divided into training set and test set；
Sample size is m=328 in training set described in step 1, and sample size is K=40 in test set；
Step 2：Using training set sample as first layer autocoder input data, pass through layer-by-layer greedy coaching method training
Each layer weight coefficient and each layer bias vector of first layer autocoder are obtained, by the further feature of first layer autocoder
Each layer of second layer autocoder is obtained by successively greedy coaching method training as second layer autocoder input data
Weight coefficient and each layer bias vector construct two layers of characteristic layer by first layer autocoder and second layer autocoder
The sparse autoencoder network of stack；
Being as first layer autocoder input data using training set sample described in step 2 will instruction described in step 1
Practice and concentrates sample x(i)(i∈[1,328],x(i)Expression characteristic dimension is s1=440 vector), as input data composition first
Layer input layer；
Described in step 2 by successively greedy coaching method training obtain first layer autocoder each layer weight coefficient and
Each layer bias vector be：
To x(i)It performs the encoding operation to obtain the further feature y of first layer autocoder characteristic layer(i)：
y(i)=s (W1x(i)+b1)
Wherein, W1For first layer autocoder characteristic layer weight coefficient, b1For the biasing of first layer autocoder characteristic layer
Vector matrix, s () are activation primitive, and the present invention is sigmoid function；
To the further feature y of first layer autocoder characteristic layer(i)It is decoded operation and obtains first layer autocoder
Output layer reconstructs vector
Wherein, W '1For first layer autocoder output layer weight coefficient, b '1It is inclined for first layer autocoder output layer
Vector matrix is set, s () is activation primitive, and the present invention is sigmoid function；
Constructing first layer autocoder cross entropy function model is:
Wherein, m is the quantity of sample in training set described in step 1, and W is the weight coefficient of autoencoder network model, and b is certainly
The bias vector matrix of coding network model；
In first layer autocoder characteristic layer, sparse constraint is added to each neuron and realizes sparse expression, neuron j
Average activity be：
Wherein, fθ(x(i)) it be input data is sample x(i)When neuron j activity, defining average activity coefficient is
ρ=0.05, average activityThe relative entropy for being ρ with average activity coefficient is：
The sparse self-encoding encoder loss function of first layer autocoder is:
Wherein, s2=400 be the quantity of neuron in first layer autocoder characteristic layer, and j is neuron serial number, β=
0.1 is sparse coefficient；
When self-encoding encoder loss function minimum sparse to first layer autocoder by using layer-by-layer greedy coaching method into
Row Optimization Solution obtains first layer autocoder characteristic layer weight coefficient W1, first layer autocoder characteristic layer is biased towards
Moment matrix b1, first layer autocoder output layer weight coefficient W '1And first layer autocoder output layer bias vector square
Battle array b '1；
Lead to described in step 2 using the further feature of first layer autocoder as second layer autocoder input data
It crosses successively greedy coaching method training and obtains each layer weight coefficient and each layer bias vector of second layer autocoder：
By the further feature y of first layer autocoder characteristic layer after layer-by-layer greedy coaching method training(i)As input data
As second layer autocoder input layer, the further feature of second layer autocoder characteristic layer is obtained by encoding operation
For：
y'(i)=s (W2y(i)+b2)
Wherein, W2For second layer autocoder characteristic layer weight coefficient, b2For the biasing of second layer autocoder characteristic layer
Vector matrix, s () are activation primitive, and the present invention is sigmoid function；
To the further feature y' of second layer autocoder characteristic layer(i)It is decoded operation, obtains second layer autocoding
Device output layer reconstructs vector：
Wherein, W2' it is second layer autocoder output layer weight coefficient, b'2It is inclined for second layer autocoder output layer
Vector matrix is set, s () is activation primitive, and the present invention is sigmoid function；
The sparse self-encoding encoder loss function of second layer autocoder is constructed according to step 2, certainly in conjunction with the second layer
The quantity s of neuron in dynamic encoder feature layer3=60, it is dilute to second layer autocoder by using layer-by-layer greedy coaching method
