Time relay image recognition method and detection system

Abstract

A time relay image recognition method and detection system, including: image preprocessing, feature extraction, classification identification and other steps, the image feature vector is input into the pre-trained classifier, and the class identification result of the relay is output. The detection system comprises an image recognition unit, a control unit, a main control unit and a rack. The time relay image recognition method and detection system are provided. The time relay image recognition method based on probability LS-SVM has better robustness, can improve the definition and signal-to-noise ratio of the image, and makes the relay image recognition more accurate.

Description

Technical Field

[0001] This invention relates to the field of relay detection methods and equipment, specifically to a time relay image recognition method and detection system. Background Technology

[0002] In the field of industrial automation control, time relays are key timing components, and the accuracy of their performance parameters directly affects the reliability of the control system. Traditional time relay testing methods mainly rely on manual operation and visual interpretation, which suffers from low testing efficiency and strong subjectivity. Although some automated testing equipment has emerged on the market, these devices still have significant shortcomings in terms of testing accuracy, ease of operation, and functional completeness. Currently, more advanced time relay testing devices typically use a single camera to capture images of the relay display, combined with a simple mechanical structure to achieve basic testing functions. These devices generally consist of a conveyor mechanism, an industrial camera, control circuitry, and image processing software. The workflow involves manually placing the relay under test, initiating the test, having the camera capture images of the relay display screen and perform simple image recognition, and finally displaying the test results.

[0003] Commonly used image recognition techniques for time relays fall into two main categories: template matching and feature classification. Template matching is fast, but it lacks robustness to rotational and scaling changes in the time relay. Feature classification methods typically extract features of the target object, which are often rotationally and translationally invariant, and then input these features into machine learning algorithms for detection and recognition. However, before inputting the features, image denoising is also necessary to obtain a clearer image with a higher signal-to-noise ratio.

[0004] Accurately reconstructing real images under noise and outlier interference is crucial for image recognition of time relays. Currently, major image denoising techniques fall into four categories: filtering, sparse representation, external prior, and low-rank clustering. Filtering-based denoising algorithms, due to their close-up templates, often ignore image edge information. Sparse representation doesn't necessarily guarantee image sparsity or noise non-sparseness. Low-rank clustering fails to achieve its intended effect under high noise levels. External prior, on the other hand, focuses on summarizing image patterns based on inherent image attributes. The most representative work is the Gaussian Mixture Model (GMM). SVM and deep learning also belong to this category. These algorithms can establish relatively accurate mapping relationships but are also susceptible to interference; therefore, robustness is a key issue. Thus, improving the accuracy of time relay image recognition is of great significance in this field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a time relay image recognition method, specifically including the following steps: Image preprocessing S100: Acquire the original image of the relay to be identified, and use a robust modeling method based on probabilistic reasoning to denoise the original image to obtain a denoised image; wherein, the robust modeling method based on probabilistic reasoning estimates the statistical distribution of the model parameters and uses the expected value or other statistical features of the model parameters to replace the deterministic parameters, so as to suppress the influence of random noise and uncertainty during the acquisition process on the denoising model. Feature extraction S200: Extract geometrically invariant image features from the denoised image to obtain an image feature vector; In the classification and recognition step S300, the image feature vector is input into a pre-trained classifier, and the category recognition result of the relay is output.

[0006] Furthermore, the robust modeling method based on probabilistic reasoning is a probabilistic least squares support vector machine; the statistical distribution of the estimated model parameters specifically includes: dividing the training image data into multiple data subsets by multiple sampling, training the model parameter samples independently for each subset, and constructing the probability distribution function of the parameters based on all parameter samples.

[0007] Furthermore, the multiple sampling method is selected from one of the following: random resampling, bootstrap sampling, and cross-validation grouping; the probability distribution function is a normal distribution, and its distribution parameters are estimated from the mean and variance of the parameter samples.

