技术领域technical field
本发明属于电气领域，具体涉及一种基于局部特征的PCNN电力故障区域检测方法。The invention belongs to the electrical field, and in particular relates to a PCNN power fault area detection method based on local features.
背景技术Background technique
红外热成像因具有非接触、远距离、被动检测等诸多优点，目前已经成为电力部门实施电气设备故障状态检测的重要手段。然而，红外故障检测主要依赖运维人员定期巡视，这对于工作人员而言，设备红外检测需要投入大量的时间，存在效率低、易漏检以及管理成本相对高等缺点。为此，自动红外检测受到了广大工作人员的高度重视，例如徐雪涛[1]提出改进PCNN(Pulse-coupled neural network)模型的图像分割方法提取红外图像中故障区域，但由于PCNN内部参数设置灵活，且因不同红外电气设备故障图像的差异，使得在实际应用中需要人为进行调整。以上模型在对红外故障图像进行处理时，受到参数及阈值设置等方面的影像，使得分割后的图像不能完全提取出电力故障区域，从而疏漏对故障系统的检测和维护，由此引发电气系统事故，造成电气设备损坏和人员伤亡。Infrared thermal imaging has many advantages such as non-contact, long-distance, passive detection, etc., and has become an important means for the power sector to implement fault status detection of electrical equipment. However, infrared fault detection mainly relies on regular inspections by operation and maintenance personnel. For the staff, infrared detection of equipment requires a lot of time, and has disadvantages such as low efficiency, easy missed detection, and relatively high management costs. For this reason, automatic infrared detection has been highly valued by the majority of staff. For example, Xu Xuetao[1] proposed an image segmentation method to improve the PCNN (Pulse-coupled neural network) model to extract fault areas in infrared images. However, due to the flexible setting of internal parameters of PCNN, Moreover, due to the difference in fault images of different infrared electrical equipment, manual adjustment is required in practical applications. When the above model is processing the infrared fault image, it is affected by the image of parameters and threshold settings, so that the segmented image cannot completely extract the power fault area, thus neglecting the detection and maintenance of the fault system, which leads to electrical system accidents , causing damage to electrical equipment and casualties.
因此，研究高效的电力设备故障自动红外检测方法非常必要。Therefore, it is very necessary to study efficient automatic infrared detection methods for power equipment faults.
相关参考文献如下：The relevant references are as follows:
[1]徐雪涛.基于红外成像技术的电气设备故障诊断[D].华北电力大学,2014.[1] Xu Xuetao. Fault diagnosis of electrical equipment based on infrared imaging technology [D]. North China Electric Power University, 2014.
发明内容Contents of the invention
本发明的目的是提供一种基于局部特征的PCNN电力故障区域检测方法，以自动高效的提取电力故障设备红外图像中故障区域，预防电力事故的发生。The purpose of the present invention is to provide a PCNN power fault area detection method based on local features to automatically and efficiently extract the fault area in the infrared image of the power fault equipment to prevent the occurrence of power accidents.
本发明所采用的技术方案是：一种基于局部特征的PCNN电力故障区域检测方法，包括以下步骤，The technical scheme that the present invention adopts is: a kind of PCNN power failure area detection method based on local feature, comprises the following steps,
步骤1，获得故障电气设备的红外图像，其中电力故障区域即为红外图像中亮度较高的区域；Step 1, obtain the infrared image of the faulty electrical equipment, wherein the power failure area is the area with higher brightness in the infrared image;
步骤2，将原始红外图像输入到基于局部特征的PCNN模型中，实现故障区域的提取，具体实现方式如下，Step 2, input the original infrared image into the PCNN model based on local features to realize the extraction of the fault area, the specific implementation method is as follows,
步骤2.1，将原始红外图像作为输入赋给基于局部特征的PCNN模型，并根据图像最高亮度区域获得初始化神经元脉冲发放区域Y(0),其中Y(0)＝{ij|Fij(n)＝＝Th}，Th为图像最高灰度值，按式(一)设置权重Mij,kl和Wij,kl，Step 2.1, assign the original infrared image as input to the PCNN model based on local features, and obtain the initial neuron pulse firing area Y(0) according to the highest brightness area of the image, where Y(0)={ij|Fij (n) ==Th }, Th is the highest gray value of the image, set the weight Mij,kl and Wij,kl according to formula (1),
其中，σh为高斯尺度，设置为1或更大，C为归一化系数；Among them, σh is the Gaussian scale, set to 1 or greater, and C is the normalization coefficient;
步骤2.2，基于局部特征的PCNN模型按式(二)～(八)进行迭代计算，In step 2.2, the PCNN model based on local features is iteratively calculated according to formulas (2) to (8),
Uij(n)＝Fij(n)·[1+βLij(n)] (四)Uij (n)＝Fij (n)·[1+βLij (n)] (4)
X＝{ij|Lij(n)＞0}∪{ij|Yij(n)＝0} (七)X＝{ij|Lij (n)＞0}∪{ij|Yij (n)＝0} (7)
其中，Fij(n)为反馈输入，Lij(n)为连接输入，Iij表示神经元ij所对应的图像灰度值，n为迭代索引，下标ij、kl分别代表红外灰度图像像素位置所对应的神经元，其值为对应灰度值Iij，VF和VL分别为放大系数，αF和αL为衰减系数，Uij为神经元ij的内部活动项，β为连接系数，Eg为全局阈值，Ωc为先前点火区域，当Uij(n)＞Eg(n-1)时，神经元发生点火并形成脉冲Yij(n)，即点火区域，神经元集合Sc是在集合X内与点火区域Yij(n)相似的神经元，由模糊聚类模型进行归类得到；Among them, Fij (n) is the feedback input, Lij (n) is the connection input, Iij represents the image gray value corresponding to the neuron ij, n is the iteration index, and the subscripts ij and kl represent the infrared gray image respectively The neuron corresponding to the pixel position, its value is the corresponding gray value Iij , VF and VL are the amplification coefficients, αF and αL are the attenuation coefficients, Uij is the internal activity item of neuron ij, and β is Connection coefficient, Eg is the global threshold, Ωc is the previous firing area, when Uij (n) > Eg (n-1), the neuron fires and forms a pulse Yij (n), that is, the firing area, neuron The cell set Sc is the neuron similar to the ignition area Yij (n) in the set X, and is classified by the fuzzy clustering model;
步骤2.3，设置停止规则，针对点火区域Yij(n)的边界，结合图像梯度算子，Step 2.3, set the stop rule, aiming at the boundary of the ignition region Yij (n), combined with the image gradient operator,
其中，Gx，Gy代表在像素点上水平和垂直梯度方向的梯度值，分别由水平方向和垂直方向两个相邻点像素灰度值之差决定，当图像梯度G大于某一阈值时，停止迭代；Among them, Gx , Gy represent the gradient value of the horizontal and vertical gradient directions on the pixel point, which are respectively determined by the difference between the gray value of two adjacent pixels in the horizontal direction and vertical direction. When the image gradient G is greater than a certain threshold , stop iteration;
步骤2.4，将点火区域Yij(n)作为输出，实现故障区域的检测。In step 2.4, the ignition region Yij (n) is used as the output to realize the detection of the fault region.
进一步，所述步骤2.2中由模糊聚类模型进行归类得到神经元集合Sc的实现方式如下，Further, in the step 2.2, the neuron setSc is obtained by classifying by the fuzzy clustering model in the following way,
根据模糊聚类模型，建立目标函数为，According to the fuzzy clustering model, the objective function is established as,
其中，Φ(·)代表非线性映射函数，p为模糊因子，uik表示第k个点到第i个聚类中心的隶属度，c为聚类的类别数，N是待聚类的数据个数，vi为第i个聚类中心，xk对应描述待聚类的数据特征，Nk表示该像素xr的8邻域，邻域像素的影响由α/NR控制，而NR为邻域Nk像素的个数，α为人工设定参数；Among them, Φ( ) represents the nonlinear mapping function, p is the fuzzy factor, uik represents the degree of membership from the kth point to the i-th clustering center, c is the number of clustering categories, and N is the data to be clustered number, vi is the i-th clustering center, xk corresponds to describe the data features to be clustered, Nk represents the 8 neighbors of the pixel xr , the influence of the neighborhood pixels is controlled by α/NR , and NR is the number of Nk pixels in the neighborhood, and α is a parameter set manually;
为了求解隶属度uik，将式(九)中的约束条件通过拉格朗日乘子λ转换为无约束的优化目标函数，可得，In order to solve the degree of membership uik , the constraints in formula (9) are transformed into an unconstrained optimization objective function through the Lagrangian multiplier λ, which can be obtained as,
然后，对上式求得到，Then, for the above formula get,
式中In the formula
Dik＝||Φ(xk)-Φ(vi)||2＝2-2K(xk,vi)；K(xk,vi)为高斯核函数，定义为，Dik ＝||Φ(xk )-Φ(vi )||2 ＝2-2K(xk ,vi ); K(xk ,vi ) is a Gaussian kernel function defined as,
K(xk,vi)＝Cs·exp(-||xk-vi||2/σ2) (十二)K(xk ,vi )＝Cs exp(-||xk -vi ||2 /σ2 ) (12)
其中，Cs为归一化常数，σ为尺度；Among them, Cs is the normalization constant, σ is the scale;
最后，根据最终求解得到的隶属度uik，将集合X中与点火区域Yij(n)相似的神经元和非相似神经元分离。Finally, according to the degree of membership uik obtained from the final solution, the neurons similar to the firing area Yij (n) in the set X are separated from the non-similar neurons.