It is optimized when dredging self-encoding encoder loss function minimum, obtains second layer autocoder characteristic layer weight coefficient W2,
Two layers of autocoder characteristic layer bias vector matrix b2, second layer autocoder output layer weight coefficient W2' and the second layer
Autocoder output layer bias vector matrix b'2；
The stack of two layers of characteristic layer is constructed described in step 2 by first layer autocoder and second layer autocoder
The sparse autoencoder network of formula：
It is after successively greedy coaching method training, first layer autocoder input layer is dilute as the stack of two layers of characteristic layer
The input layer for dredging autoencoder network is sparse from coding net using first layer autocoder characteristic layer as the stack of two layers of characteristic layer
The fisrt feature layer of network is sparse from coding net using second layer characteristic layer autocoder characteristic layer as the stack of two layers of characteristic layer
The second feature layer of network, using second layer autocoder output layer as the defeated of the sparse autoencoder network of the stack of two layers of characteristic layer
Layer out constructs the sparse autoencoder network of stack of two layers of characteristic layer；
Step 3：Using the further feature of second layer autocoder described in step 2 as input data with training
Softmax classifier, by gradient descent algorithm Optimization Solution Softmax classifier, using Softmax classifier after training as
Softmax classifies layer with the output layer of the sparse autoencoder network of the stack for replacing two layers of characteristic layer；
By the further feature y ' of second autocoder described in step 2 described in step 3(i)As input data to instruct
Practice Softmax classification layer, for a k=4 classification problem, needs to export a k=4 dimensional vector and carry out the general of this k=4 estimation
Rate value, it is assumed that function hθ(x)：
Wherein, θ is the matrixes that whole parameters of model are a k × (n+1) dimension, and subscript refers to a certain specific neuron
Parameter, subscript T represent the transposition of matrix, and n is y '(i)Characteristic dimension；
It is by gradient descent algorithm Optimization Solution Softmax classifier described in step 3：
To input data y '(i)It carries out a Softmax to return, establishing a Softmax recurrence cost function model is：
Wherein, m=328 is the quantity of sample in training set described in step 1, and i is sample serial number, and k=4 is the classification of classification
Quantity, j are classification sequence number, and θ is whole parameters of model, and subscript refers to the parameter of the Softmax classification a certain specific neuron of layer,
N is y '(i)Characteristic dimension, γ=3e^ (- 3) is weight term attenuation coefficient；
Carrying out derivation to cost function is：
Wherein, m=328 is the quantity of sample in training set described in step 1, and i is sample serial number, and k=4 is the classification of classification
Quantity, j are classification sequence number, and θ is whole parameters of model, and subscript refers to the parameter of the Softmax classification a certain specific neuron of layer,
N is y '(i)Characteristic dimension, γ=3e^ (- 3) is weight term attenuation coefficient；
It is minimised as optimization aim with J (θ), by gradient descent algorithm Optimization Solution, thus Softmax after being trained
Classifier；
Step 4：According to the fisrt feature layer of the sparse autoencoder network of the stack of two layers of characteristic layer, second layer characteristic layer and
Softmax classification layer building preliminary classification network, is optimized after preliminary classification network is optimized by back-propagation algorithm and is classified
Network；
The fisrt feature layer of the sparse autoencoder network of the stack of two layers of characteristic layer described in step 4 is：
y(i)=s (W1x(i)+b1)
Wherein, W1For first layer characteristic layer weight coefficient, b1For first layer characteristic layer bias vector matrix, s () is activation
Function, the present invention are sigmoid function；
The second layer characteristic layer of the sparse autoencoder network of the stack of two layers of characteristic layer is：
y'(i)=s (W2y(i)+b2)
Wherein, W2For second layer autocoder characteristic layer weight coefficient, b2For the biasing of second layer autocoder characteristic layer
Vector matrix, s () are activation primitive, and the present invention is sigmoid function；
The sparse autoencoder network of the stack of two layers of characteristic layer Softmax classification layer be：