[0008] Furthermore, the step of replacing deterministic parameters with the expected value of model parameters specifically includes: deriving the probability density function of the kernel function based on the probability distribution function of the parameters, and calculating the expected value of the kernel function; constructing an objective function in a probabilistic sense using the expected value of the kernel function, and solving it to obtain the denoising model.

[0009] Furthermore, before the image preprocessing S100, the method further includes: locating and segmenting the time relay region of the original image, and extracting the region of interest containing the main body of the time relay.

[0010] Furthermore, the geometrically invariant image feature is an oriented gradient histogram feature; the geometric invariance includes rotation invariance and translation invariance.

[0011] Furthermore, the extraction process of the directional gradient histogram features includes: dividing the denoised image into multiple cell units, calculating the gradient direction histogram in each cell unit, forming adjacent cell units into blocks and performing normalization processing, and concatenating the feature vectors of all blocks to obtain a global feature vector.

[0012] Furthermore, the classifier is a support vector machine classifier; the kernel function of the support vector machine classifier is selected from one of the following: linear kernel function, polynomial kernel function, radial basis function.

[0013] Furthermore, the support vector machine classifier adopts a soft-margin form, and controls the balance between the classification margin and the cost of misclassification through a penalty parameter; the training process of the support vector machine classifier includes: transforming the original optimization problem into a dual problem, solving the dual problem to obtain support vectors and Lagrange factors, and constructing a classification decision function.

[0014] A time relay detection system is also proposed, comprising: The image recognition unit includes a camera for acquiring real-time images of the time relay and executes the aforementioned time relay image recognition method to process the acquired real-time images. The control unit is used to drive the camera to move along the X, Y, and Z axes in order to accurately acquire real-time images from the time relay; The main control unit is used to process the recognition results of the image recognition unit, judge and output the detection results according to the detection requirements, and can issue control commands to the manipulation unit. The rack is used to mount the image recognition unit, the control unit and the main control unit.

[0015] Compared with existing technologies, the technical solution of this application has the following beneficial effects: The time relay image recognition method and detection system proposed in this invention, based on probabilistic LS-SVM, has better robustness, can improve image clarity and signal-to-noise ratio, and makes relay image recognition more accurate. Attached Figure Description

[0016] Figure 1 Overall methodology and process; Figure 2 Flowchart of probabilistic LS-SVM-based denoising; Figure 3 Feature extraction and SVM classification process; Figure 4 Schematic diagram of a time relay detection system. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Combination Figures 1-3 The present invention provides a detailed description of the time relay image recognition method proposed in this invention.

[0019] An unknown system with random noise can be represented as: (2-1), in, x Represents input, y Represents output, Represents random noise. f Represents the model function. Assume a given dataset... The system is modeled using LS-SVM as follows: (2-2), in, w It is a weight vector. b It is a bias term. It is a projection function. This is the model output.

[0020] Model (2-2) can be solved by solving the following optimization problem: (2-3), in, e i It is an error term. This is the regularization parameter. To solve the above optimization problem, the Lagrange equation is constructed as follows: (2-4), It is the Lagrange factor. Solving the Lagrange equation yields: (2-5), (2-5) can be transformed into the following equation: (2-6), in, , It is a kernel function matrix and has .

[0021] By solving equation (2-6), the final LS-SVM model can be expressed as: (2-7), Substituting (2-5) into (2-4), the optimization function becomes: (2-8), LS-SVM has good denoising performance; however, this algorithm is also susceptible to data uncertainty. When collecting input or output data, limitations such as the collection environment and equipment can cause some collected data to deviate from the actual data. This can affect the model, leading to a form (2-1). Research has found that both input and output uncertainty cause changes in the kernel function K. Because the kernel function parameters are obtained through cross-validation, this means that regardless of changes in the input and output, the values ​​of the kernel function parameters will change. This demonstrates that the kernel function can directly reflect the uncertainty of the model. For example, given a certain level of uncertainty... ε i Below, for the Gaussian kernel function There will be a corresponding kernel function parameter. The LS-SVM model. All uncertainties will lead to a set of parameters. This will construct a distribution for the parameters and kernel function. Therefore, uncertainty... ε i The randomness of the kernel function is the main source of this randomness. However, in traditional least squares support vector machines, the model is based on a deterministic kernel function and the randomness of the kernel function is not considered, which often reduces the robustness of modeling uncertain systems.