进一步的，所述步骤2.3中，当图像梯度G大于20时，停止迭代。Further, in the step 2.3, when the image gradient G is greater than 20, the iteration is stopped.
进一步的，利用红外成像仪获得故障电气设备的红外图像。Further, an infrared imager is used to obtain an infrared image of the faulty electrical equipment.
本发明利用的技术原理为PCNN：The technical principle that the present invention utilizes is PCNN:
PCNN模型是一种模仿猫等哺乳动物视觉同步发放脉冲的神经网络，相比于其他神经网络应用于图像处理，PCNN不需要复杂训练，且每一个神经元依次对应图像I中的一个像素(i,j)，构成一个单层的图像处理系统，继而依靠神经元内在机制例如耦合、阈值以及反馈等方式完成感兴趣区域的提取。The PCNN model is a neural network that imitates cats and other mammals to send pulses synchronously. Compared with other neural networks applied to image processing, PCNN does not require complicated training, and each neuron corresponds to a pixel in the image I in turn (i ,j), constitute a single-layer image processing system, and then rely on the internal mechanism of neurons such as coupling, threshold and feedback to complete the extraction of the region of interest.
神经元的输入主要由反馈输入F和连接输入L两部分组成，接受来自外部的图像灰度信息以及邻域神经元的脉冲信息。特别地，每一部分输入都包含一个类似指数衰减的漏电积分器，以确保神经元的当前状态和先前状态之间具有关联性，其表达如式(1)～(2)所示，The input of the neuron is mainly composed of two parts: the feedback input F and the connection input L, which accept the image gray information from the outside and the pulse information of the neighboring neurons. In particular, each part of the input contains a leakage integrator similar to exponential decay to ensure that there is a correlation between the current state of the neuron and the previous state, and its expression is shown in equations (1)-(2),
其中，Iij表示神经元ij所对应的图像灰度值，n为迭代索引，下标ij、kl等代表灰度图像像素位置所对应的神经元，其值为对应灰度值Iij，VF和VL分别为放大系数，αF和αL为衰减系数，M和W为权重矩阵，用于连接8邻域神经元Nij，通常将其设置为相邻神经元的欧氏距离的倒数，Among them, Iij represents the gray value of the image corresponding to the neuron ij, n is the iteration index, subscripts ij, kl, etc. represent the neuron corresponding to the pixel position of the gray image, and its value is the corresponding gray value Iij , VF and VL are amplification coefficients respectively, αF and αL are attenuation coefficients, M and W are weight matrices, which are used to connect 8 neighboring neurons Nij , which are usually set as the Euclidean distance of adjacent neurons reciprocal,
主要负责将邻域神经元发放的脉冲信息传递给中心神经元。It is mainly responsible for transmitting the pulse information issued by the neighboring neurons to the central neuron.
然后，输入F和L在非线性耦合方式作用下，使得神经元内在特性发生变化。通常，该神经元内部的特性描述为，Then, under the action of nonlinear coupling mode, the input F and L make the intrinsic characteristics of the neuron change. Typically, the properties inside this neuron are described as,
Uij(n)＝Fij(n)·[1+βLij(n)] (4)Uij (n)＝Fij (n)·[1+βLij (n)] (4)
其中连接系数β控制邻域神经元的内部活动强度。当神经元内部活动Uij大于其自身的动态阈值Eij时，神经元会发生点火并形成脉冲，输出为1，即Among them, the connection coefficient β controls the internal activity intensity of the neighboring neurons. When the internal activity Uij of a neuron is greater than its own dynamic threshold Eij , the neuron will ignite and form a pulse, and the output is 1, that is
式中神经元阈值E也具有类似指数衰减的漏电积分器特性，其描述为In the formula, the neuron threshold E also has the characteristics of a leakage integrator similar to an exponential decay, which is described as
Eij(n)＝exp(-αE)Eij(n-1)+VEYij(n-1) (6)Eij (n)＝exp(-αE )Eij (n-1)+VE Yij (n-1) (6)
由式(6)可知，在神经元发生点火之后，其动态阈值会瞬间增加常数VE，然后在衰减因子αE的影响下，阈值呈指数衰减直至该神经元再次发生点火。因此每一个神经元都会有一定的点火频率。由于神经元的邻域连接，点火的神经元会触发邻域相似的神经元会产生同步振荡现象，即一个神经元点火，会捕获其周围与其相似的神经元并发生同步点火，从而构成PCNN提取图像相似区域的内在机制，适合提取电气故障区域所表现出来的亮度区域。It can be known from formula (6) that after the firing of a neuron, its dynamic threshold will instantly increase by a constantVE , and then under the influence of the decay factorαE , the threshold decays exponentially until the neuron fires again. Therefore, each neuron will have a certain firing frequency. Due to the neighborhood connection of neurons, the firing neurons will trigger similar neurons in the neighborhood to produce synchronous oscillations, that is, when a neuron fires, it will capture its surrounding neurons similar to it and fire synchronously, thus forming a PCNN extraction. Intrinsic mechanism of similar regions in images, suitable for extracting brightness regions exhibited by electrical fault regions.
本发明的有益效果是：本发明提出了一种有效的红外图像故障区域提取方法，该方法依据PCNN模型自身所具有的同步点火特性，将模型参数进行了优化配置，特别地，动态阈值以及连接系数的优化设置是依据故障区域所表现高亮、相似特性，同时再结合故障区域与非故障区域之间存在的一些特性，提出并采用了邻域变化的局部特征，进而为PCNN模型的迭代设置停止规则。该方法能够自动高效的识别电气设备故障区域。The beneficial effects of the present invention are: the present invention proposes an effective infrared image fault area extraction method, which optimizes the model parameters according to the synchronous ignition characteristics of the PCNN model itself, especially the dynamic threshold and connection The optimal setting of the coefficients is based on the highlight and similar characteristics of the fault area, and at the same time combined with some characteristics between the fault area and the non-fault area, the local characteristics of the neighborhood change are proposed and used, and then the iterative setting of the PCNN model stop rule. The method can automatically and efficiently identify fault areas of electrical equipment.
附图说明Description of drawings
图1是本发明实施例流程图；Fig. 1 is a flowchart of an embodiment of the present invention;
图2为本发明实施例电力设备故障的红外检测图像；Fig. 2 is the infrared detection image of the power equipment fault of the embodiment of the present invention;
图3为改进的PCNN模型处理结果；Fig. 3 is the improved PCNN model processing result;
图4为本发明实施例在不添加局部特征约束下的处理结果；Fig. 4 is the processing result of the embodiment of the present invention without adding local feature constraints;
图5是本发明实施例的故障区域提取结果。Fig. 5 is the result of fault area extraction according to the embodiment of the present invention.
具体实施方式Detailed ways
为了便于本领域普通技术人员理解和实施本发明，下面结合附图及实施例对本发明作进一步的详细描述，应当理解，此处所描述的实施示例仅用于说明和解释本发明，并不用于限定本发明。In order to facilitate those of ordinary skill in the art to understand and implement the present invention, the present invention will be described in further detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the implementation examples described here are only used to illustrate and explain the present invention, and are not intended to limit this invention.
本发明的基于局部特征的PCNN电力故障区域检测方法依据PCNN模型自身所具有的同步点火特性，将模型参数进行了优化配置，特别地，动态阈值以及连接系数的优化设置是依据故障区域所表现高亮、相似特性，同时再结合故障区域与非故障区域之间存在的一些特性，提出并采用了邻域变化的局部特征，进而为PCNN模型的迭代设置停止规则。该方法能够自动高效的识别电气设备故障区域。The PCNN power fault area detection method based on local features of the present invention optimizes the model parameters according to the synchronous ignition characteristics of the PCNN model itself. In particular, the optimal setting of the dynamic threshold and connection coefficient is based on the high Brightness and similarity characteristics, combined with some characteristics between the faulty area and the non-faulty area, proposed and adopted the local characteristics of the neighborhood change, and then set the stopping rule for the iteration of the PCNN model. The method can automatically and efficiently identify fault areas of electrical equipment.
本实施例对真实的电力设备红外图像故障区域提取处理，采用基于局部特征的PCNN方法，对电力设备故障区域进行检测，具体流程如图1所示，包括以下步骤：This embodiment extracts and processes real power equipment infrared image fault areas, and uses the PCNN method based on local features to detect power equipment fault areas. The specific process is shown in Figure 1, including the following steps:
(1)初始化，将原图像作为输入赋给神经元网络，并根据图像最高亮度区域获得初始化神经元脉冲发放区域Y(0),其中Y(0)＝{ij|Fij(n)＝＝Th}，Th为图像最高灰度值，按式(一)设置权重M和W，(1) Initialization, assign the original image as input to the neuron network, and obtain the initialized neuron pulse firing area Y(0) according to the highest brightness area of the image, where Y(0)={ij|Fij (n)== Th }, Th is the highest gray value of the image, and the weights M and W are set according to formula (1),
其中，σh为高斯尺度，设置为1或更大，C为归一化系数；Among them, σh is the Gaussian scale, set to 1 or greater, and C is the normalization coefficient;
(2)迭代，按式(二)～(八)进行迭代计算，(2) Iteration, carry out iterative calculation according to formula (two)～(eight),
Uij(n)＝Fij(n)·[1+βLij(n)] (四)Uij (n)＝Fij (n)·[1+βLij (n)] (4)
X＝{ij|Lij(n)＞0}∪{ij|Yij(n)＝0} (七)X＝{ij|Lij (n)＞0}∪{ij|Yij (n)＝0} (7)
其中，Eg为全局阈值，Ωc为先前点火区域。神经元集合Sc是在集合X内与点火区域Yij(n)相似的神经元，其由模糊聚类模型进行归类得到。本发明在模糊聚类模型的基础上，添加空间信息，避免噪声以及边缘模糊等引起的干扰，并将特征空间通过非线性映射函数Φ(·)投影，建立目标函数为，where Eg is the global threshold and Ωc is the previously fired region. The neuron set Sc is the neurons similar to the firing area Yij (n) in the set X, and it is classified by the fuzzy clustering model. On the basis of the fuzzy clustering model, the present invention adds spatial information to avoid interference caused by noise and edge blurring, and projects the feature space through the nonlinear mapping function Φ(·), and establishes the objective function as,
其中，p为模糊因子，uik表示第k个点到第i个聚类中心的隶属度，c为聚类的类别数，N是待聚类的数据个数，vi为第i个聚类中心，xk对应描述待聚类的数据特征，本例为像素灰度值，Nk表示该像素xr的8邻域，邻域像素的影响由α/NR控制，而NR为邻域Nk像素的个数，α为人工设定参数，本实施例设置为1。为了求解隶属度uik，将式(九)中的约束条件通过拉格朗日乘子λ转换为无约束的优化目标函数，可得，Among them, p is the fuzzy factor, uik represents the membership degree from the kth point to the i-th cluster center, c is the number of cluster categories, N is the number of data to be clustered, vi is the i-th cluster Class center, xk corresponds to describe the data features to be clustered, in this example, the gray value of the pixel, Nk represents the 8 neighbors of the pixel xr , the influence of the neighborhood pixels is controlled by α/NR , andNR is The number of Nk pixels in the neighborhood, α is a parameter set manually, which is set to 1 in this embodiment. In order to solve the degree of membership uik , the constraints in formula (9) are transformed into an unconstrained optimization objective function through the Lagrangian multiplier λ, which can be obtained as,
然后，对上式求得到，Then, for the above formula get,
式中Dik＝||Φ(xk)-Φ(vi)||2＝2-2K(xk,vi)。通常，K(xk,vi)为高斯核函数，定义为，In the formula Dik =||Φ(xk )−Φ(vi )||2 =2−2K(xk ,vi ). Usually, K(xk ,vi ) is a Gaussian kernel function defined as,
K(xk,vi)＝Cs·exp(-||xk-vi||2/σ2) (十二)K(xk ,vi )＝Cs exp(-||xk -vi ||2 /σ2 ) (12)
其中，Cs为归一化常数，σ为尺度。where Cs is the normalization constant and σ is the scale.
由此，根据最终求解得到的隶属度uik，便可将集合X中与点火区域Yij(n)相似的神经元和非相似神经元分离，本例记点火区域Yij(n)相似的神经元集合为Sc。Thus, according to the membership degree uik obtained from the final solution, the neurons similar to the ignition area Yij (n) in the set X can be separated from the non-similar neurons. In this example, the neurons similar to the ignition area Yij (n) The set of neurons is Sc .
(3)停止规则：(3) Stopping rules:
每一次PCNN迭代，得到脉冲点火区域Y，本发明针对其点火区域边界，结合图像梯度算子，Each PCNN iteration obtains the pulse ignition area Y, and the present invention aims at its ignition area boundary, combined with the image gradient operator,
其中，Gx，Gy代表在像素点上水平和垂直梯度方向的梯度值，分别由水平方向和垂直方向两个相邻点像素灰度值之差决定。显然，当G越大，越可能是故障区域的边界，本实施例中设定图像梯度G为20。Among them, Gx , Gy represent the gradient values in the horizontal and vertical gradient directions on the pixel point, which are determined by the difference between the gray value of two adjacent pixel pixels in the horizontal direction and vertical direction, respectively. Obviously, when G is larger, it is more likely to be the boundary of the fault area. In this embodiment, the image gradient G is set to 20.
当迭代运行满足该停止规则则停止迭代。When the iteration operation satisfies the stop rule, the iteration is stopped.
(4)输出：将点火区域Yij(n)作为输出，实现故障区域的检测。(4) Output: The ignition region Yij (n) is used as the output to realize the detection of the fault region.
本发明模型中需要初始化式(一)核尺度系数σh＝1.0，聚类算法中α＝0.1；σ＝150以及设置初始的点火神经元具有最高的反馈输入。In the model of the present invention, it is necessary to initialize formula (1) kernel scale coefficient σh =1.0, α=0.1; σ=150 in the clustering algorithm and set the initial firing neuron to have the highest feedback input.
为了验证本发明方法的有效性，在真实红外图像上测试模型的故障区域提取性能。此外，选用改进的PCNN模型[1]、不添加局部特征的PCNN模型分割方法与本发明的添加局部特征的PCNN方法进行了对比。所有算法均在Intel(R)CoreTM 2Duo i5CPU 4GB内存PC机Matlab7。10(2010b)上编程实现。图3列出了一些具有代表性红外检测图像。实验中，改进的PCNN模型的内在参数按其文献[1]设置为：VL＝0.01；α＝0.1；β＝0.1；T0＝255；最大迭代次数为25次。本发明模型中需要初始化核尺度系数σh＝1.0，聚类算法中α＝0.1；σ＝150以及设置初始的点火神经元具有最高的反馈输入。In order to verify the effectiveness of the method of the present invention, the fault region extraction performance of the model is tested on real infrared images. In addition, the improved PCNN model[1] , the PCNN model segmentation method without adding local features and the PCNN method adding local features of the present invention were compared. All algorithms are programmed on Intel(R) CoreTM 2Duo i5CPU 4GB memory PC Matlab7.10(2010b). Figure 3 lists some representative infrared inspection images. In the experiment, the internal parameters of the improved PCNN model are set as follows according to its literature [1]: VL =0.01; α=0.1; β=0.1; T0 =255; the maximum number of iterations is 25. In the model of the present invention, it is necessary to initialize the kernel scale coefficient σh =1.0, in the clustering algorithm α=0.1; σ=150 and set the initial firing neuron to have the highest feedback input.
从图3中可以看出，改进的PCNN模型，由于采用了总体类绝对差法决定PCNN迭代规则，并将阈值进行了适当简化，从而针对一些高亮区域，能够得到较好的提取，然而对应一些次高亮度的故障区域，则容易将其忽略，产生欠分割，例如图3中第3幅图的分割结果，其原因一方面受制于其固定的一些参数设置，其次由于停止迭代规则在参数的约束下，也会引起模型不能完整的提取故障区域，例如图3中第2、4幅图像的分割结果。而本发明方法，一方面将PCNN模型参数进行了优化设置，首先将阈值与区域特性关联，其次通过聚类方法替代连接系数以确保PCNN同步点火，使得模型具备了自适应迭代特性，然后再采用局部邻域特性，以控制PCNN同步点火范围。其次，本发明方法结合了图像中电气设备故障的灰度分布特性，提出了局部特性来提升本文模型的适用性。相比于不添加局部特征约束，图4给出了相应的提取结果。从图中可看出，部分图像故障区域并没有完整地从图像中划分出来，出现了欠分割现象，而结合局部特征约束，最终得到较为理想的故障区域提取结果。It can be seen from Figure 3 that the improved PCNN model uses the overall class absolute difference method to determine the iteration rules of PCNN and simplifies the threshold appropriately, so that some highlighted regions can be better extracted, but the corresponding Some sub-brightness fault areas are easy to be ignored, resulting in under-segmentation. Under the constraints of , it will also cause the model to fail to completely extract the fault area, such as the segmentation results of the 2nd and 4th images in Figure 3. In the method of the present invention, on the one hand, the parameters of the PCNN model are optimized. First, the threshold value is associated with the regional characteristics, and secondly, the connection coefficient is replaced by the clustering method to ensure that the PCNN is ignited synchronously, so that the model has adaptive iterative characteristics, and then adopts Local neighborhood features to control PCNN synchronization firing range. Secondly, the method of the present invention combines the gray distribution characteristics of electrical equipment faults in the image, and proposes local characteristics to improve the applicability of the model in this paper. Compared with not adding local feature constraints, Figure 4 shows the corresponding extraction results. It can be seen from the figure that some image fault areas are not completely separated from the image, and under-segmentation occurs. However, combined with local feature constraints, a relatively ideal fault area extraction result is finally obtained.
应当理解的是，本说明书未详细阐述的部分均属于现有技术。It should be understood that the parts not described in detail in this specification belong to the prior art.
应当理解的是，上述针对较佳实施例的描述较为详细，并不能因此而认为是对本发明专利保护范围的限制，本领域的普通技术人员在本发明的启示下，在不脱离本发明权利要求所保护的范围情况下，还可以做出替换或变形，均落入本发明的保护范围之内，本发明的请求保护范围应以所附权利要求为准。It should be understood that the above-mentioned descriptions for the preferred embodiments are relatively detailed, and should not therefore be considered as limiting the scope of the patent protection of the present invention. Within the scope of protection, replacements or modifications can also be made, all of which fall within the protection scope of the present invention, and the scope of protection of the present invention should be based on the appended claims.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710653562.XACN107292883B (en) | 2017-08-02 | 2017-08-02 | A PCNN Power Fault Area Detection Method Based on Local Features |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710653562.XACN107292883B (en) | 2017-08-02 | 2017-08-02 | A PCNN Power Fault Area Detection Method Based on Local Features |
| Publication Number | Publication Date |
|---|---|
| CN107292883A CN107292883A (en) | 2017-10-24 |
| CN107292883Btrue CN107292883B (en) | 2019-10-25 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710653562.XAActiveCN107292883B (en) | 2017-08-02 | 2017-08-02 | A PCNN Power Fault Area Detection Method Based on Local Features |
| Country | Link |
|---|---|
| CN (1) | CN107292883B (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108197637A (en)* | 2017-12-11 | 2018-06-22 | 国网上海市电力公司 | A kind of high pressure cabinet switch detection method based on deep learning |
| CN109492617A (en)* | 2018-11-29 | 2019-03-19 | 国网上海市电力公司 | Power cable diagnostic method based on K-means clustering algorithm |
| CN110276776B (en)* | 2019-06-17 | 2021-06-15 | 北京博睿维讯科技有限公司 | An Adaptive Object Detection Method Based on SPCNN |
| CN110782437B (en)* | 2019-10-18 | 2022-05-06 | 国网电力科学研究院武汉南瑞有限责任公司 | An Improved PCNN Power Fault Image Spatial Location Method Based on Boundary Features |
| CN112710914B (en)* | 2020-12-16 | 2022-06-17 | 西华大学 | Fault diagnosis method of intelligent substation considering control center fault information tampering |
| CN118657775A (en)* | 2024-08-20 | 2024-09-17 | 南京中鑫智电科技有限公司 | An intelligent diagnosis method for transformer bushing based on Faster RCNN algorithm |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103679670A (en)* | 2012-09-25 | 2014-03-26 | 中国航天科工集团第二研究院二〇七所 | A PCNN multisource image fusion method based on an improved model |
| CN104569666A (en)* | 2014-12-25 | 2015-04-29 | 重庆大学 | Power transformer fault prediction method based on electricity-graph model |
| CN104601109A (en)* | 2014-12-25 | 2015-05-06 | 重庆大学 | Photovoltaic hot spot effect detection method for electricity-graph model |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20030076992A1 (en)* | 2001-06-26 | 2003-04-24 | Banish Michele R. | Neural network based element, image pre-processor, and method of pre-processing using a neural network |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103679670A (en)* | 2012-09-25 | 2014-03-26 | 中国航天科工集团第二研究院二〇七所 | A PCNN multisource image fusion method based on an improved model |
| CN104569666A (en)* | 2014-12-25 | 2015-04-29 | 重庆大学 | Power transformer fault prediction method based on electricity-graph model |
| CN104601109A (en)* | 2014-12-25 | 2015-05-06 | 重庆大学 | Photovoltaic hot spot effect detection method for electricity-graph model |
| Title |
|---|
| 《Computational Mechanisms of Pulse-Coupled Neural Networks: A Comprehensive Review》;Kun Zhan，et al;《Springer Link》;20160719;第1-16页* |
| Publication number | Publication date |
|---|---|
| CN107292883A (en) | 2017-10-24 |
| Publication | Publication Date | Title |
|---|---|---|
| CN107292883B (en) | A PCNN Power Fault Area Detection Method Based on Local Features | |
| CN106595551B (en) | Ice covering thickness detection method in powerline ice-covering image based on deep learning | |
| Liu et al. | Insulator detection in aerial images based on faster regions with convolutional neural network | |
| CN108009628B (en) | Anomaly detection method based on generation countermeasure network | |
| CN112200178B (en) | Transformer substation insulator infrared image detection method based on artificial intelligence | |
| CN103886344B (en) | A kind of Image Fire Flame recognition methods | |
| CN113486877B (en) | Power equipment infrared image real-time detection and diagnosis method based on light artificial intelligent model | |
| Jiang et al. | Visible image recognition of power transformer equipment based on mask R-CNN | |
| CN114137358B (en) | Power transmission line fault diagnosis method based on graph convolution neural network | |
| CN103942557B (en) | A kind of underground coal mine image pre-processing method | |
| CN106596579A (en) | Insulator contamination condition detection method based on multispectral image information fusion | |
| CN111428748A (en) | An infrared image insulator identification and detection method based on HOG feature and SVM | |
| CN112013966A (en) | Power equipment infrared image processing method based on measured temperature | |
| Zhu et al. | Defect identification of wind turbine blade based on multi‐feature fusion residual network and transfer learning | |
| CN110992378B (en) | Method and system for dynamically updating visual tracking aerial photography based on rotor flying robot | |
| Pei et al. | UV‐flashover evaluation of porcelain insulators based on deep learning | |
| Maraaba et al. | A neural network-based estimation of the level of contamination on high-voltage porcelain and glass insulators | |
| CN113034465B (en) | A thermal fault monitoring method, device and medium for power equipment based on infrared images | |
| Feng et al. | Infrared image recognition technology based on visual processing and deep learning | |
| CN110992306B (en) | Method and device for segmenting deteriorated insulator in infrared image based on deep learning | |
| Chen et al. | Attention mechanism-based CNN for surface damage detection of wind turbine blades | |
| Jiskani et al. | Aerial identification of flashed over faulty insulator using binary image classification | |
| Tiancheng et al. | Infrared image segmentation method of current transformer based on deeplabv3+ network | |
| Zhang | Computer image processing and neural network technology for thermal energy diagnosis of boiler plants | |
| CN217181103U (en) | Monitoring system of high-voltage switch cabinet |
| 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 | ||
| CB03 | Change of inventor or designer information | Inventor after:Gu Kaikai Inventor after:Huang Hua Inventor after:Fu Chenzhao Inventor after:Hu Zhengyong Inventor after:Cheng Lin Inventor after:Xu Xiaolu Inventor after:Cai Wei Inventor after:Zhou Zhengqin Inventor after:Ni Hui Inventor after:Xu Jinxia Inventor after:Zhou Dongguo Inventor after:Zhao Kun Inventor before:Gu Kaikai Inventor before:Huang Hua Inventor before:Fu Chenzhao Inventor before:Hu Zhengyong Inventor before:Zhou Dongguo Inventor before:Xu Xiaolu Inventor before:Cai Wei Inventor before:Zhou Zhengqin Inventor before:Ni Hui Inventor before:Xu Jinxia Inventor before:Cheng Lin Inventor before:Zhao Kun | |
| CB03 | Change of inventor or designer information | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |