TWI810988B - Method of enhancing classification of electroencephalography signals by time-frequency domain channel weighted technique and system thereof - Google Patents
Method of enhancing classification of electroencephalography signals by time-frequency domain channel weighted technique and system thereof Download PDFInfo
- Publication number
- TWI810988B TWI810988B TW111123718A TW111123718A TWI810988B TW I810988 B TWI810988 B TW I810988B TW 111123718 A TW111123718 A TW 111123718A TW 111123718 A TW111123718 A TW 111123718A TW I810988 B TWI810988 B TW I810988B
- Authority
- TW
- Taiwan
- Prior art keywords
- time
- frequency
- motor imagery
- imagery
- weighted
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 26
- 230000002708 enhancing effect Effects 0.000 title claims abstract description 9
- 238000000537 electroencephalography Methods 0.000 title abstract description 15
- 238000012360 testing method Methods 0.000 claims abstract description 78
- 238000006243 chemical reaction Methods 0.000 claims abstract description 32
- 238000012549 training Methods 0.000 claims abstract description 18
- 238000010586 diagram Methods 0.000 claims description 53
- 210000004556 brain Anatomy 0.000 claims description 43
- 238000004364 calculation method Methods 0.000 claims description 22
- 238000011176 pooling Methods 0.000 claims description 16
- 238000010832 independent-sample T-test Methods 0.000 claims description 4
- 238000005728 strengthening Methods 0.000 claims description 4
- 238000010998 test method Methods 0.000 claims 2
- 230000000875 corresponding effect Effects 0.000 abstract description 13
- 238000005516 engineering process Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 230000009466 transformation Effects 0.000 description 3
- 238000007796 conventional method Methods 0.000 description 2
- 230000033764 rhythmic process Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007177 brain activity Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
Images
Landscapes
- Control Of Electric Motors In General (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
- Networks Using Active Elements (AREA)
Abstract
Description
本發明是關於一種腦電波訊號分類之方法及其系統,特別是關於一種時間頻域通道加權強化腦電波訊號分類之方法及其系統。The present invention relates to a method and system for classifying brain wave signals, in particular to a method and system for classifying brain wave signals with time-frequency domain channel weighting enhancement.
生理訊號會因個體差異而產生出不同的結果,再加上腦部活動的不穩定性和訊號的低雜訊比,會導致運動想像分類的效能受限。而在習知的運動想像相關研究中,大多數的方法並沒有充分考慮到腦電波訊號(Electroencephalography Signal;EEG Signal)在頻率、時間及空間域的特徵訊息,這些模型的結構未能有效提取具有判別力的特徵,進而導致分類的效能受限。由此可知,目前市場上缺乏一種充分考慮時頻域、能有效提取具有判別力之特徵且提高分類效能的腦電波訊號分類之方法及其系統,故相關業者均在尋求其解決之道。Physiological signals will produce different results due to individual differences. In addition, the instability of brain activity and the low signal-to-noise ratio of the signal will limit the performance of motor imagery classification. In the conventional studies on motor imagery, most of the methods did not fully consider the characteristic information of the electroencephalography signal (EEG Signal) in the frequency, time and space domains, and the structure of these models could not effectively extract the Discriminative features, which in turn lead to limited performance in classification. It can be seen that there is currently a lack of a brainwave signal classification method and system in the market that fully considers the time-frequency domain, can effectively extract discriminative features, and improve classification performance. Therefore, relevant companies are looking for solutions.
因此,本發明之目的在於提供一種時間頻域通道加權強化腦電波訊號分類之方法及其系統,其透過分析不同運動想像任務的腦電波訊號,從中擷取出關鍵的特徵,並結合時間頻域通道加權,可以有效分辨相應動作的腦電波形式,且可避免習知技術無法有效提取具有判別力之特徵而導致分類效能受限的問題。Therefore, the object of the present invention is to provide a method and system for classifying brainwave signals with time-frequency domain channel weighting enhancement. By analyzing brainwave signals of different motor imagery tasks, key features are extracted from them and combined with time-frequency domain channels. Weighting can effectively distinguish the brain wave form of the corresponding action, and can avoid the problem that the conventional technology cannot effectively extract discriminative features, resulting in limited classification performance.
依據本發明的方法態樣之一實施方式提供一種時間頻域通道加權強化腦電波訊號分類之方法,其包含訓練步驟與測試步驟。訓練步驟包含卷積運算步驟、通道權重產生步驟、第一時頻轉換步驟、時頻權重產生步驟及加權時頻產生步驟。卷積運算步驟包含驅動處理器對第一腦電波訊號進行卷積運算,並提取第一腦電波訊號之複數第一特徵訊息。通道權重產生步驟包含驅動處理器對此些第一特徵訊息進行池化運算,並學習此些第一特徵訊息之重要程度而產生對應此些第一特徵訊息之複數通道權重係數。第一時頻轉換步驟包含驅動處理器對第一腦電波訊號進行連續小波轉換而產生複數第一時頻圖,此些第一時頻圖包含第一運動想像平均特徵時頻圖與第二運動想像平均特徵時頻圖。時頻權重產生步驟包含驅動處理器依據T檢定計算第一運動想像平均特徵時頻圖與第二運動想像平均特徵時頻圖之間的差異而產生時頻權重。加權時頻產生步驟包含驅動處理器將第一運動想像平均特徵時頻圖與時頻權重相乘而產生第一運動想像加權特徵時頻圖,並將第二運動想像平均特徵時頻圖與時頻權重相乘而產生第二運動想像加權特徵時頻圖。再者,測試步驟包含通道權重計算步驟、第二時頻轉換步驟、時頻權重計算步驟及測試分類步驟。通道權重計算步驟包含驅動處理器計算第二腦電波訊號及此些通道權重係數而產生通道權重訊號。第二時頻轉換步驟包含驅動處理器對通道權重訊號進行連續小波轉換而產生第二時頻圖。時頻權重計算步驟包含驅動處理器將第二時頻圖與時頻權重相乘而產生加權測試時頻圖。測試分類步驟包含驅動處理器計算加權測試時頻圖、第一運動想像加權特徵時頻圖及第二運動想像加權特徵時頻圖而產生分類結果。One embodiment of the method aspect according to the present invention provides a method for time-frequency domain channel weighting enhancement of brain wave signal classification, which includes a training step and a testing step. The training step includes a convolution operation step, a channel weight generation step, a first time-frequency conversion step, a time-frequency weight generation step and a weighted time-frequency generation step. The convolution operation step includes driving the processor to perform convolution operation on the first electroencephalogram signal, and extracting a plurality of first characteristic information of the first electroencephalogram signal. The channel weight generating step includes driving the processor to perform pooling operation on the first feature information, and learning the importance of the first feature information to generate complex channel weight coefficients corresponding to the first feature information. The first time-frequency conversion step includes driving the processor to perform continuous wavelet conversion on the first brain wave signal to generate complex first time-frequency images, these first time-frequency images include the first motion imagery average feature time-frequency image and the second motion Imagine an average feature time-frequency plot. The time-frequency weight generating step includes driving the processor to calculate the difference between the first motor imagery average characteristic time-frequency diagram and the second motor imagery average characteristic time-frequency diagram according to the T test to generate the time-frequency weight. The weighted time-frequency generation step includes driving the processor to multiply the first motor imagery average feature time-frequency graph with the time-frequency weight to generate the first motor imagery weighted feature time-frequency graph, and multiply the second motor imagery average feature time-frequency graph with the time-frequency weight Frequency weights are multiplied to generate a second motor imagery weighted feature time-frequency map. Furthermore, the testing step includes a channel weight calculation step, a second time-frequency conversion step, a time-frequency weight calculation step and a test classification step. The channel weight calculation step includes driving the processor to calculate the second electroencephalogram signal and the channel weight coefficients to generate channel weight signals. The second time-frequency conversion step includes driving the processor to perform continuous wavelet conversion on the channel weight signal to generate a second time-frequency diagram. The time-frequency weight calculation step includes driving the processor to multiply the second time-frequency diagram by the time-frequency weight to generate a weighted test time-frequency diagram. The test classification step includes driving the processor to calculate the weighted test time-frequency map, the first motor imagery weighted feature time-frequency map and the second motor imagery weighted feature time-frequency map to generate a classification result.
藉此,本發明的時間頻域通道加權強化腦電波訊號分類之方法透過分析不同運動想像任務的腦電波訊號,從中擷取出關鍵的特徵,並結合時間頻域通道加權,可以有效分辨相應動作的腦電波形式,且可避免習知技術無法有效提取具有判別力之特徵而導致分類效能受限的問題。In this way, the method of time-frequency domain channel weighting of the present invention to strengthen the classification of brain wave signals can effectively distinguish the corresponding action by analyzing the brain wave signals of different motor imagery tasks, extracting key features, and combining time-frequency domain channel weighting. Brain wave form, and can avoid the problem of limited classification performance caused by the inability of conventional techniques to effectively extract discriminative features.
依據本發明的結構態樣之一實施方式提供一種時間頻域通道加權強化腦電波訊號分類之系統,其包含儲存單元與處理器。儲存單元用以存取第一腦電波訊號與第二腦電波訊號。處理器連接儲存單元並接收第一腦電波訊號與第二腦電波訊號,處理器經配置以實施包含以下步驟之操作:訓練步驟與測試步驟。訓練步驟包含卷積運算步驟、通道權重產生步驟、第一時頻轉換步驟、時頻權重產生步驟及加權時頻產生步驟。卷積運算步驟包含對第一腦電波訊號進行卷積運算,並提取第一腦電波訊號之複數第一特徵訊息。通道權重產生步驟包含對此些第一特徵訊息進行池化運算,並學習此些第一特徵訊息之重要程度而產生對應此些第一特徵訊息之複數通道權重係數。第一時頻轉換步驟包含對第一腦電波訊號進行連續小波轉換而產生複數第一時頻圖,此些第一時頻圖包含第一運動想像平均特徵時頻圖與第二運動想像平均特徵時頻圖。時頻權重產生步驟包含依據T檢定計算第一運動想像平均特徵時頻圖與第二運動想像平均特徵時頻圖之間的差異而產生時頻權重。加權時頻產生步驟包含將第一運動想像平均特徵時頻圖與時頻權重相乘而產生第一運動想像加權特徵時頻圖,並將第二運動想像平均特徵時頻圖與時頻權重相乘而產生第二運動想像加權特徵時頻圖。此外,測試步驟包含通道權重計算步驟、第二時頻轉換步驟、時頻權重計算步驟及測試分類步驟。通道權重計算步驟包含計算第二腦電波訊號及此些通道權重係數而產生通道權重訊號。第二時頻轉換步驟包含對通道權重訊號進行連續小波轉換而產生第二時頻圖。時頻權重計算步驟包含將第二時頻圖與時頻權重相乘而產生加權測試時頻圖。測試分類步驟包含計算加權測試時頻圖、第一運動想像加權特徵時頻圖及第二運動想像加權特徵時頻圖而產生分類結果。One embodiment of the structural aspect according to the present invention provides a system for time-frequency domain channel weighting enhanced brain wave signal classification, which includes a storage unit and a processor. The storage unit is used for accessing the first electroencephalogram signal and the second electroencephalogram signal. The processor is connected to the storage unit and receives the first brain wave signal and the second brain wave signal, and the processor is configured to implement the operation including the following steps: a training step and a testing step. The training step includes a convolution operation step, a channel weight generation step, a first time-frequency conversion step, a time-frequency weight generation step and a weighted time-frequency generation step. The convolution operation step includes performing convolution operation on the first electroencephalogram signal, and extracting plural first characteristic information of the first electroencephalogram signal. The channel weight generating step includes performing a pooling operation on the first feature information, and learning the importance of the first feature information to generate complex channel weight coefficients corresponding to the first feature information. The first time-frequency conversion step includes performing continuous wavelet conversion on the first brain wave signal to generate a plurality of first time-frequency images, these first time-frequency images include a first motor imagery average feature time-frequency map and a second motor imagery average feature Time-Frequency Diagram. The time-frequency weight generating step includes calculating the difference between the first motor imagery average characteristic time-frequency diagram and the second motor imagery average characteristic time-frequency diagram according to the T-test to generate the time-frequency weight. The weighted time-frequency generating step comprises multiplying the first motion imagery average feature time-frequency map with the time-frequency weight to generate the first motion imagery weighted feature time-frequency map, and combining the second motion imagery average feature time-frequency map with the time-frequency weight Multiplied to generate a second motor imagery weighted feature time-frequency map. In addition, the testing step includes a channel weight calculation step, a second time-frequency conversion step, a time-frequency weight calculation step and a test classification step. The channel weight calculation step includes calculating the second electroencephalogram signal and the channel weight coefficients to generate channel weight signals. The second time-frequency conversion step includes performing continuous wavelet conversion on the channel weight signal to generate a second time-frequency diagram. The time-frequency weight calculation step includes multiplying the second time-frequency map by the time-frequency weight to generate a weighted test time-frequency map. The test classification step includes calculating the weighted test time-frequency map, the first motor imagery weighted feature time-frequency map and the second motor imagery weighted feature time-frequency map to generate a classification result.
藉此,本發明的時間頻域通道加權強化腦電波訊號分類之系統透過分析不同運動想像任務的腦電波訊號,從中擷取出關鍵的特徵,並結合時間頻域通道加權,可以有效分辨相應動作的腦電波形式,且可避免習知技術無法有效提取具有判別力之特徵而導致分類效能受限的問題。In this way, the time-frequency domain channel weighting system of the present invention enhances brain wave signal classification by analyzing the brain wave signals of different motor imagery tasks, extracting key features from them, and combining time-frequency domain channel weighting, it can effectively distinguish the corresponding action Brain wave form, and can avoid the problem of limited classification performance caused by the inability of conventional techniques to effectively extract discriminative features.
以下將參照圖式說明本發明之複數個實施例。為明確說明起見,許多實務上的細節將在以下敘述中一併說明。然而,應瞭解到,這些實務上的細節不應用以限制本發明。也就是說,在本發明部分實施例中,這些實務上的細節是非必要的。此外,為簡化圖式起見,一些習知慣用的結構與元件在圖式中將以簡單示意的方式繪示之;並且重複之元件將可能使用相同的編號表示之。Several embodiments of the present invention will be described below with reference to the drawings. For the sake of clarity, many practical details are included in the following narrative. It should be understood, however, that these practical details should not be used to limit the invention. That is, in some embodiments of the present invention, these practical details are unnecessary. In addition, for the sake of simplifying the drawings, some commonly used structures and elements will be shown in a simple and schematic way in the drawings; and repeated elements may be denoted by the same reference numerals.
此外,本文中當某一元件(或單元或模組等)「連接」於另一元件,可指所述元件是直接連接於另一元件,亦可指某一元件是間接連接於另一元件,意即,有其他元件介於所述元件及另一元件之間。而當有明示某一元件是「直接連接」於另一元件時,才表示沒有其他元件介於所述元件及另一元件之間。而第一、第二、第三等用語只是用來描述不同元件,而對元件本身並無限制,因此,第一元件亦可改稱為第二元件。且本文中之元件/單元/電路之組合非此領域中之一般周知、常規或習知之組合,不能以元件/單元/電路本身是否為習知,來判定其組合關係是否容易被技術領域中之通常知識者輕易完成。In addition, when a certain element (or unit or module, etc.) is "connected" to another element herein, it may mean that the element is directly connected to another element, or it may mean that a certain element is indirectly connected to another element , that is, there are other elements interposed between the element and another element. And when it is stated that an element is "directly connected" to another element, it means that there is no other element interposed between the element and another element. The terms first, second, third, etc. are used to describe different components, and have no limitation on the components themselves. Therefore, the first component can also be called the second component. Moreover, the combination of components/units/circuits in this article is not a combination that is generally known, conventional or conventional in this field. Whether the components/units/circuits themselves are known or not can be used to determine whether the combination relationship is easily recognized by those in the technical field. Usually knowledgeable people do it easily.
請參閱第1圖,第1圖係繪示本發明之第一實施例之時間頻域通道加權強化腦電波訊號分類之方法100的流程示意圖。時間頻域通道加權強化腦電波訊號分類之方法100包含訓練步驟S0與測試步驟S2。訓練步驟S0用以訓練第一腦電波訊號110,測試步驟S2用以測試並分類第二腦電波訊號120。Please refer to FIG. 1 . FIG. 1 is a schematic flowchart of a
訓練步驟S0包含卷積運算步驟S01、通道權重產生步驟S02、第一時頻轉換步驟S03、時頻權重產生步驟S04及加權時頻產生步驟S05。The training step S0 includes a convolution operation step S01, a channel weight generation step S02, a first time-frequency conversion step S03, a time-frequency weight generation step S04, and a weighted time-frequency generation step S05.
卷積運算步驟S01包含驅動處理器對第一腦電波訊號110進行卷積運算,並提取第一腦電波訊號110之複數第一特徵訊息。詳細地說,在卷積運算步驟S01中,處理器利用不同尺寸的複數卷積核對第一腦電波訊號110進行卷積運算,以提取不同尺寸的此些第一特徵訊息。第一腦電波訊號110屬於訓練資料。在一實施例中,此些卷積核的尺寸可分別為1×3、1×5、1×11,但本發明不以此為限。The convolution operation step S01 includes driving the processor to perform convolution operation on the
通道權重產生步驟S02包含驅動處理器對此些第一特徵訊息進行池化運算,並學習此些第一特徵訊息之重要程度而產生對應此些第一特徵訊息之複數通道權重係數112。詳細地說,通道權重產生步驟S02係對此些第一特徵訊息進行校正,針對每個通道上的特徵圖進行池化,再透過學習不同通道特徵的重要程度,進一步關注資訊量大的通道特徵,並抑制對當前任務用處不大的通道特徵,最後獲得通道權重係數112;換言之,通道權重係數112代表各通道的重要程度。在一實施例中,池化運算可為一全域性平均池化(Global Average Pooling;GAP)運算,但本發明不以此為限。The channel weight generating step S02 includes driving the processor to perform a pooling operation on the first feature information, and learning the importance of the first feature information to generate the complex
第一時頻轉換步驟S03包含驅動處理器對第一腦電波訊號110進行一連續小波轉換而產生複數第一時頻圖114,此些第一時頻圖114包含第一運動想像平均特徵時頻圖1142與第二運動想像平均特徵時頻圖1144。詳細地說,運動想像係指在腦中想像模擬自己執行運動時的狀態。第一運動想像平均特徵時頻圖1142代表多個第一運動想像特徵時頻圖之平均,其可對應一左手運動想像任務;第二運動想像平均特徵時頻圖1144代表多個第二運動想像特徵時頻圖之平均,其可對應一右手運動想像任務。連續小波轉換可為一莫萊(Morlet)小波轉換,其透過時間和頻率的局部變換,可以有效地從暫態的第一腦電波訊號110中提取特徵訊息。第一腦電波訊號110轉換至時頻域後,可繪製成二維的時頻圖,從而觀察本次試驗在某一頻率、某段時間內的能量大小變化,亦能透過觀察多次試驗來估計有利於運動想像分析的事件相關同步腦波律動(Event-Related Synchronization;ERS)與事件相關非同步腦波律動(Event-Related Desynchronization;ERD),以及mu波與beta波的變化。連續小波轉換符合下列式子(1):
(1)。
其中,
代表母小波經由伸縮
a及平移
b係數所產生的小波基底函數
。
代表輸入的第一腦電波訊號110,
代表經連續小波轉換後的結果,
s代表不同任務的運動想像狀態,
C代表通道。
The first time-frequency conversion step S03 includes driving the processor to perform a continuous wavelet conversion on the
時頻權重產生步驟S04包含驅動處理器依據一T檢定計算第一運動想像平均特徵時頻圖1142與第二運動想像平均特徵時頻圖1144之間的一差異而產生一時頻權重116。詳細地說,T檢定可為一獨立樣本T檢定(Independent Samples T-Test)。時頻權重產生步驟S04包含驅動處理器計算第一運動想像平均特徵時頻圖1142與第二運動想像平均特徵時頻圖1144之任一者之一平均值與一標準差;及驅動處理器依據T檢定計算平均值與標準差而求得時頻權重116。The time-frequency weight generating step S04 includes driving the processor to calculate a difference between the first motor imagery average characteristic time-frequency diagram 1142 and the second motor imagery average characteristic time-frequency diagram 1144 according to a T test to generate a time-
平均值依據複數運動想像任務狀態、複數總試驗次數、複數通道及複數時間點計算求得。平均值表示為 ,此些運動想像任務狀態表示為 s,此些總試驗次數表示為 ,此些通道表示為 C,此些時間點表示為 b。平均值符合下列式子(2): (2)。 The average value was calculated based on the complex motor imagery task status, the complex total number of trials, the complex channels and the complex time points. The average value is expressed as , the state of these motor imagery tasks is denoted as s , and the total number of trials is denoted as , these channels are denoted as C , and these time points are denoted as b . The average value conforms to the following formula (2): (2).
標準差依據平均值、此些運動想像任務狀態、此些總試驗次數、此些通道及此些時間點計算求得。標準差表示為 ,且符合下列式子(3): (3)。 The standard deviation is calculated based on the mean value, the motor imagery task status, the total number of trials, the channels and the time points. The standard deviation is expressed as , and conform to the following formula (3): (3).
時頻權重116依據第一運動想像任務狀態、第二運動想像任務狀態、第一運動想像試驗總數、第二運動想像試驗總數、此些通道及此些時間點計算求得。時頻權重116表示為
,第一運動想像任務狀態表示為
s 1,第二運動想像任務狀態表示為
s 2,第一運動想像試驗總數表示為
,第二運動想像試驗總數表示為
,此些通道表示為
C,此些時間點表示為
b,時頻權重116符合下列式子(4):
(4)。
上述所求得的
代表兩組運動想像任務之間的差異程度。本發明可以透過觀察兩組資料間之不同時間點
b及不同頻率的差異狀況(例如:
局部最大值的點代表在時間點
b,左手與右手的運動想像狀態之間存在著最大差異),並將此值作為時頻的權重。
The time-
加權時頻產生步驟S05包含驅動處理器將第一運動想像平均特徵時頻圖1142與時頻權重116相乘而產生第一運動想像加權特徵時頻圖1182,並將第二運動想像平均特徵時頻圖1144與時頻權重116相乘而產生第二運動想像加權特徵時頻圖1184。第一運動想像加權特徵時頻圖1182對應左手運動想像任務,第二運動想像加權特徵時頻圖1184對應右手運動想像任務。The weighted time-frequency generation step S05 includes driving the processor to multiply the first motion imagery average feature time-
測試步驟S2包含通道權重計算步驟S21、第二時頻轉換步驟S22、時頻權重計算步驟S23及測試分類步驟S24。The test step S2 includes a channel weight calculation step S21, a second time-frequency conversion step S22, a time-frequency weight calculation step S23, and a test classification step S24.
通道權重計算步驟S21包含驅動處理器計算第二腦電波訊號120及此些通道權重係數112而產生通道權重訊號122。第二腦電波訊號120屬於測試資料,通道權重係數112來自於訓練步驟S0之通道權重產生步驟S02。The channel weight calculation step S21 includes driving the processor to calculate the
第二時頻轉換步驟S22包含驅動處理器對通道權重訊號122進行連續小波轉換而產生第二時頻圖124。連續小波轉換可同第一時頻轉換步驟S03之式子(1),其細節不再贅述。The second time-frequency conversion step S22 includes driving the processor to perform continuous wavelet conversion on the
時頻權重計算步驟S23包含驅動處理器將第二時頻圖124與時頻權重116相乘而產生加權測試時頻圖126。時頻權重116來自於訓練步驟S0之時頻權重產生步驟S04;換言之,加權測試時頻圖126為經過T統計加權的時頻圖。The time-frequency weight calculation step S23 includes driving the processor to multiply the second time-
測試分類步驟S24包含驅動處理器計算加權測試時頻圖126、第一運動想像加權特徵時頻圖1182及第二運動想像加權特徵時頻圖1184而產生分類結果128。詳細地說,在測試分類步驟S24中,處理器計算加權測試時頻圖126、第一運動想像加權特徵時頻圖1182及第二運動想像加權特徵時頻圖1184而求得二相關係數
、
,並比對此二相關係數
、
而產生分類結果128,此二相關係數
、
之任一者符合下列式子(5):
(5)。
其中,
代表各此二相關係數
、
,
代表一總試驗次數,
s代表複數運動想像任務狀態,
x代表第一運動想像加權特徵時頻圖1182與第二運動想像加權特徵時頻圖1184之一者,
y代表加權測試時頻圖126,
i代表1~
之一正整數。舉例來說,若
x代表第一運動想像加權特徵時頻圖1182(對應左手運動想像),則
代表相關係數
;若
x代表第二運動想像加權特徵時頻圖1184(對應右手運動想像),則
代表相關係數
。二相關係數
、
用以分別評估各次試驗的加權測試時頻圖126和左、右手運動想像平均時頻圖(即第一運動想像平均特徵時頻圖1142、第二運動想像平均特徵時頻圖1144)之間的相關程度。對某次試驗的加權測試時頻圖126與左手運動想像平均時頻圖(即第一運動想像平均特徵時頻圖1142)進行計算,可求得相關係數
;對某次試驗的加權測試時頻圖126與右手運動想像平均時頻圖(即第二運動想像平均特徵時頻圖1144)進行計算,可求得相關係數
。如果相關係數
大於等於相關係數
,則分類結果128代表此次試驗被預測為左手運動想像;相反地,如果相關係數
小於相關係數
,則分類結果128代表此次試驗被預測為右手運動想像。
The test classification step S24 includes driving the processor to calculate the weighted test time-
式子(5)的相關係數 表示某次試驗經T統計加權的時頻圖(即加權測試時頻圖126)與經T統計加權的左、右手運動想像平均時頻圖(即第一運動想像加權特徵時頻圖1182、第二運動想像加權特徵時頻圖1184)之間的相關程度。藉此,本發明可以透過相關係數 來觀察本次試驗與何種類型的運動想像任務之關係較為密切,並推測其分類。此外,透過分析不同運動想像任務的腦電波訊號,從中擷取出關鍵的特徵,可以有效分辨相應動作的腦電波形式。 Correlation coefficient of formula (5) Represent the time-frequency diagram weighted by T statistics (i.e. weighted test time-frequency diagram 126) and the average time-frequency diagram (i.e. the first weighted feature time-frequency diagram 1182, the first weighted characteristic time-frequency diagram 1182) of left and right hand motion imagery weighted by T statistics in a test The degree of correlation between the two motor imagery weighted feature time-frequency diagrams 1184). In this way, the present invention can use the correlation coefficient To observe which type of motor imagery task is closely related to this experiment, and speculate on its classification. In addition, by analyzing the brain wave signals of different motor imagery tasks and extracting key features from them, the brain wave forms of corresponding actions can be effectively distinguished.
請一併參閱第1圖與第2圖,其中第2圖係繪示本發明之第二實施例之時間頻域通道加權強化腦電波訊號分類之系統200的示意圖。時間頻域通道加權強化腦電波訊號分類之系統200包含儲存單元210與處理器220。儲存單元210用以存取第一腦電波訊號110與第二腦電波訊號120。處理器220連接儲存單元210並接收第一腦電波訊號110與第二腦電波訊號120,處理器220經配置以實施第1圖的時間頻域通道加權強化腦電波訊號分類之方法100,亦即處理器220經配置以執行卷積運算222、池化運算224、連續小波轉換226及T檢定228。儲存單元210可包含可儲存供處理器220執行之資訊和指令的隨機存取記憶體(Random Access Memory;RAM)或其它型式的動態儲存裝置。處理器220可包含任何型式的處理器、微處理器、或可編譯並執行指令的場效型可編程邏輯陣列(Field Programmable Gate Array;FPGA)。處理器220可包含單一裝置(例如單核心)或一組裝置(例如多核心)。藉此,本發明之時間頻域通道加權強化腦電波訊號分類之系統200透過分析不同運動想像任務的腦電波訊號,從中擷取出關鍵的特徵,並結合時間頻域通道加權,可以有效分辨相應動作的腦電波形式,且可避免習知技術無法有效提取具有判別力之特徵而導致分類效能受限的問題。Please refer to FIG. 1 and FIG. 2 together. FIG. 2 is a schematic diagram of a
由上述實施方式可知,本發明具有下列優點:其一,可對特徵訊息進行校正,針對每個通道上的特徵圖進行池化,再透過學習不同通道特徵的重要程度,進一步關注資訊量大的通道特徵,並抑制對當前任務用處不大的通道特徵,進而有效提取具有判別力之特徵。其二,可以透過相關係數來觀察各次試驗與何種類型的運動想像任務之關係較為密切,並推測其分類。其三,透過分析不同運動想像任務的腦電波訊號,從中擷取出關鍵的特徵,並結合時間頻域通道加權,可以有效分辨相應動作的腦電波形式,且可避免習知技術無法有效提取具有判別力之特徵而導致分類效能受限的問題。It can be seen from the above embodiments that the present invention has the following advantages: First, the feature information can be corrected, and the feature maps on each channel can be pooled, and then by learning the importance of different channel features, further attention can be paid to information-heavy Channel features, and suppress channel features that are not very useful for the current task, and then effectively extract discriminative features. Second, the correlation coefficient can be used to observe which types of motor imagery tasks are closely related to each test, and to speculate on its classification. Third, by analyzing the brain wave signals of different motor imagery tasks, key features can be extracted from them, combined with time-frequency domain channel weighting, the brain wave form of the corresponding action can be effectively distinguished, and the conventional technology can not effectively extract the discriminative features. The characteristics of power lead to the problem of limited classification efficiency.
雖然本發明已以實施方式揭露如上,然其並非用以限定本發明,任何熟習此技藝者,在不脫離本發明之精神和範圍內,當可作各種之更動與潤飾,因此本發明之保護範圍當視後附之申請專利範圍所界定者為準。Although the present invention has been disclosed above in terms of implementation, it is not intended to limit the present invention. Anyone skilled in this art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection of the present invention The scope shall be defined by the appended patent application scope.
100:時間頻域通道加權強化腦電波訊號分類之方法 110:第一腦電波訊號 112:通道權重係數 114:第一時頻圖 1142:第一運動想像平均特徵時頻圖 1144:第二運動想像平均特徵時頻圖 116:時頻權重 1182:第一運動想像加權特徵時頻圖 1184:第二運動想像加權特徵時頻圖 120:第二腦電波訊號 122:通道權重訊號 124:第二時頻圖 126:加權測試時頻圖 128:分類結果 :相關係數 S0:訓練步驟 S01:卷積運算步驟 S02:通道權重產生步驟 S03:第一時頻轉換步驟 S04:時頻權重產生步驟 S05:加權時頻產生步驟 S2:測試步驟 S21:通道權重計算步驟 S22:第二時頻轉換步驟 S23:時頻權重計算步驟 S24:測試分類步驟 200:時間頻域通道加權強化腦電波訊號分類之系統 210:儲存單元 220:處理器 222:卷積運算 224:池化運算 226:連續小波轉換 228:T檢定100: Time-frequency domain channel weighting method to strengthen the classification of brain wave signals 110: The first brain wave signal 112: Channel weight coefficient 114: The first time-frequency graph 1142: The first motor imagery average feature time-frequency graph 1144: The second motor imagery Average feature time-frequency diagram 116: time-frequency weight 1182: first motor imagery weighted feature time-frequency graph 1184: second motor imagery weighted feature time-frequency graph 120: second brain wave signal 122: channel weight signal 124: second time-frequency Figure 126: Weighted Test Time-Frequency Figure 128: Classification Results : correlation coefficient S0: training step S01: convolution operation step S02: channel weight generation step S03: first time-frequency conversion step S04: time-frequency weight generation step S05: weighted time-frequency generation step S2: testing step S21: channel weight calculation Step S22: Second Time-Frequency Conversion Step S23: Time-Frequency Weight Calculation Step S24: Test Classification Step 200: Time-Frequency Domain Channel Weighting Enhanced EEG Signal Classification System 210: Storage Unit 220: Processor 222: Convolution Operation 224: Pooling Operation 226: Continuous Wavelet Transformation 228: T Test
第1圖係繪示本發明之第一實施例之時間頻域通道加權強化腦電波訊號分類之方法的流程示意圖;以及 第2圖係繪示本發明之第二實施例之時間頻域通道加權強化腦電波訊號分類之系統的示意圖。 FIG. 1 is a schematic flow chart showing the method for time-frequency domain channel weighting and strengthening of brain wave signal classification according to the first embodiment of the present invention; and FIG. 2 is a schematic diagram of a system for time-frequency domain channel weighting and enhanced brain wave signal classification according to the second embodiment of the present invention.
100:時間頻域通道加權強化腦電波訊號分類之方法 100: Time-Frequency Domain Channel Weighting Enhancement Method for EEG Signal Classification
110:第一腦電波訊號 110: The first brain wave signal
112:通道權重係數 112: channel weight coefficient
114:第一時頻圖 114: The first time-frequency diagram
1142:第一運動想像平均特徵時頻圖 1142: Time-Frequency Diagram of Average Feature of First Motor Imagery
1144:第二運動想像平均特徵時頻圖 1144: Time-Frequency Diagram of Average Feature of Second Motor Imagery
116:時頻權重 116: Time-frequency weight
1182:第一運動想像加權特徵時頻圖 1182: Time-Frequency Map of the First Motion Imagery Weighted Feature
1184:第二運動想像加權特徵時頻圖 1184: Time-Frequency Diagram of Second Motor Imagery Weighted Feature
120:第二腦電波訊號 120: The second brain wave signal
122:通道權重訊號 122: Channel weight signal
124:第二時頻圖 124: The second time-frequency diagram
126:加權測試時頻圖 126: Weighted test time-frequency diagram
128:分類結果 128: Classification result
r s :相關係數 r s : correlation coefficient
S0:訓練步驟 S0: training step
S01:卷積運算步驟 S01: Convolution operation steps
S02:通道權重產生步驟 S02: Channel weight generation steps
S03:第一時頻轉換步驟 S03: the first time-frequency conversion step
S04:時頻權重產生步驟 S04: Time-frequency weight generation steps
S05:加權時頻產生步驟 S05: weighted time-frequency generation steps
S2:測試步驟 S2: Test steps
S21:通道權重計算步驟 S21: Channel weight calculation steps
S22:第二時頻轉換步驟 S22: the second time-frequency conversion step
S23:時頻權重計算步驟 S23: Time-frequency weight calculation steps
S24:測試分類步驟 S24: Test classification step
Claims (10)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW111123718A TWI810988B (en) | 2022-06-24 | 2022-06-24 | Method of enhancing classification of electroencephalography signals by time-frequency domain channel weighted technique and system thereof |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW111123718A TWI810988B (en) | 2022-06-24 | 2022-06-24 | Method of enhancing classification of electroencephalography signals by time-frequency domain channel weighted technique and system thereof |
Publications (2)
Publication Number | Publication Date |
---|---|
TWI810988B true TWI810988B (en) | 2023-08-01 |
TW202400086A TW202400086A (en) | 2024-01-01 |
Family
ID=88585562
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
TW111123718A TWI810988B (en) | 2022-06-24 | 2022-06-24 | Method of enhancing classification of electroencephalography signals by time-frequency domain channel weighted technique and system thereof |
Country Status (1)
Country | Link |
---|---|
TW (1) | TWI810988B (en) |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114492500A (en) * | 2021-12-07 | 2022-05-13 | 浙江工业大学 | Motor imagery electroencephalogram signal classification method based on one-dimensional convolution kernel |
CN114521903A (en) * | 2022-02-15 | 2022-05-24 | 南京邮电大学 | Electroencephalogram attention recognition system and method based on feature selection |
-
2022
- 2022-06-24 TW TW111123718A patent/TWI810988B/en active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114492500A (en) * | 2021-12-07 | 2022-05-13 | 浙江工业大学 | Motor imagery electroencephalogram signal classification method based on one-dimensional convolution kernel |
CN114521903A (en) * | 2022-02-15 | 2022-05-24 | 南京邮电大学 | Electroencephalogram attention recognition system and method based on feature selection |
Also Published As
Publication number | Publication date |
---|---|
TW202400086A (en) | 2024-01-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2019090879A1 (en) | Analog circuit fault diagnosis method based on cross wavelet features | |
Kirch et al. | Detection of changes in multivariate time series with application to EEG data | |
Maechler et al. | VLSI design of approximate message passing for signal restoration and compressive sensing | |
Weinberger et al. | Graph Laplacian regularization for large-scale semidefinite programming | |
US9724005B2 (en) | Real-time multi-channel EEG signal processor based on on-line recursive independent component analysis | |
US9031816B2 (en) | Independent component analysis processor | |
Zou et al. | Removing muscle artifacts from EEG data via underdetermined joint blind source separation: A simulation study | |
CN112906335B (en) | Passivity correction method and device for integrated circuit system | |
Murali et al. | An efficient adaptive filter architecture for improving the seizure detection in EEG signal | |
TWI810988B (en) | Method of enhancing classification of electroencephalography signals by time-frequency domain channel weighted technique and system thereof | |
CN112884062B (en) | Motor imagery classification method and system based on CNN classification model and generated countermeasure network | |
US11544281B2 (en) | Query-oriented approximate query processing based on machine learning techniques | |
US20140316746A1 (en) | Real-time signal processing system and method based on multi-channel independent component analysis | |
CN111508525B (en) | Full-reference audio quality evaluation method and device | |
Jiang et al. | A data-driven high-resolution time-frequency distribution | |
Du et al. | Gearbox fault diagnosis method based on improved MobileNetV3 and transfer learning | |
Ding et al. | Multivariate functional response low‐rank regression with an application to brain imaging data | |
Boudehane et al. | Parallelization scheme for canonical polyadic decomposition of large-scale high-order tensors | |
Cárdenas et al. | Ecg arrhythmia classification for comparing pre-trained deep learning models | |
Baali et al. | Efficient hardware implementation of the l 1—Regularized least squares for IoT edge computing | |
Zhang et al. | Design and application of electrocardiograph diagnosis system based on multifractal theory | |
Canbay et al. | An area efficient real time implementation of dual tree complex wavelet transform in field programmable gate arrays | |
Sundararajan | Factor modeling of multivariate time series: A frequency components approach | |
CN110197219B (en) | Hardware implementation method of Bayes classifier supporting data classification | |
Zhu et al. | Multivariate varying coefficient model and its application in neuroimaging data |