CN112542161B - BP neural network voice recognition method based on double-layer PID optimization - Google Patents
BP neural network voice recognition method based on double-layer PID optimization Download PDFInfo
- Publication number
- CN112542161B CN112542161B CN202011455918.7A CN202011455918A CN112542161B CN 112542161 B CN112542161 B CN 112542161B CN 202011455918 A CN202011455918 A CN 202011455918A CN 112542161 B CN112542161 B CN 112542161B
- Authority
- CN
- China
- Prior art keywords
- layer
- neural network
- output
- error
- pid
- 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.)
- Expired - Fee Related
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 73
- 238000000034 method Methods 0.000 title claims abstract description 28
- 238000005457 optimization Methods 0.000 title claims abstract description 13
- 238000012549 training Methods 0.000 claims abstract description 30
- 230000008569 process Effects 0.000 claims abstract description 7
- 238000012360 testing method Methods 0.000 claims description 9
- 230000006870 function Effects 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 7
- 239000000284 extract Substances 0.000 claims description 4
- 238000000605 extraction Methods 0.000 claims description 3
- 238000002360 preparation method Methods 0.000 claims description 2
- 210000002569 neuron Anatomy 0.000 description 16
- 238000012545 processing Methods 0.000 description 8
- 210000004205 output neuron Anatomy 0.000 description 7
- 238000010586 diagram Methods 0.000 description 5
- 230000010355 oscillation Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000012937 correction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000004069 differentiation Effects 0.000 description 2
- 230000010354 integration Effects 0.000 description 2
- 239000002245 particle Substances 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/08—Speech classification or search
- G10L15/16—Speech classification or search using artificial neural networks
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
-
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
- G10L15/063—Training
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Feedback Control In General (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a BP neural network voice recognition method based on double-layer PID optimization, which takes an FPGA as a platform of a voice signal input method, adjusts a weight threshold value and a learning rate by using a PID algorithm, and adopts K in the double-layer PID algorithm P ,K I And K D Error E of three parameters according to system and training result g (k) The automatic adjustment is carried out, so that the weight threshold convergence of the hidden layer and the output layer is more stable, and the data fluctuation of the system is reduced; the outer PID algorithm synchronizes the updating of the learning rate with the training process of the neural network, provides a larger updating intensity in the early stage of the training of the neural network to enable the neural network to iterate quickly, reduces the updating intensity in the later stage of the training of the neural network to prevent the data from deviating from the correct value, enables the learning rate to provide a larger updating intensity in the early stage to help the system to update quickly, reduces the updating intensity in the later stage to prevent the data from deviating from the correct value, and has higher voice recognition correct rate.
Description
The technical field is as follows:
the invention relates to an artificial intelligence algorithm, in particular to a BP neural network speech recognizer optimized by double-layer PID
Background art:
as an important application direction in the field of artificial intelligence, speech recognition has been studied by many learners, such as speech recognition under deep learning, and a speech recognition method under a support vector machine. In the aspect of learning algorithm, the BP algorithm has been named after a strong nonlinear mapping and a simple structure, and has been used for a long time in the context of speech recognition, but it also has some defects, such as that a weight threshold and a learning rate cannot be determined during network initialization, convergence fluctuation is large if the setting is too large, and convergence speed is slow if the setting is small. The existing weight threshold updating formula almost depends on a negative gradient algorithm, although the negative gradient algorithm accelerates the convergence of the weight threshold to a certain degree, the numerical fluctuation caused by the negative gradient algorithm is too large, and the normal convergence of the weight threshold is influenced to a certain degree. Meanwhile, the updating of the learning rate is basically manually adjusted and set in continuous experiments, the existing variable learning rate algorithm only simply and linearly reduces the learning rate, the influence of the learning rate is reduced when the algorithm is carried out to the later stage, the output deviation from a correct result due to fluctuation caused by overlarge updating force is avoided, but the effect of the algorithm on the correct rate is not achieved, and the condition of low music identification correct rate can occur when a general BP neural network structure is used for identifying music data.
CN103639211A discloses a BP neural network and PID parameter optimization roll gap control method and system, and provides a method for optimizing PID structure by using a neural network, wherein PID parameters are more stable by the method, and the algorithm is more stable.
CN110488600A discloses an LQR optimization type brushless DC motor speed regulation neural network PID controller, and provides an algorithm for optimizing neural network regulation PID by using an LQR algorithm, so that the DC motor control is more stable.
CN104834215A discloses a BP neural network PID control algorithm optimized by variation particle swarm, and provides a BP neural network PID control algorithm optimized by variation particle swarm, and the method enables PID algorithm output to be more stable.
The invention content is as follows:
the invention aims to provide a BP neural network speech recognition method based on double-layer PID optimization aiming at the defects of the prior art.
The invention idea is as follows: and the FPGA with the voice recognition function is used as a platform of a voice signal input method. The double-layer PID is divided into an inner layer and an outer layer, the inner layer PID algorithm reduces data fluctuation caused by a negative gradient algorithm, so that the weight threshold updating process is converged smoothly, and the oscillation generated during convergence is reduced; meanwhile, the outer-layer PID algorithm synchronizes the updating of the learning rate with the training process of the neural network, provides a larger updating intensity in the early stage of the neural network training to enable the neural network to iterate rapidly, reduces the updating intensity in the later stage of the neural network training to prevent data from deviating from a correct value, and has higher voice recognition accuracy.
The purpose of the invention is realized by the following technical scheme:
firstly, three layers of BP neural networks are built, namely an input layer, a hidden layer and an output layer. Randomly generating a weighting factor W between an output layer and a hidden layer ij (k) And W jg (k) And an activation function parameter a between two layers j (k) And b g (k) Selecting a learning rate eta (k), and setting k to be 1;
secondly, the FPGA platform extracts voice data of the voice to be recognized, and the BP neural network is used for extracting the extracted data X i (k) Analyzing and calculating the output O of the BP neural network output layer g (k) While outputting Y as desired g (k) Calculating error E g (k) Using error E g (k) Proportional parameter K in Proportion Integration Differentiation (PID) algorithm P Integral parameter K I Differential parameter K D Adjusting η (k);
then, using the error E g (k) And the adjusted eta (k) corrects the weighting coefficient W of the output layer and the hidden layer of the BP neural network ij (k) And W jg (k) And an activation function parameter a between two layers j (k) And b g (k) Until the output error of the BP neural network output layer meets the requirement, finally judging whether the input voice is a set voice signal;
the method comprises the following steps:
A. carrying out training initialization preparation, initializing a neural network structure, and acquiring a sample set for training;
B. carrying out feature extraction on the sample set to obtain a feature set;
C. training a neural network by taking the feature set as a training set, and using the expected output Y in the training process g (k) And the actual output O g (k) Obtain an error E g (k) Error E g (k) Adjusting the learning rate η (k) as a parameter in the outer PID algorithm, and then continuing to use the error E g (k) And the adjusted learning rate eta (k +1) is taken as the weight W of the inner PID algorithm parameter to the input layer ij (k) Threshold value a j (k) And output layer weight W jg (k) Threshold b g (k) Carrying out adjustment;
D. testing the adjusted neural network, extracting the characteristics of the test sample according to the step (2), and identifying the test sample by the neural network in the step (3) to obtain the error E in the step (3) g (k) When error E g (k) If the value is less than a certain threshold value, the neural network training is finished, and whether the input voice signal is a set voice signal or not is judged.