For a k=4 classification problem, need to export the probability value that a k=4 dimensional vector carrys out this k=4 estimation, it is false
If function hθ(x) as follows：
θ is whole parameters of model, is the matrix of a k × (n+1) dimension, and subscript refers to the ginseng of a certain specific neuron
Number, subscript T represent the transposition of matrix, and n is y '(i)Characteristic dimension；
Preliminary classification network is constructed described in step 4 is：
Using the fisrt feature layer of the sparse autoencoder network of stack as the first layer characteristic layer of preliminary classification network, by stack
Second layer characteristic layer of the second layer characteristic layer of sparse autoencoder network as preliminary classification network is sparse from coding net by stack
Softmax classification layer of the Softmax classification layer of network as preliminary classification network；
Sorter network after preliminary classification network is optimized is optimized by back-propagation algorithm described in step 4：
The first layer characteristic layer of sorter network after optimization：
Wherein,For optimization after first layer characteristic layer weight coefficient,For first layer characteristic layer bias vector square after optimization
Battle array, s () are activation primitive, and the present invention is sigmoid function；
The second layer characteristic layer of sorter network after optimization：
Wherein,For optimization after second layer characteristic layer weight coefficient,For second layer characteristic layer bias vector square after optimization
Battle array, s () are activation primitive, and the present invention is sigmoid function；
The Softmax classification layer of sorter network after optimization：
Wherein,Whole parameters for model after optimization are the matrixes of a k × (n+1) dimension, and subscript refers to a certain specific mind
Parameter through member, subscript T represent the transposition of matrix, and n is y '(i)Characteristic dimension；
Step 5：Using test set as sorter network input data after optimization, building neuronal quantity optimizes first layer feature
Layer and neuronal quantity optimize second layer characteristic layer, and it is excellent to optimize first layer characteristic layer, neuronal quantity by neuronal quantity
The Softmax classification layer of sorter network obtains the sorter network of neuronal quantity optimization after changing second layer characteristic layer and optimizing；
Building neuronal quantity described in step 5 optimizes first layer characteristic layer：
Sorter network obtains classification results after test set described in step 1 is input to optimization, net of classifying after traversal optimization
The neuronal quantity of the first layer characteristic layer of sorter network after neuronal quantity, that is, adjusting and optimizing of the first layer characteristic layer of network, Gu
Determine other parameters, sick age classification results and the practical sick age of test set described in step 1 are compared, will be compared with practical sick age
Neuronal quantity of the smallest neuronal quantity of resultant error as the first layer characteristic layer of sorter network after optimization, obtains nerve
First quantity optimization first layer characteristic layer；
Building neuronal quantity described in step 5 optimizes second layer characteristic layer：
Sorter network after the neuronal quantity that test set described in step 1 is input to first layer characteristic layer is optimized, traversal
The second layer characteristic layer of sorter network after neuronal quantity, that is, adjusting and optimizing of the second layer characteristic layer of sorter network after optimization
Neuronal quantity, fixed other parameters compare sick age classification results and the practical sick age of test set described in step 1, will
Nerve with the practical sick age the smallest neuronal quantity of comparing result error as the second layer characteristic layer of sorter network after optimization
First quantity obtains neuronal quantity optimization second layer characteristic layer；
It should be understood that the above-mentioned description for preferred embodiment is more detailed, can not therefore be considered to this
The limitation of invention patent protection range, those skilled in the art under the inspiration of the present invention, are not departing from power of the present invention
Benefit requires to make replacement or deformation under protected ambit, fall within the scope of protection of the present invention, this hair
It is bright range is claimed to be determined by the appended claims.
