Multifunctional calling extension set
Technical field
The present invention relates to Multifunctional calling extension sets, belong to field of medical device.
Background technique
Calling extension set mainly cooperates nurse station host to use, and intercommunication communicates between host.
Since the bed of hospital is more, network bandwidth is unable to satisfy the requirement of more extension set simultaneous communications.Do not have in the prior art
It is provided with a solution and carries out screening, compression for the transmission data to calling extension set to reduce the demand to bandwidth.
Summary of the invention
For that will reduce demand of the calling extension set to bandwidth, the invention proposes a kind of Multifunctional calling extension sets.
Technical solution of the present invention is as follows:
Multifunctional calling extension set, including main control chip further include the microphone being respectively connected with main control chip, raise
Sound device and receiving module, the main control chip also pass through level cache and are connected with discrimination module, and the discrimination module passes through two
Grade caching is connected with sending module;
The discrimination module is based on fpga chip, has built in the fpga chip for differentiating invalid voice data slot
BP neural network, the BP neural network differentiates the method and step of invalid voice data slot are as follows:
(A) voice that main control chip includes microphone is converted into voice data, by 50Hz or less in the voice data and
The frequency range of 1200Hz or more all filters out, then the voice data after filtering out is divided into speech-sound data sequence as unit of 3s and is incited somebody to action
The speech-sound data sequence is stored in level cache;
Main control chip is handled as follows respectively from the element being successively read in speech-sound data sequence in level cache:
(A-1) the ensemble average decibel value for remembering the element is x1, whole code rate is x2;
(A-2) frequency-domain analysis is carried out to the element, using 50Hz as starting point, calculates every change rate for crossing 50Hz decibel value, note
Recording frequency values of first change rate greater than 0.1dB/Hz is x3, frequency values of first change rate less than -0.1dB/Hz be x4;
If not finding qualified x3, then by x3It is set as 50Hz, if not finding qualified x4, then by x4It is set as 1200Hz;
(A-3) x is calculated3To x4The average decibel value of frequency range is x5;
(A-4) by x1、x2、x3、x4And x5It stores as one group of input data into level cache;
(B) each group input data in level cache is sequentially delivered to sentence in the BP neural network of discrimination module
Not;
The BP neural network is disposed with input layer, pretreatment layer, middle layer and output layer along input to output direction;
The input layer includes for inputting x1Input unit one, for inputting x2Input unit two, for inputting x3
Input unit three, for inputting x4Input unit four and for inputting x5Input unit five;
The pretreatment layer includes pretreatment unit one, pretreatment unit two, pretreatment unit three and pretreatment unit four;
The middle layer includes temporary location one, temporary location two and temporary location three;
The output layer includes output unit;
The input layer, pretreatment layer, middle layer and output layer are respectively the 1st layer, the 2nd layer, the 3rd layer of BP neural network
With the 4th layer;
The input unit one, input unit two, input unit three, input unit four and input unit five are respectively the 1st
Unit the 1st, Unit the 2nd, Unit the 3rd, Unit the 4th and the Unit the 5th of layer;
The pretreatment unit one, pretreatment unit two, pretreatment unit three and pretreatment unit four are respectively the 2nd layer
Unit the 1st, Unit the 2nd, Unit the 3rd and Unit the 4th;
The temporary location one, temporary location two and temporary location three be respectively the 3rd layer Unit the 1st, Unit the 2nd and
Unit the 3rd;
Unit the 1st that the output unit is the 4th layer;
If the output valve of l layers of i-th cell isBias term isActivation primitive is fi (l)(), l layers
Unit sum is n(l), the output valve of l layers of jth unitWeight when being transferred to l+1 layers of i-th cell is
Then for the 1st layer:
For the 2nd to 4 layer:
IfWithPerseverance is 0;
BP neural network judges whether the element is invalid voice data slot according to the input data, if invalid language
The element is then replaced with blank sound data by sound data slot;
(C) discrimination module will be replaced processed speech-sound data sequence and is sent in L2 cache.
Further: the activation primitive of the pretreatment layer each unit are as follows:
Further: the activation primitive of the middle layer and output layer each unit are as follows:
fi (l)(x)=max (0, x-0.2).
Further, the training method of BP neural network are as follows: ambient noise frequency be 20Hz, 80Hz, 150Hz,
It is recorded respectively in 200Hz, 360Hz, 500Hz, 750Hz, 1000Hz, 1500Hz, 2000Hz and the respective environment without voice
The invalid sample voice data of 100 duration 30s, and ambient noise frequency be 20Hz, 80Hz, 150Hz, 200Hz, 360Hz,
It 500Hz, 750Hz, 1000Hz, 1500Hz, 2000Hz and respectively all has and records 100 duration 30s in the environment of voice respectively
Effective sample voice data;2000 sample voice data are respectively divided into speech-sound data sequence as unit of 3s,
20000 elements of all speech-sound data sequences are subjected to random ordering and are arranged to make up sample sequence, are successively read in sample sequence
Element: for each element, remember that the ensemble average decibel value of the element is x1, whole code rate is x2, frequency domain is carried out to the element
Analysis calculates every change rate for crossing 50Hz decibel value using 50Hz as starting point, records the frequency that first change rate is greater than 0.1dB/Hz
Rate value is x3, frequency values of first change rate less than -0.1dB/Hz be x4If not finding qualified x3, then by x3If
It is set to 50Hz, if not finding qualified x4, then by x4It is set as 1200Hz, calculates x3To x4The average decibels of frequency range are
x5, by x1、x2、x3、x4And x5As one group of training sample input data;20000 groups of training sample input datas are combined into each member
Invalidating expected outcome corresponding to plain original keeps BP neural network training when training WithPerseverance is 0.
Further: the main control chip is also connected with FM module.
Further: the main control chip is also connected with display screen.
Further: the main control chip is also connected with memory module.
Compared with the existing technology, the invention has the following advantages that (1) present invention has the discrimination module based on FPGA, energy
Whether enough determined using trained neural network algorithm according to the characteristic information of voice data segment is voice content blank
Invalid voice data slot, and invalid voice data slot is replaced with into blank sound data, has compressed transmitting audio data
Size reduces demand of the calling extension set to bandwidth;(2) this detection method differentiates voice data using neural network,
Have the advantages that None-linear approximation ability is strong, judging efficiency is high and accuracy rate is high;(3) pretreatment layer is introduced in neural network,
Since everyone voice frequency range is more concentrated, part flexible strategy are carried out in pretreatment layer to force setting, and will
First change rate is greater than the frequency values x of 0.1dB/Hz3It is less than the frequency values x of -0.1dB/Hz with first change rate4Both
Correlation is more apparent but the characteristic information that can not be completely integrated has carried out incomplete merging treatment, then again will pretreatment
The result of layer is output in middle layer, ensure that x in subsequent calculating process3And x4Always possess certain correlation, to mention
The high accuracy of judging result, while also improving trained efficiency;(4) the activation primitive setting of pretreatment layer fully considers
X3And x4Two incomplete merging treatments of characteristic information are in terms of computational efficiency, differential solve difficulty and correlation reservation
It is required that having the advantages that solution, training effectiveness are high and judgment accuracy is high.
Detailed description of the invention
Fig. 1 is the structural schematic diagram of Multifunctional calling extension set proposed by the invention.
Fig. 2 is the structural schematic diagram of BP neural network.
Specific embodiment
The technical solution that the invention will now be described in detail with reference to the accompanying drawings:
Such as Fig. 1, Multifunctional calling extension set, including main control chip further include the wheat being respectively connected with main control chip
Gram wind, loudspeaker and receiving module, the main control chip also pass through level cache and are connected with discrimination module, the discrimination module
Sending module is connected with by L2 cache;The main control chip is also connected with FM module, liquid crystal display and memory module, institute
It states memory module and can be built-in storage chip and be also possible to access the flash disk device of this calling extension set by USB interface.
Patient can also listen to FM broadcast, or appreciate storage by this calling extension set and host communication by loudspeaker
Music in module.Liquid crystal display may also display the relevant informations such as the name of patient, age.
The discrimination module is based on fpga chip, has built in the fpga chip for differentiating invalid voice data slot
BP neural network, the BP neural network differentiates the method and step of invalid voice data slot are as follows:
(A) voice that main control chip includes microphone is converted into voice data, it is contemplated that voice frequency range be 65 to
1100Hz all filters out the frequency range of 50Hz or less and 1200Hz or more in the voice data, then by the voice data after filtering out
It is divided into speech-sound data sequence as unit of 3s and the speech-sound data sequence is stored in level cache;
Main control chip is handled as follows respectively from the element being successively read in speech-sound data sequence in level cache:
(A-1) the ensemble average decibel value for remembering the element is x1, whole code rate is x2;
(A-2) frequency-domain analysis is carried out to the element, using 50Hz as starting point, calculates every change rate for crossing 50Hz decibel value, note
Recording frequency values of first change rate greater than 0.1dB/Hz is x3, frequency values of first change rate less than -0.1dB/Hz be x4;
If not finding qualified x3, then by x3It is set as 50Hz, if not finding qualified x4, then by x4It is set as 1200Hz;
(A-3) x is calculated3To x4The average decibel value of frequency range is x5;
(A-4) by x1、x2、x3、x4And x5It stores as one group of input data into level cache;
(B) each group input data in level cache is sequentially delivered to sentence in the BP neural network of discrimination module
Not;
The BP neural network is disposed with input layer, pretreatment layer, middle layer and output layer along input to output direction;
The input layer includes for inputting x1Input unit one, for inputting x2Input unit two, for inputting x3
Input unit three, for inputting x4Input unit four and for inputting x5Input unit five;
The pretreatment layer includes pretreatment unit one, pretreatment unit two, pretreatment unit three and pretreatment unit four;
The middle layer includes temporary location one, temporary location two and temporary location three;
The output layer includes output unit;
The input layer, pretreatment layer, middle layer and output layer are respectively the 1st layer, the 2nd layer, the 3rd layer of BP neural network
With the 4th layer;
The input unit one, input unit two, input unit three, input unit four and input unit five are respectively the 1st
Unit the 1st, Unit the 2nd, Unit the 3rd, Unit the 4th and the Unit the 5th of layer;
The pretreatment unit one, pretreatment unit two, pretreatment unit three and pretreatment unit four are respectively the 2nd layer
Unit the 1st, Unit the 2nd, Unit the 3rd and Unit the 4th;
The temporary location one, temporary location two and temporary location three be respectively the 3rd layer Unit the 1st, Unit the 2nd and
Unit the 3rd;
Unit the 1st that the output unit is the 4th layer;
If the output valve of l layers of i-th cell isBias term isActivation primitive is fi (l)(), l layers
Unit sum is n(l), the output valve of l layers of jth unitWeight when being transferred to l+1 layers of i-th cell is
Then for the 1st layer:
For the 2nd to 4 layer:
IfWithThe reason of perseverance is 0, this setting is: each
The voice frequency range of people is typically more concentrated, i.e. x3With x4Difference is not too big, therefore carries out in pretreatment layer to part flexible strategy
Force setting, and by x3And x4Both correlations are more apparent but the characteristic information that can not be completely integrated has carried out non-fully
The result of pretreatment layer, is then output in middle layer again, ensure that x in subsequent calculating process by the merging treatment of property3With
x4Always possess certain correlation, to improve the accuracy of judging result, while also improving trained efficiency;
BP neural network judges whether the element is invalid voice data slot according to the input data, if invalid language
The element is then replaced with blank sound data by sound data slot;
(C) discrimination module will be replaced processed speech-sound data sequence and is sent in L2 cache.
The activation primitive of the pretreatment layer each unit are as follows:
The setting of the activation primitive has fully considered x3And x4Effect is being calculated after two incomplete merging treatments of characteristic information
Rate, differential solve the requirement that difficulty and correlation retain aspect, have that solution, training effectiveness are high and that judgment accuracy is high is excellent
Point;
The activation primitive of the middle layer and output layer each unit are as follows:
fi (1)(x)=max (0, x-0.2).
The training method of the BP neural network are as follows: ambient noise frequency be 20Hz, 80Hz, 150Hz, 200Hz,
100 durations are recorded respectively in 360Hz, 500Hz, 750Hz, 1000Hz, 1500Hz, 2000Hz and the respective environment without voice
The invalid sample voice data of 30s, and ambient noise frequency be 20Hz, 80Hz, 150Hz, 200Hz, 360Hz, 500Hz,
It 750Hz, 1000Hz, 1500Hz, 2000Hz and respectively all has and records the effective of 100 duration 30s in the environment of voice respectively
Sample voice data;The personnel that when sampling, should be selected different environment as far as possible, and arrange sound ray mutually different participate in voice
Recording;2000 sample voice data are respectively divided into speech-sound data sequence as unit of 3s, by all voice numbers
Random ordering, which is carried out, according to 20000 elements of sequence is arranged to make up sample sequence, the element being successively read in sample sequence: for each
Element remembers that the ensemble average decibel value of the element is x1, whole code rate is x2, frequency-domain analysis is carried out to the element, is with 50Hz
Point calculates every change rate for crossing 50Hz decibel value, and recording frequency values of first change rate greater than 0.1dB/Hz is x3, first
Frequency values of the change rate less than -0.1dB/Hz are x4If not finding qualified x3, then by x3It is set as 50Hz, if not finding
Qualified x4, then by x4It is set as 1200Hz, calculates x3To x4The average decibels of frequency range are x5, by x1、x2、x3、x4With
x5As one group of training sample input data;By it is effective corresponding to 20000 groups of training sample input data combination each elements originals/
Invalid expected outcome keeps BP neural network training when training
WithPerseverance is 0.
There is the voice that interim sound rests, thus transmits due to health status etc., in communication often in patient
Data volume is larger, occupies substantial portion of bandwidth.In order to which the segment that voice content in voice data is blank is removed, with
Reduce the demand to bandwidth, this programme is determined in vain by using discrimination module according to the characteristic information of voice data segment
Voice data segment is led to by L2 cache again after invalid voice data slot is then replaced with the minimum blank sound data of volume
It crosses sending module and is transmitted to host, have compressed the size of transmitting audio data, reduce demand of the calling extension set to bandwidth.