[0022] Based on the above analysis, it can be seen that when subjected to uncertainty... ε i When this happens, the kernel function and optimization function will also change: (2-9), (2-10), in, and These are the kernel function K and the parameters. Changes caused by uncertainty. Solving (2-10) yields: (2-11), Due to ε i Due to the influence of [various factors], the model parameters often change, resulting in an unstable model. To eliminate these influences and obtain the most reliable model, we introduce the expectation value instead of the original value: (2-12), (2-13), As can be seen from the above formula derivation, replacing the original kernel function with the expected kernel function can effectively eliminate the uncertainty of input and output. Therefore, this paper proposes a probabilistic LS-SVM method for robust modeling of stochastic uncertainties and unknown nonlinear systems, such as... Figure 2-4 As shown, the distribution of the model is first constructed using probabilistic inference. Using this distribution information, a new objective function is developed in a probabilistic sense. By solving this objective function, a robust LS-SVM model can be obtained, which can effectively model unknown nonlinear systems even under random uncertainty.

[0023] The following details the probabilistic LS-SVM image denoising method.

[0024] Based on the above analysis, by incorporating the model's probability information into equation (2-8), the objective function becomes: (2-14), in, It is a kernel function matrix and , It is a regularization factor in a probabilistic sense. It is the probability density function of the Gaussian kernel matrix. It is the expectation of K.

[0025] This objective function has the following characteristics: It is a regularization term in a probabilistic sense, used to ensure the generalization ability of the model and avoid overfitting.

[0026] It is both an empirical risk term and a structural risk term, which contains both model probability information and error information.

[0027] It is the global error term of the model, which ensures the model's accuracy and predictive ability.

[0028] From a probabilistic perspective, the probability distribution of the model can better reflect the inherent characteristics of the original system under random noise. Therefore, the established model can effectively represent unknown nonlinear systems even in the presence of noise and uncertainty.

[0029] To solve the optimization problem (2-14), we first need to solve for K and The expectation, due to The method for calculating the expectation of K is similar to that for K, so this chapter mainly discusses how to obtain the probability density function of K and solve for its expectation. To obtain the probability density function of K... First, we need to evaluate the parameters of the kernel function K. Therefore, a probability distribution LS-SVM algorithm needs to be established to achieve this goal.

[0030] Step 1: Divide the entire data into q groups using a resampling method.

[0031] Step 2: Divide each set of data into s subsets on average.

[0032] Step 3: Construct a model for each subset using traditional LS-SVM and obtain the corresponding kernel function parameters. Using this method, we can obtain [the following] for each group. .

[0033] Step 4: Calculate the values ​​in each group The mean of all groups can be obtained in this way. .

[0034] Step 5: Building The probability density function. According to the central limit theorem, when the amount of data is large enough... It will follow a normal distribution: (2-15), in, and These are the mean and variance of a normal distribution. They can be derived from maximum likelihood estimation: (2-16), (2-17), In the above algorithm process, In each group The mean of the model is used, which has two advantages: replacing the original values ​​with the mean effectively reduces the impact of noise and outliers, making the model more accurate; the mean satisfies the requirements of the central limit theorem, reducing the difficulty of solving the model.

[0035] By resampling and solving using the above algorithm, we obtained the kernel function parameters. The distribution of parameters will be determined next. The distribution of the kernel function K is obtained from the distribution of the kernel function K.

[0036] After obtaining the distribution of kernel function parameters, for the kernel function... Its distribution function is as follows: (2-18), The probability density function is as follows: (2-19), in, F Represents the distribution function.

[0037] because It follows a normal distribution, so The probability density function is: (2-20), As is well known, the value of the kernel function is always positive, therefore for The distribution function is: (2-21), when When the distribution function is: (2-22), Differentiating (2-22), we can obtain k The probability density function: (2-23), Therefore, we can obtain k Expectations: (2-24), In fact, formula (2-24) only applies to continuous distribution functions, while for discrete distribution functions: (2-25), in S j Discretization k The value of the first j In each region, we will k The value is discretized from [0,1] into N regions. P ( k j The calculation method for ) is as follows: (2-26), in,, j_low and j_high It is the first k The Middle j The boundaries of each region.

[0038] Thus, we obtained K and K using the resampling method. To obtain the expected value, review the objective function (2-14) and solve for it: (2-27), (2-28), Combining equations (2-27) and (2-28), we get: (2-29), in, , .make: , Solving equation (2-29) simultaneously yields the following result: (2-30), The final valuation function becomes: (2-31), As shown in (2-11), when the system is affected by noise or uncertainty, the traditional LS-SVM evaluation function and the true function are respectively: (2-32), (2-33), As shown in (2-32) and (2-33), the estimated value and the true value of the traditional LS-SVM will have a certain difference. The bias increases when the noise level is high, which significantly affects the accuracy of the model. For probabilistic LS-SVM, combining (2-31) and (2-25), we can obtain: (2-34), From the above three equations, it is easy to see that traditional LS-SVM deviates from the true value, while probabilistic LS-SVM approximates the true value more closely. Therefore, probabilistic LS-SVM has a better approximation ability.

[0039] When subjected to noise, the deviation of traditional LS-SVM is: (2-35), The output of probabilistic LS-SVM is: (2-36), The deviation at this point becomes: (2-37), It is obvious that Therefore, probabilistic LS-SVM has stronger robustness.

[0040] The above probabilistic LS-SVM image denoising method can be summarized as follows: Input: Noisy image X .

[0041] Output: Real image Y .

[0042] Step 1: For a given training set The input is divided into q groups by resampling, and each group is trained with traditional LS-SVM to obtain the corresponding kernel function parameters.

[0043] Step 2: Construct the probability density function of K based on (2-19).

[0044] Step 3: Calculate the expectation of the kernel function K according to (2-25).

[0045] Step 4: Construct a new probability density function (2-14) and solve it to obtain a new model (2-31).

[0046] After image denoising, a time relay image recognition classifier is established.

[0047] After denoising, the signal-to-noise ratio, contrast, and sharpness of the original image are significantly improved. The second step, feature extraction, then proceeds to the time relay image recognition stage. Since HOG features have excellent properties and can effectively represent image features, they are selected as the extracted features. Finally, a classifier for time relay image recognition needs to be established. Because SVM has good classification performance and strong generalization ability, this paper selects the SVM algorithm as the classifier for time relay image recognition.

[0048] The SVM programming problem can be solved by solving the following optimization problem: (2-38), The constraints are: (2-39), in, C This is a penalty parameter used to adjust the degree of influence of the classification margin and the number of misclassified points. As slack variables, introduce a Lagrange factor. Construct the Lagrange function: , Taking the derivative with respect to each parameter, we can obtain: (2-41), Substitute (2-41) into (2-40) and transform the original problem into its dual problem: (2-42), Among them, kernel function The final SVM classifier expression is: (2-43), This embodiment also relates to an automatic detection system for a time relay, see details below. Figure 4 .

[0049] The system includes an image recognition unit, including a camera 1 for acquiring real-time images of a time relay, and processes the acquired real-time images using the image recognition method described above. The control unit 2 is equipped with a joystick 21 for jogging the time relay. The joystick is driven by a drive device to move along the X, Y, and Z axes to jog the time relay. Specifically, in this embodiment, a camera 1 is mounted on the top of the device, with two cameras positioned on either side of the top, for capturing images of the time relay. The jogging operation of the joystick 21 is driven by a Z-axis drive mechanism (which can be a pneumatic or electric cylinder). The joystick 21 can be driven by a lead screw and a drive motor to move along the X and Y axes to reach the operating area of ​​the time relay.

[0050] The main control unit is used to process the recognition results of the image recognition unit, judge and output the detection results according to the detection requirements, and can issue control commands to the manipulation unit. The frame 3 is used to mount the image recognition unit, the control unit, and the main control unit. In this embodiment, the frame 3 is a frame structure, with the top for mounting the camera 1, the middle part for mounting the control unit 2, and the bottom for supporting the time relay to be measured.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for recognizing time relay images, characterized in that, Specifically, the steps include the following: Image preprocessing S100: Acquire the original image of the relay to be identified, and use a robust modeling method based on probabilistic reasoning to denoise the original image to obtain a denoised image; wherein, the robust modeling method based on probabilistic reasoning estimates the statistical distribution of the model parameters and uses the expected value or other statistical features of the model parameters to replace the deterministic parameters, so as to suppress the influence of random noise and uncertainty during the acquisition process on the denoising model. Feature extraction S200: Extract geometrically invariant image features from the denoised image to obtain an image feature vector; In the classification and recognition step S300, the image feature vector is input into a pre-trained classifier, and the category recognition result of the relay is output.

2. The time relay image recognition method as described in claim 1, characterized in that, The robust modeling method based on probabilistic reasoning is a probabilistic least squares support vector machine; the statistical distribution of the estimated model parameters specifically includes: dividing the training image data into multiple data subsets by multiple sampling, training the model parameter samples independently for each subset, and constructing the probability distribution function of the parameters based on all parameter samples.

3. The time relay image recognition method as described in claim 2, characterized in that, The multiple sampling method is selected from one of the following: random resampling, bootstrap sampling, and cross-validation grouping; the probability distribution function is a normal distribution, and its distribution parameters are estimated from the mean and variance of the parameter samples.

4. The time relay image recognition method as described in claim 2, characterized in that, The specific steps of replacing deterministic parameters with the expected value of model parameters include: deriving the probability density function of the kernel function based on the probability distribution function of the parameters, and calculating the expected value of the kernel function; constructing an objective function in a probabilistic sense using the expected value of the kernel function, and solving for the denoising model.

5. The time relay image recognition method as described in claim 1, characterized in that, Before the image preprocessing S100, the method further includes: locating and segmenting the time relay region of the original image, and extracting the region of interest containing the main body of the time relay.

6. The time relay image recognition method as described in claim 1, characterized in that, The geometrically invariant image feature is the directional gradient histogram feature; the geometric invariance includes rotation invariance and translation invariance.

7. The time relay image recognition method as described in claim 6, characterized in that, The extraction process of the directional gradient histogram features includes: dividing the denoised image into multiple cell units, calculating the gradient direction histogram in each cell unit, forming adjacent cell units into blocks and performing normalization processing, and concatenating the feature vectors of all blocks to obtain a global feature vector.

8. The time relay image recognition method as described in claim 1, characterized in that, The classifier is a support vector machine classifier; the kernel function of the support vector machine classifier is selected from one of the following: linear kernel function, polynomial kernel function, radial basis function.

9. The time relay image recognition method as described in claim 8, characterized in that, The support vector machine classifier adopts a soft margin form, and controls the balance between the classification margin and the cost of misclassification through a penalty parameter. The training process of the support vector machine classifier includes: transforming the original optimization problem into a dual problem, solving the dual problem to obtain support vectors and Lagrange factors, and constructing a classification decision function.

10. A time relay detection system, characterized in that, include: The image recognition unit includes a camera for acquiring real-time images of a time relay and executes the time relay image recognition method as described in any one of claims 1-9 to process the acquired real-time images. The control unit is equipped with a joystick for jogging the time relay. The joystick is driven by a drive device to move along the X, Y, and Z axes to jog the time relay. The main control unit is used to process the recognition results of the image recognition unit, judge and output the detection results according to the detection requirements, and can issue control commands to the manipulation unit. The rack is used to mount the image recognition unit, the control unit and the main control unit.

Images