Claims (6)
1. a kind of Raman spectrum data classification method based on autoencoder network, which is characterized in that include the following steps：
Step 1：The Raman spectrogram of the Alzheimer's blood platelet in different sick ages is extracted by optical tweezer Raman system, and
Raman spectrogram progress background is reduced, baseline correction, smooth and average treatment, the Raman spectrogram after processing is divided at random
For training set and test set；
Step 2：Using training set sample as first layer autocoder input data, obtained by layer-by-layer greedy coaching method training
Each layer weight coefficient and each layer bias vector of first layer autocoder, using the further feature of first layer autocoder as
Second layer autocoder input data obtains each layer weight of second layer autocoder by successively greedy coaching method training
Coefficient and each layer bias vector construct the stack of two layers of characteristic layer by first layer autocoder and second layer autocoder
The sparse autoencoder network of formula；
Step 3：The further feature of second layer autocoder described in step 2 is trained Softmax points as input data
Class device, by gradient descent algorithm Optimization Solution Softmax classifier, using Softmax classifier after training as Softmax points
Class layer is with the output layer of the sparse autoencoder network of the stack for replacing two layers of characteristic layer；
Step 4：According to the fisrt feature layer of the sparse autoencoder network of the stack of two layers of characteristic layer, second layer characteristic layer and
Softmax classification layer building preliminary classification network, is optimized after preliminary classification network is optimized by back-propagation algorithm and is classified
Network；
Step 5：Using test set as optimization after sorter network input data, building neuronal quantity optimization first layer characteristic layer with
And neuronal quantity optimizes second layer characteristic layer, optimizes first layer characteristic layer, neuronal quantity optimization the by neuronal quantity
The Softmax classification layer of sorter network obtains the sorter network of neuronal quantity optimization after two layers of characteristic layer and optimization.
2. the Raman spectrum data classification method according to claim 1 based on autoencoder network, it is characterised in that：Step
Sample size is m in training set described in 1, and sample size is K in test set.
3. the Raman spectrum data classification method according to claim 1 based on autoencoder network, it is characterised in that：Step
It using training set sample as first layer autocoder input data is by sample x in training set described in step 1 described in 2(i)
(i∈[1,m],x(i)Expression characteristic dimension is s1Vector), as input data form first layer input layer；
Each layer weight coefficient and each layer of first layer autocoder are obtained by successively greedy coaching method training described in step 2
Bias vector is：
To x(i)It performs the encoding operation to obtain the further feature y of first layer autocoder characteristic layer(i)：
y(i)=s (W1x(i)+b1)
Wherein, W1For first layer autocoder characteristic layer weight coefficient, b1For first layer autocoder characteristic layer bias vector
Matrix, s () are activation primitive, and the present invention is sigmoid function；
To the further feature y of first layer autocoder characteristic layer(i)It is decoded operation and obtains the output of first layer autocoder
Layer reconstruct vector
Wherein, W '1For first layer autocoder output layer weight coefficient, b '1It is biased towards for first layer autocoder output layer
Moment matrix, s () are activation primitive, and the present invention is sigmoid function；
Constructing first layer autocoder cross entropy function model is:
Wherein, m is the quantity of sample in training set described in step 1, and W is the weight coefficient of autoencoder network model, and b is to encode certainly
The bias vector matrix of network model；
In first layer autocoder characteristic layer, sparse constraint is added to each neuron and realizes sparse expression, neuron j's is flat
Equal activity is：
Wherein, fθ（x(i)) it be input data is sample x(i)When neuron j activity, defining average activity coefficient is ρ, is put down
Equal activityThe relative entropy for being ρ with average activity coefficient is：
The sparse self-encoding encoder loss function of first layer autocoder is:
Wherein, s2It is the quantity of neuron in first layer autocoder characteristic layer, j is neuron serial number, and β is sparse coefficient；
It is carried out when self-encoding encoder loss function minimum sparse to first layer autocoder by using layer-by-layer greediness coaching method excellent
Change and solve, obtains first layer autocoder characteristic layer weight coefficient W1, first layer autocoder characteristic layer bias vector square
Battle array b1, first layer autocoder output layer weight coefficient W '1And first layer autocoder output layer bias vector matrix
b′1；
Described in step 2 using the further feature of first layer autocoder as second layer autocoder input data by by
The greedy coaching method training of layer obtains each layer weight coefficient and each layer bias vector of second layer autocoder：
By the further feature y of first layer autocoder characteristic layer after layer-by-layer greedy coaching method training(i)As input data conduct
Second layer autocoder input layer is by the further feature that encoding operation obtains second layer autocoder characteristic layer：
y'(i)=s (W2y(i)+b2)
Wherein, W2For second layer autocoder characteristic layer weight coefficient, b2For second layer autocoder characteristic layer bias vector
Matrix, s () are activation primitive, and the present invention is sigmoid function；
To the further feature y' of second layer autocoder characteristic layer(i)It is decoded operation, it is defeated to obtain second layer autocoder
Layer reconstruct vector is out：
Wherein, W '2For second layer autocoder output layer weight coefficient, b '2It is biased towards for second layer autocoder output layer
Moment matrix, s () are activation primitive, and the present invention is sigmoid function；
The sparse self-encoding encoder loss function of second layer autocoder is constructed according to step 2, is compiled automatically in conjunction with the second layer
The quantity s of neuron in code device characteristic layer3, sparse to second layer autocoder from coding by using layer-by-layer greedy coaching method
It is optimized when device loss function minimum, obtains second layer autocoder characteristic layer weight coefficient W2, the second layer it is automatic
Encoder feature layer bias vector matrix b2, second layer autocoder output layer weight coefficient W '2And second layer autocoding
Device output layer bias vector matrix b '2；
The stack for constructing two layers of characteristic layer by first layer autocoder and second layer autocoder described in step 2 is dilute
Dredge autoencoder network：
It is sparse certainly using first layer autocoder input layer as the stack of two layers of characteristic layer after successively greedy coaching method training
The input layer of coding network, using first layer autocoder characteristic layer as the sparse autoencoder network of the stack of two layers of characteristic layer
Fisrt feature layer, using second layer characteristic layer autocoder characteristic layer as the sparse autoencoder network of the stack of two layers of characteristic layer
Second feature layer, using second layer autocoder output layer as the output of the sparse autoencoder network of the stack of two layers of characteristic layer
Layer constructs the sparse autoencoder network of stack of two layers of characteristic layer.
4. the Raman spectrum data classification method according to claim 1 based on autoencoder network, it is characterised in that：Step
The 3 further feature y ' by second autocoder described in step 2(i)As input data to train Softmax to classify
Layer needs to export the probability value that a k dimensional vector carrys out this k estimation, it is assumed that function h for a k classification problemθ(x)：
Wherein, θ is the matrixes that whole parameters of model are a k × (n+1) dimension, and subscript refers to the ginseng of a certain specific neuron
Number, subscript T represent the transposition of matrix, and n is y '(i)Characteristic dimension；
It is by gradient descent algorithm Optimization Solution Softmax classifier described in step 3：
To input data y '(i)It carries out a Softmax to return, establishing a Softmax recurrence cost function model is：
Wherein, m is the quantity of sample in training set described in step 1, and i is sample serial number, and k is the categorical measure of classification, and j is classification
Serial number, θ are whole parameters of model, and subscript refers to the parameter of the Softmax classification a certain specific neuron of layer, and n is y '(i)Spy
Dimension is levied, γ is weight term attenuation coefficient；
Carrying out derivation to cost function is：
Wherein, m is the quantity of sample in training set described in step 1, and i is sample serial number, and k is the categorical measure of classification, and j is classification
Serial number, θ are whole parameters of model, and subscript refers to the parameter of the Softmax classification a certain specific neuron of layer, and n is y '(i)Spy
Dimension is levied, γ is weight term attenuation coefficient；
It is minimised as optimization aim with J (θ), by gradient descent algorithm Optimization Solution, so that Softmax classifies after being trained
Device.