In said step C, error E g (k) And adjusting the learning rate as an outer layer PID algorithm parameter, wherein an outer layer PID adjusting formula is as follows:
where s is 1,2, …, k, k is the current iteration number, N 3 Taking N as the number of output layers 3 3, whereinE gD (k)=E g (k)-E g (k-1)。
In said step C, the error E is used g (k) And the adjusted learning rate eta (k +1) is taken as the weight W of the inner PID algorithm parameter pair between the input layer and the hidden layer ij (k) Hidden layer threshold a j (k) And the weight W between the hidden layer and the output layer jg (k) Output layer threshold b g (k) Make a toneThe PID regulating formula of the outer layer is as follows,
the updating formula of the weight between the input layer and the hidden layer is as follows:
the hidden layer threshold is formulated as:
N 3 taking N as the number of output layers 3 =3,
The updating formula of the weight between the hidden layer and the output layer is as follows:
the update formula of the output layer threshold is as follows:
where s is 1,2, …, k, k is the current iteration number, N 3 Taking N as the number of output layers 3 3, whereinE gD (k)=E g (k)-E g (k-1)。
Has the advantages that: adjusting weight threshold and learning rate using PID algorithm, K in double-layer PID algorithm P ,K I And K D Error E of three parameters according to system and training result g (k) The automatic adjustment is carried out, so that the weight threshold convergence of the hidden layer and the output layer is more stable, and the data fluctuation of the system is reduced; the learning rate can provide larger updating intensity in the early stage to help the system to update quickly, and can reduce the updating intensity in the later stage to prevent the data from deviating from the correct value and realize voice recognitionThe accuracy is higher.
Description of the drawings:
FIG. 1 is a running diagram of a speech recognition method based on an FPGA platform
FIG. 2 is a flowchart of an algorithm for updating learning rate of weight threshold by two-level PID
FIG. 3 is a diagram of a BP neural network structure
FIG. 4 is a flow chart of neural network training
FIG. 5 is a simulation display diagram of four music characteristic signals
FIG. 6 is a graph of convergence weight threshold for training of general neural network structure
FIG. 7 is a graph of convergence of training weight threshold of neural network structure after double-layer PID optimization
FIG. 8 is a graph of the convergence of learning rate in training of general neural network structure
FIG. 9 is a neural network structure training learning rate convergence diagram after double-layer PID optimization
FIG. 10 is a graph of the accuracy of the training structure of the general neural network structure for four music characteristic signals
FIG. 11 is a diagram of the accuracy of a neural network structure training structure under the double-layer PID result optimization of four music characteristic signals.
The specific implementation mode is as follows:
the invention is described in further detail below with reference to the following figures and examples:
a BP neural network voice recognition method with double-layer PID optimization,
firstly, three layers of BP neural networks are built, namely an input layer, a hidden layer and an output layer. Randomly generating a weighting factor W between an output layer and a hidden layer ij (k) And W jg (k) And an activation function parameter a between two layers j (k) And b g (k) Selecting a learning rate eta (k), and setting k to 1;
secondly, the FPGA platform extracts voice data of the voice to be recognized, and the BP neural network is used for extracting the extracted data X i (k) Analyzing, and calculating output O of BP neural network output layer g (k) While outputting Y as desired g (k) Calculating error E g (k) Using error E g (k) Proportional parameter K in Proportion Integration Differentiation (PID) algorithm P Integral parameter K I Differential parameter K D Adjusting η (k);
then, using the error E g (k) And the adjusted weighting coefficient W of the output layer and the hidden layer of the eta (k) correction BP neural network ij (k) And W jg (k) And an activation function parameter a between two layers j (k) And b g (k) Until the output error of the BP neural network output layer meets the requirement, finally, judging whether the input voice is a set voice signal;
the method comprises the following steps:
a: initializing an FPGA platform based on a double-layer PID optimization BP neural network, and taking three input layer neurons, six hidden layer neurons and three output layer neurons as shown in FIG. 3. Denote the ith input layer neuron input data as X i (k) (ii) a Recording the weight value between the ith input layer neuron and the jth hidden layer neuron as W ij (k) (ii) a Let the hidden layer threshold for the jth hidden layer neuron be a j (k) (ii) a Recording the weight value between the jth hidden layer neuron and the g output layer neuron as W jg (k) (ii) a Denote the g-th output layer neuron threshold as b g (k) (ii) a Let the g output layer neuron output value be O g (k) The expected output value and the output value error value are recorded as Y g (k) And E g (k) (ii) a The learning rate is noted as η (k). Wherein i is the number of neurons in the input layer, j is the number of neurons in the hidden layer, g is the number of neurons in the output layer, and k is the current iteration number. Generating proportional parameter K in inner PID and outer PID algorithm simultaneously P Integral parameter K I Differential parameter K D 。
B: the FPGA platform extracts voice data from a voice signal by AD conversion, then performs feature extraction on the extracted voice data by the MFCC method to obtain a feature vector set, and records an input layer vector as z (k) ═ X 1 (k),X 2 (k),X 3 (k) Wherein X) is i For each of the three neurons in the input layer (i ═ 1,2,3), the input layer first iteration takes as input Z (1) ═ X 1 (1),X 2 (1),X 3 (1) Of the first one of them)The first input of a neuron is X 1 (1) After Z (k) is input, firstly, data weight processing is carried out on the hidden layer, and the processing result is recorded as:
variable N 1 For inputting the number of layers, take N 1 =3;
Continue to process the weight value processing result net j Processing hidden layer threshold value data, recorded as H j =f(net j -a j ) Wherein f is an activation function, a j To hide the layer threshold, H j Is the output layer input value.
Output layer input value H j (j is 1,2,3,4,5,6), data weighting processing is performed on the output layer, and the processing result is expressed as:
variable N 2 Taking N to hide the number of layers 2 =6;
Then, the threshold processing is continuously carried out on the output layer weight processing result and is recorded as O g =G(net g -b g ) G is an activation function, b g As output layer threshold, O g And outputting the output value of the output layer, namely the final output of the neural network. Error recording E g (k)=Y g (k)-O g (k) Wherein the first output neuron outputs an error E 1 (1)=Y 1 (1)-O 1 (1) Second output neuron output error E 2 (1)=Y 2 (1)-O 2 (1) Third output neuron output error E 3 (1)=Y 3 (1)-O 3 (1)。
C: carrying out reverse error propagation by the BP neural network, and updating a learning rate eta (k); hidden layer weight W ij (k) And a threshold value a j (k) (ii) a Output layer weight W jg (k) And a threshold value b g (k) In that respect The inner-layer PID updating formula of the learning rate eta (k) is as follows:
where s is 1,2, …, k, k is the current iteration number, N 3 Taking N as the number of output layers 3 =3;
WhereinThe purpose of summation and averaging in the formula is to convert a plurality of adjustment values into one, so that eta can be subjected to the change of the whole neural network, the parameter is used as a proportional adjustment parameter, eta can be rapidly adjusted, and the second iteration of the first output neuron is taken as an example: the parameter is used as an integral adjustment parameter, error accumulation is carried out on time, a stable adjustment is maintained, inertia during parameter updating is reduced, and system oscillation is reduced. Take the second iteration of the first output neuron as an example: e 1 (2)+E 1 (1) (ii) a Wherein E gD (k)=E g (k)-E g (k-1), the parameter is used as a differential regulation parameter, the output error change of the neural network can be reflected in advance, and an effective early correction signal is introduced into the system, so that the action speed of the system is accelerated, and the regulation time is shortened. Take the second iteration of the first output neuron as an example: e 1D (2)=E 1 (2)-E 1 (1)。
Take the inner-layer PID update formula of the second iteration of the first output neuron as an example:
the learning rate updated by using the inner PID updating algorithm can provide a larger updating intensity at the early stage of the operation of the neural network, so that the system can quickly iterate and correct parameters, the updating intensity is reduced at the later stage of the operation of the system, and the result is prevented from deviating from the correct value due to data oscillation;
after the learning rate is updated, the updated learning rate eta (k) and the output error E are used g (k) As a parameter in an inner PID updating formula, then updating a weight between an input layer and a hidden layer, a hidden layer threshold, a weight between the hidden layer and an output layer threshold;
the updating formula of the weight between the input layer and the hidden layer is as follows:
the formula for updating the weight between the input layer and the hidden layer is exemplified by the second iteration:
the updating formula of the hidden layer threshold value is as follows:
N 3 taking N as the number of output layers 3 =3。
The hidden layer threshold formula is exemplified by the second iteration:
the updating formula of the weight between the hidden layer and the output layer is as follows:
the weight formula between the hidden layer and the output layer is exemplified by the second iteration:
the updating formula of the output layer threshold value is as follows:
the update formula of the output layer threshold takes the second iteration as an example:
the negative gradient algorithm only considers the current state of the neural network, but does not consider the past state and the future state, and the weight threshold value adjustment data in the iteration process has large fluctuation and is easy to deviate from correct output. And updating the weight threshold value by using an inner-layer PID algorithm, and then updating the weight threshold value according to the error return value, so that the weight threshold value can be stably iterated, and the purposes of stronger system stability and higher output result accuracy are achieved.
D: and testing the adjusted neural network algorithm, extracting the characteristics of the test sample according to the step two, then testing according to the step two, and if the error is lower than a certain threshold value, finishing the training and outputting the result.
According to the simulation experiment result, the same data is processed by using a common neural network and a neural network with a double-layer PID structure optimized, four types of characteristic signals are shown in FIG. 5, and it can be seen that the second type of music characteristic signal is very similar to the third type of music characteristic signal, so that the second type of music characteristic signal or the third type of music characteristic signal may not be ideal in final identification accuracy.
The weight threshold of the hidden layer and the weight threshold of the output layer of the common neural network are shown in fig. 6, the convergence of the neural network with a double-layer PID structure is shown in fig. 7, and it can be seen that the output convergence of the weight threshold optimized by using the double-layer PID is more stable.
The learning rate after the training of the common neural network structure is shown in fig. 8, and the learning rate of the neural network training under the optimization of the double-layer PID structure is shown in fig. 9, so that the obvious convergence can be seen, and compared with the fixed learning rate, the learning rate is obviously more scientific by continuously converging according to the feedback result.
The operation result of the neural network with the common structure is shown in fig. 10, and the output result of the neural network with the optimized double-layer PID structure is shown in fig. 11, so that the accuracy of the neural network with the optimized double-layer PID structure for the third music identification is far higher than that of the neural network with the common structure, and meanwhile, the neural network with the optimized double-layer PID structure has almost no influence on the other three types of music identification, and the overall accuracy is greatly improved.
Claims (3)
1. A BP neural network speech recognition method of double-deck PID optimization, characterized by that:
firstly, building three layers of BP neural networks which are respectively an input layer, a hidden layer and an output layer, and randomly generating a weighting coefficient W between the output layer and the hidden layer ij (k) And W jg (k) And an activation function parameter a between two layers j (k) And b g (k) Selecting a learning rate eta (k), and setting k to be 1;
secondly, the FPGA platform extracts voice data of the voice to be recognized, and the BP neural network is used for extracting the extracted data X i (k) Analyzing and calculating the output O of the BP neural network output layer g (k) While outputting Y as desired g (k) Calculating error E g (k) Using error E g (k) With a proportional parameter K P Integral parameter K I Differential parameter K D Carrying out PID regulation on eta (k);
after the learning rate is updated, the updated learning rate eta (k) and the output error E are used g (k) As a parameter in an inner PID updating formula, then updating a weight between an input layer and a hidden layer, a hidden layer threshold, a weight between the hidden layer and an output layer threshold until an output error of a BP neural network output layer meets requirements, and finally judging whether the input voice is a set voice signal;
the method comprises the following steps:
A. carrying out training initialization preparation, initializing a neural network structure, and acquiring a sample set for training;
B. carrying out feature extraction on the sample set to obtain a feature set;
C. training a neural network by taking the feature set as a training set, and using the expected output Y in the training process g (k) And the actual output O g (k) Obtain an error E g (k) Error E g (k) Adjusting the learning rate η (k) as a parameter in the outer PID algorithm, and then continuing to use the error E g (k) And the adjusted learning rate eta (k +1) is taken as the weight W of the inner PID algorithm parameter to the input layer ij (k) Threshold value a j (k) And the output layer weight W jg (k) Threshold b g (k) Carrying out adjustment;
D. testing the adjusted neural network, extracting the characteristics of the test sample according to the step (2), and identifying the test sample by the neural network in the step (3) to obtain the error E in the step (3) g (k) When error E g (k) If the value is less than a certain threshold value, the neural network training is finished, and whether the input voice signal is a set voice signal or not is judged.
2. The method according to claim 1, wherein in the step C, the error E is determined by the method g (k) And adjusting the learning rate as an outer layer PID algorithm parameter, wherein an outer layer PID adjusting formula is as follows:
where s is 1,2, …, k, k is the current iteration number, N 3 Taking N as the number of output layers 3 3, wherein
3. The method according to claim 1, wherein in step C, an error E is used g (k) And the adjusted learning rate eta (k +1) is taken as the weight W of the inner PID algorithm parameter pair between the input layer and the hidden layer ij (k) Hidden layer threshold a j (k) And the weight W between the hidden layer and the output layer jg (k) Output layer threshold b g (k) The regulation is carried out, the outer layer PID regulation formula is as follows,
the updating formula of the weight between the input layer and the hidden layer is as follows:
the hidden layer threshold is formulated as:
N 3 taking N as the number of output layers 3 =3,
The updating formula of the weight between the hidden layer and the output layer is as follows:
the updating formula of the output layer threshold value is as follows:
where s is 1,2, …, k, k is the current iteration number, N 3 Taking N as the number of output layers 3 3, wherein E gD (k)=E g (k)-E g (k-1)。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011455918.7A CN112542161B (en) | 2020-12-10 | 2020-12-10 | BP neural network voice recognition method based on double-layer PID optimization |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011455918.7A CN112542161B (en) | 2020-12-10 | 2020-12-10 | BP neural network voice recognition method based on double-layer PID optimization |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112542161A CN112542161A (en) | 2021-03-23 |
CN112542161B true CN112542161B (en) | 2022-08-12 |
Family
ID=75018429
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011455918.7A Expired - Fee Related CN112542161B (en) | 2020-12-10 | 2020-12-10 | BP neural network voice recognition method based on double-layer PID optimization |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112542161B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113267993A (en) * | 2021-04-22 | 2021-08-17 | 上海大学 | Network training method and device based on collaborative learning |
CN113411456B (en) * | 2021-06-29 | 2023-05-02 | 中国人民解放军63892部队 | Voice quality assessment method and device based on voice recognition |
CN117012184A (en) * | 2022-10-31 | 2023-11-07 | 腾讯科技(深圳)有限公司 | Voice recognition method and related device |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101667012A (en) * | 2008-09-03 | 2010-03-10 | 长春工程学院 | Method for controlling reinforcement learning adaptive proportion integration differentiation-based distribution static synchronous compensator |
US9053431B1 (en) * | 2010-10-26 | 2015-06-09 | Michael Lamport Commons | Intelligent control with hierarchical stacked neural networks |
CN108073985A (en) * | 2016-11-14 | 2018-05-25 | 张素菁 | A kind of importing ultra-deep study method for voice recognition of artificial intelligence |
CN108445742A (en) * | 2018-02-07 | 2018-08-24 | 广东工业大学 | A kind of intelligent PID control method of gas suspension platform |
CN109034390A (en) * | 2018-08-07 | 2018-12-18 | 河北工业大学 | Phase angular amplitude PID adaptive approach based on BP neural network Three-Dimensional Magnetic feature measurement |
CN109991842A (en) * | 2019-03-14 | 2019-07-09 | 合肥工业大学 | Piano tone tuning method and system neural network based |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109492883A (en) * | 2018-10-18 | 2019-03-19 | 山东工业职业学院 | A kind of efficiency artificial intelligence on-line analysis system and method |
-
2020
- 2020-12-10 CN CN202011455918.7A patent/CN112542161B/en not_active Expired - Fee Related
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101667012A (en) * | 2008-09-03 | 2010-03-10 | 长春工程学院 | Method for controlling reinforcement learning adaptive proportion integration differentiation-based distribution static synchronous compensator |
US9053431B1 (en) * | 2010-10-26 | 2015-06-09 | Michael Lamport Commons | Intelligent control with hierarchical stacked neural networks |
CN108073985A (en) * | 2016-11-14 | 2018-05-25 | 张素菁 | A kind of importing ultra-deep study method for voice recognition of artificial intelligence |
CN108445742A (en) * | 2018-02-07 | 2018-08-24 | 广东工业大学 | A kind of intelligent PID control method of gas suspension platform |
CN109034390A (en) * | 2018-08-07 | 2018-12-18 | 河北工业大学 | Phase angular amplitude PID adaptive approach based on BP neural network Three-Dimensional Magnetic feature measurement |
CN109991842A (en) * | 2019-03-14 | 2019-07-09 | 合肥工业大学 | Piano tone tuning method and system neural network based |
Non-Patent Citations (2)
Title |
---|
Research on the Application of PID Control with Neural Network and Parameter Adjustment Method of PID Controller;Jiayu Liu,等;《Association for Computing Machinery》;20181231;第72-76页 * |
多变量自适应PID型神经网络控制器及其设计方法;丛爽,等;《信息与控制》;20061031;第35卷(第5期);第568-573页 * |
Also Published As
Publication number | Publication date |
---|---|
CN112542161A (en) | 2021-03-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112542161B (en) | BP neural network voice recognition method based on double-layer PID optimization | |
US20180046915A1 (en) | Compression of deep neural networks with proper use of mask | |
CN109671423B (en) | Non-parallel text-to-speech conversion method under limited training data | |
CN107729999A (en) | Consider the deep neural network compression method of matrix correlation | |
CN111477247B (en) | Speech countermeasure sample generation method based on GAN | |
CN107689224A (en) | The deep neural network compression method of reasonable employment mask | |
CN109740481B (en) | CNN and LSTM combined atrial fibrillation signal classification system based on jump connection | |
CN112766399A (en) | Self-adaptive neural network training method for image recognition | |
CN112330487A (en) | Photovoltaic power generation short-term power prediction method | |
CN107273971B (en) | Feed-forward neural network structure self-organization method based on neuron significance | |
CN113177355A (en) | Power load prediction method | |
CN108734116B (en) | Face recognition method based on variable speed learning deep self-coding network | |
CN110858484A (en) | Voice recognition method based on voiceprint recognition technology | |
CN108446506B (en) | Uncertain system modeling method based on interval feedback neural network | |
CN111444787B (en) | Fully intelligent facial expression recognition method and system with gender constraint | |
CN112069876A (en) | Handwriting recognition method based on adaptive differential gradient optimization | |
CN109960307B (en) | MPPT active disturbance rejection control method for photovoltaic off-grid inverter | |
CN113807005B (en) | Bearing residual life prediction method based on improved FPA-DBN | |
Seman et al. | The optimization of artificial neural networks connection weights using genetic algorithms for isolated spoken Malay parliamentary speeches | |
JP3039408B2 (en) | Sound classification method | |
Gouvea et al. | Diversity-based model reference for genetic algorithms in dynamic environment | |
CN114596567A (en) | Handwritten digit recognition method based on dynamic feedforward neural network structure and growth rate function | |
Toprak et al. | Searching Optimal Values of Identification and Controller Design Horizon Lengths, and Regularization Parameters in NARMA Based Online Learning Controller Design | |
Attya et al. | PID Controller Design and Simulation for Aircraft Roll Control Based on Evolutionary Technique Using MATLAB | |
Ghule | Implementation of Optimal Hidden Neuronsusing a fuzzy Controller |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20220812 |