5. the Raman spectrum data classification method according to claim 1 based on autoencoder network, it is characterised in that：Step
The fisrt feature layer of the sparse autoencoder network of the stack of two layers of characteristic layer described in 4 is：
y(i)=s (W1x(i)+b1)
Wherein, W1For first layer characteristic layer weight coefficient, b1For first layer characteristic layer bias vector matrix, s () is activation letter
Number, the present invention are sigmoid function；
The second layer characteristic layer of the sparse autoencoder network of the stack of two layers of characteristic layer is：
y'(i)=s (W2y(i)+b2)
Wherein, W2For second layer autocoder characteristic layer weight coefficient, b2For second layer autocoder characteristic layer bias vector
Matrix, s () are activation primitive, and the present invention is sigmoid function；
The sparse autoencoder network of the stack of two layers of characteristic layer Softmax classification layer be：
For a k classification problem, need to export the probability value that a k dimensional vector carrys out this k estimation, it is assumed that function hθ(x) such as
Under：
θ is whole parameters of model, is the matrix of a k × (n+1) dimension, and subscript refers to the parameter of a certain specific neuron,
Subscript T represents the transposition of matrix, and n is y '(i)Characteristic dimension；
Preliminary classification network is constructed described in step 4 is：
It is using the fisrt feature layer of the sparse autoencoder network of stack as the first layer characteristic layer of preliminary classification network, stack is sparse
Second layer characteristic layer of the second layer characteristic layer of autoencoder network as preliminary classification network, by the sparse autoencoder network of stack
Softmax classification layer of the Softmax classification layer as preliminary classification network；
Sorter network after preliminary classification network is optimized is optimized by back-propagation algorithm described in step 4：
The first layer characteristic layer of sorter network after optimization：
Wherein,For optimization after first layer characteristic layer weight coefficient,For first layer characteristic layer bias vector matrix after optimization, s
() is activation primitive, and the present invention is sigmoid function；
The second layer characteristic layer of sorter network after optimization：
Wherein,For optimization after second layer characteristic layer weight coefficient,For second layer characteristic layer bias vector matrix after optimization, s
() is activation primitive, and the present invention is sigmoid function；
The Softmax classification layer of sorter network after optimization：
Wherein,Whole parameters for model after optimization are the matrixes of a k × (n+1) dimension, and subscript refers to a certain specific neuron
Parameter, subscript T represents the transposition of matrix, and n is y '(i)Characteristic dimension.
6. the Raman spectrum data classification method according to claim 1 based on autoencoder network, it is characterised in that：Step
Building neuronal quantity described in 5 optimizes first layer characteristic layer：
Sorter network obtains classification results after test set described in step 1 is input to optimization, sorter network after traversal optimization
The neuronal quantity of the first layer characteristic layer of sorter network, fixes it after neuronal quantity, that is, adjusting and optimizing of first layer characteristic layer
His parameter compares sick age classification results and test set described in step 1 practical sick age, will be with practical sick age comparing result
Neuronal quantity of the smallest neuronal quantity of error as the first layer characteristic layer of sorter network after optimization, obtains neuron number
Amount optimization first layer characteristic layer；
Building neuronal quantity described in step 5 optimizes second layer characteristic layer：
Sorter network after the neuronal quantity that test set described in step 1 is input to first layer characteristic layer is optimized, traversal optimization
Afterwards after neuronal quantity, that is, adjusting and optimizing of the second layer characteristic layer of sorter network the second layer characteristic layer of sorter network nerve
First quantity, fixed other parameters compare sick age classification results and the practical sick age of test set described in step 1, will be with reality
Neuron number of the border disease age the smallest neuronal quantity of comparing result error as the second layer characteristic layer of sorter network after optimization
Amount obtains neuronal quantity optimization second layer characteristic layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810857856.9A CN108921233B (en) | 2018-07-31 | 2018-07-31 | Raman spectrum data classification method based on self-coding network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810857856.9A CN108921233B (en) | 2018-07-31 | 2018-07-31 | Raman spectrum data classification method based on self-coding network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108921233A true CN108921233A (en) | 2018-11-30 |
| CN108921233B CN108921233B (en) | 2021-06-04 |
Family
ID=64393277
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810857856.9A Active CN108921233B (en) | 2018-07-31 | 2018-07-31 | Raman spectrum data classification method based on self-coding network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108921233B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109816002A (en) * | 2019-01-11 | 2019-05-28 | 广东工业大学 | The single sparse self-encoding encoder detection method of small target migrated certainly based on feature |
| CN110561192A (en) * | 2019-09-11 | 2019-12-13 | 大连理工大学 | Deep hole boring cutter state monitoring method based on stacking self-encoder |
| CN110895565A (en) * | 2019-11-29 | 2020-03-20 | 国网湖南省电力有限公司 | Method and system for classifying fault defect texts of power equipment |
| CN110974213A (en) * | 2019-12-20 | 2020-04-10 | 哈尔滨理工大学 | Electrocardiosignal identification method based on deep stack network |
| CN111325233A (en) * | 2019-10-29 | 2020-06-23 | 国网辽宁省电力有限公司电力科学研究院 | Transformer fault detection method and device |
| CN111832993A (en) * | 2020-07-13 | 2020-10-27 | 深圳市今天国际物流技术股份有限公司 | Predictive maintenance method for warehouse logistics system and related components |
| CN112348114A (en) * | 2020-11-27 | 2021-02-09 | 上海交通大学烟台信息技术研究院 | Self-detection classification method and system for Raman spectrum gas markers |
| CN114512236A (en) * | 2022-04-18 | 2022-05-17 | 山东师范大学 | Intelligent auxiliary diagnosis system for Alzheimer's disease |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104346617A (en) * | 2014-04-18 | 2015-02-11 | 南京信息工程大学 | Cell detection method based on sliding window and depth structure extraction features |
| CN106503654A (en) * | 2016-10-24 | 2017-03-15 | 中国地质大学（武汉） | A kind of face emotion identification method based on the sparse autoencoder network of depth |
| US20170235996A1 (en) * | 2015-07-28 | 2017-08-17 | Chiman KWAN | Method and system for collaborative multi-satellite remote sensing |
-
2018
- 2018-07-31 CN CN201810857856.9A patent/CN108921233B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104346617A (en) * | 2014-04-18 | 2015-02-11 | 南京信息工程大学 | Cell detection method based on sliding window and depth structure extraction features |
| US20170235996A1 (en) * | 2015-07-28 | 2017-08-17 | Chiman KWAN | Method and system for collaborative multi-satellite remote sensing |
| CN106503654A (en) * | 2016-10-24 | 2017-03-15 | 中国地质大学（武汉） | A kind of face emotion identification method based on the sparse autoencoder network of depth |
Non-Patent Citations (2)
| Title |
|---|
| PUSHKAR B 等: "Early Diagnosis of Alzheimer"s Disease: A Multi- class Deep Learning Framework with Modified k- sparse Autoencoder Classification", 《IEEE IMAGE AND VISION COMPUTING NEW ZEALAND》 * |
| 孙锡龙: "等离激元增强拉曼光谱预处理和识别算法的研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109816002A (en) * | 2019-01-11 | 2019-05-28 | 广东工业大学 | The single sparse self-encoding encoder detection method of small target migrated certainly based on feature |
| CN109816002B (en) * | 2019-01-11 | 2022-09-06 | 广东工业大学 | Single sparse self-encoder weak and small target detection method based on feature self-migration |
| CN110561192A (en) * | 2019-09-11 | 2019-12-13 | 大连理工大学 | Deep hole boring cutter state monitoring method based on stacking self-encoder |
| CN111325233A (en) * | 2019-10-29 | 2020-06-23 | 国网辽宁省电力有限公司电力科学研究院 | Transformer fault detection method and device |
| CN111325233B (en) * | 2019-10-29 | 2024-02-09 | 国网辽宁省电力有限公司电力科学研究院 | Transformer fault detection method and device |
| CN110895565A (en) * | 2019-11-29 | 2020-03-20 | 国网湖南省电力有限公司 | Method and system for classifying fault defect texts of power equipment |
| CN110974213A (en) * | 2019-12-20 | 2020-04-10 | 哈尔滨理工大学 | Electrocardiosignal identification method based on deep stack network |
| CN111832993A (en) * | 2020-07-13 | 2020-10-27 | 深圳市今天国际物流技术股份有限公司 | Predictive maintenance method for warehouse logistics system and related components |
| CN112348114A (en) * | 2020-11-27 | 2021-02-09 | 上海交通大学烟台信息技术研究院 | Self-detection classification method and system for Raman spectrum gas markers |
| CN112348114B (en) * | 2020-11-27 | 2023-05-30 | 上海交通大学烟台信息技术研究院 | Self-detection classification method and system for Raman spectrum gas marker |
| CN114512236A (en) * | 2022-04-18 | 2022-05-17 | 山东师范大学 | Intelligent auxiliary diagnosis system for Alzheimer's disease |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108921233B (en) | 2021-06-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108921233A (en) | A kind of Raman spectrum data classification method based on autoencoder network | |
| CN106683081B (en) | Brain glioma molecular marker nondestructive prediction method and system based on image omics | |
| Ahmadi et al. | QAIS-DSNN: tumor area segmentation of MRI image with optimized quantum matched-filter technique and deep spiking neural network | |
| CN109528197B (en) | Individual prediction method and system for mental diseases based on brain function map | |
| CN109165667A (en) | Based on the cerebral disease categorizing system from attention mechanism | |
| Lu et al. | Classification of Camellia (Theaceae) species using leaf architecture variations and pattern recognition techniques | |
| CN107368671A (en) | System and method are supported in benign gastritis pathological diagnosis based on big data deep learning | |
| CN108511056A (en) | Therapeutic scheme based on patients with cerebral apoplexy similarity analysis recommends method and system | |
| Qu et al. | An experimental study of data heterogeneity in federated learning methods for medical imaging | |
| Ouchtati et al. | Novel method for brain tumor classification based on use of image entropy and seven Hu’s invariant moments | |
| Koziarski et al. | DiagSet: a dataset for prostate cancer histopathological image classification | |
| CN108647707A (en) | Probabilistic neural network creation method, method for diagnosing faults and device, storage medium | |
| CN116664931A (en) | Knee osteoarthritis grading method based on quantum-to-classical migration learning | |
| Abd Elaziz et al. | An efficient artificial rabbits optimization based on mutation strategy for skin cancer prediction | |
| Liu et al. | Multi-LSTM networks for accurate classification of attention deficit hyperactivity disorder from resting-state fMRI data | |
| Harun et al. | Color image enhancement of acute leukemia cells in blood microscopic image for leukemia detection sample | |
| CN116259415A (en) | Patient medicine taking compliance prediction method based on machine learning | |
| Yerukalareddy et al. | Brain tumor classification based on mr images using GAN as a pre-trained model | |
| Mamun et al. | Lcdctcnn: Lung cancer diagnosis of ct scan images using cnn based model | |
| Peng et al. | Blood vessels segmentation by using cdnet | |
| Özdem et al. | A GA-based CNN model for brain tumor classification | |
| Wei et al. | Genetic U-Net: automatically designing lightweight U-shaped CNN architectures using the genetic algorithm for retinal vessel segmentation | |
| Liu et al. | A Cross-Lesion Attention Network for Accurate Diabetic Retinopathy Grading With Fundus Images | |
| Wan et al. | Depth-wise convolution with attention neural network (DWA) for pneumonia detection | |
| Meswal et al. | A weighted ensemble transfer learning approach for melanoma classification from skin lesion images |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |