Constructing method for identifying side column reinforcing seat of open wagon

Abstract

The invention relates to a construction method for recognizing a side column reinforcing seat of an open wagon. The construction method comprises the following steps: collecting laser radar data; constructing a feature data set; designing an annotation system; constructing and training a support vector machine classification model; constructing an online learning framework; and application. According to the invention, through laser radar accurate acquisition, feature data set construction, support vector machine model training and online frame operation, a multi-dimensional feature extraction, SVM classification and online learning mechanism is formed, real-time accurate identification of the side column reinforcing seat is realized, the obstacle avoidance requirement of crushing equipment is met, and the operation safety and efficiency are improved.

Description

Technical Field

[0001] This invention relates to a method for identifying side pillar reinforcement seats of open wagons, an automated identification technology at the intersection of intelligent detection technology and artificial intelligence, a process for automated loading and unloading of railway open wagons, and a method for constructing a method for obstacle avoidance of crushing equipment during the unloading of frozen coal. Background Technology

[0002] The side pillar reinforcement seat of an open wagon is a key load-bearing component connecting the side wall of the wagon to the underframe. In frozen coal crushing operations, its raised surface structure is highly susceptible to collision with the crushing equipment, leading to equipment damage or operational interruption. Traditional methods for identifying side pillar reinforcement seats rely on manual visual inspection or fixed sensor threshold judgment, which suffers from low efficiency, poor accuracy, and weak adaptability.

[0003] Existing machine learning-based recognition methods mostly employ offline models, which struggle to handle varying operating conditions such as different levels of wear on open wagons, interference from frozen coal deposits, and sensor drift. Furthermore, general datasets lack targeted reinforcement features, resulting in insufficient model generalization ability. Therefore, there is an urgent need to develop an online recognition method that integrates accurate data acquisition, feature-adapted datasets, and dynamically updated models to meet the high accuracy and reliability requirements of industrial scenarios. Summary of the Invention

[0004] To overcome the problems of existing technologies, this invention proposes a method for identifying side pillar reinforcement seats in open wagons. The method, based on LiDAR data acquisition, feature dataset construction, support vector machine model training, and online framework operation, achieves real-time and accurate identification of side pillar reinforcement seats, improving operational safety and efficiency.

[0005] The objective of this invention is achieved as follows: a method for identifying side pillar reinforcement seats in open wagons, the system used in the method comprising: a lidar 1 installed above the railway wagon and capable of scanning the wagon from beginning to end, the lidar being connected to a digital processor, and the steps of the method being as follows:

[0006] Step 1, LiDAR data acquisition:

[0007] Establish a global coordinate system OXYZ to ensure that the radar coordinate system is aligned with the global coordinate system. The origin O of the coordinate system is set at one end of the open wagon entering the work area. The X-axis is the direction of travel of the open wagon, the Y-axis is the left and right direction of the open wagon, and the Z-axis is the height direction of the open wagon.

[0008] Data Acquisition Process: After the open wagon stops in the work area, the guide rail platform drives the radar to move at a constant speed along the X-axis. In each acquisition cycle, it outputs a YZ plane depth data to form a three-dimensional point cloud matrix.

[0009] Step 2, Feature dataset construction:

[0010] Sub-step 201, Sliding window settings: With the X-axis as the sliding direction, set the window parameters as follows: window size: 5; window step size: 1; window data format: each window contains 5 sets of point cloud data, each set containing all points in the YZ plane at that X position;

[0011] Sub-step 202, multi-dimensional feature extraction:

[0012] Extract 15-dimensional feature vectors from each window. It covers three categories of features: geometric, statistical, and positional.

[0013]

[0014] The geometric eigenvectors include: β1, β2, β3, β4, and β5;

[0015] Where: β1 is the mean Z-value, extracted by the average Z-coordinate of all points within the window; β2 is the variance Z-value, extracted by the variance of the Z-coordinate of all points within the window; β3 is the median Y-value, extracted by the median Y-coordinate of all points within the window; β4 is the maximum Z-switching value, extracted by the maximum difference between adjacent X-positions in Z-value; β5 is the flatness, extracted by the deviation of the fitted plane of the top point set.

[0016] The statistical eigenvectors include: β6, β7, β8, and β9;

[0017] Wherein: β6 is the mean reflection intensity, extracted by the average reflection intensity of all points within the window; β7 is the standard deviation of reflection intensity; β8 is the point density, extracted by the number of points per unit volume; β9 is the percentage of effective points, extracted by the ratio of the number of effective points to the total number of points within the window.

[0018] The location feature vector includes: β 10 β 11 β 12 β 13 β 14 β 15 ;

[0019] Where: β 10 The theoretical position deviation of the X-axis distance is extracted by the following method: the difference between the current window's X-coordinate and the theoretical position of the side column reinforcement seat; β 11 The deviation of the Z-axis from the baseline is extracted by the following method: the difference between the mean Z-axis value of the current window and the baseline value; β 12 The X-coordinate of the window center is extracted using the average of the five X-positions of the window; β 13 The length of the vertical transition segment is extracted using the following method: the span of the point set with a Z-value abrupt change exceeding 50 mm along the X-axis; β14 The length of the top plane is extracted using the following method: the span of a point set with a Z-value variance < 5 mm² along the X-axis; β 15 The reflection intensity gradient is extracted by the difference in reflection intensity at the X position before and after within the window.

[0020] Step 3, Labeling System Design:

[0021] Sub-step 301, Category label:

[0022] definition: , where 1 indicates that the window contains the side column reinforcement seat, and 0 indicates that it does not;

[0023] Judgment criteria: If all three of the following conditions are met, it is marked as 1:

[0024] Vertical transition segment length: Matches width feature;

[0025] Maximum Z mutation value: matches high-level features;

[0026] Mean reflection intensity: matches material characteristics;

[0027] Sub-step 302, standard result storage format:

[0028] The annotation results of each sample are stored in association with the feature vector, and a single sample The standard storage format consists of feature vectors and annotation results, specifically... , where x is the feature vector; The annotation results;

[0029] Step 4, Support Vector Machine classification model construction and training:

[0030] Sub-step 401, Dataset:

[0031] Input: Training dataset Depend on Composed of 1 sample, ,,in, , For the sample size of the dataset, For the first One sample, For the first The feature vector of each sample For the first The annotation results for each sample;

[0032] Output: Separating hyperplane and classification decision function;

[0033] Sub-step 402, Vector Machine Optimization Problem Construction:

[0034] Select an appropriate kernel function and penalty function Construct and solve the convex quadratic programming problem:

[0035]

[0036] in, To obtain the Lagrange multipliers corresponding to the samples, the optimal solution is obtained by solving the above convex quadratic programming problem. ;T is the transpose sign; the kernel function chosen is the Gaussian kernel function:

[0037]

[0038] in, For hyperparameters;

[0039] Sub-step 403, Classification Decision Tree:

[0040] choose A component Meet the conditions ,calculate

[0041]

[0042] in, The optimal intercept of the hyperplane;

[0043] The corresponding vector machine is a Gaussian radial basis function classifier. In this case, the classification decision function is:

[0044]

[0045] in, For mathematical symbolic functions;

[0046] Step 5, Building the online learning framework:

[0047] Sub-step 501, Model accuracy evaluation and update trigger:

[0048] Periodically use new calibration data as the test dataset The test set contains 20-50 data entries. When a prediction error occurs, the model is updated.

[0049] Sub-step 502, implementation of the forgetting strategy:

[0050] To avoid a surge in computation due to the infinite expansion of the training set, a forgetting strategy based on kernel function similarity is introduced, which is implemented in the following steps:

[0051] Step 5021, Adding New Samples: Adding incorrectly predicted samples... The samples are transformed into new feature samples, labeled with the true labels, and then used as samples to be added.

[0052] Step 5022, Similarity Calculation: Calculate the similarity between the sample to be added and the current training set. Kernel function vectors for all samples:

[0053]

[0054] The larger the kernel function value, the higher the sample similarity;

[0055] Step 5023, Sample Set Update: Select The sample corresponding to the maximum value of the kernel function is from the current training dataset. Remove from the middle; remove new feature samples Add to training dataset In this process, the size of the training set is kept stable at a preset threshold to ensure computational efficiency and sample diversity;

[0056] Sub-step 503, Iterative Model Training: Retrain the SVM model based on the updated training set;

[0057] Step 6, Application: Use the trained model for actual recognition operations and continuously train it using the acquired data.

[0058] The advantages and beneficial effects of this invention are as follows: This invention uses LiDAR for precise acquisition, feature dataset construction, support vector machine model training, and online framework operation to form a multi-dimensional feature extraction, SVM classification, and online learning mechanism, thereby achieving real-time and accurate identification of side column reinforcement seats, meeting the obstacle avoidance requirements of crushing equipment, and improving operational safety and efficiency. Attached Figure Description

[0059] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0060] Figure 1 This is a schematic diagram of the lidar used in the method described in the embodiment of the present invention in the identification area of ​​the side pillar reinforcement of an open wagon;

[0061] Figure 2 This is a schematic diagram showing the position of the lidar used in the method described in the embodiment of the present invention relative to the carriage;

[0062] Figure 3 This is a schematic diagram of the principle framework of the method described in the embodiments of the present invention. Detailed Implementation

[0063] Example 1:

[0064] This embodiment describes a method for identifying side pillar reinforcement seats in open wagons. The system used in this method includes: a device installed above the railway wagon (see...). Figure 1 , 2The lidar 1, capable of scanning the carriage back and forth from the front to the rear, is connected to a digital processor. The digital processor can be an industrial control computer or other electronic device with digital processing and storage capabilities. The lidar can be mounted on a platform 2 that can move along the carriage's direction of travel. The platform is set on a guide rail aligned with the carriage's direction of travel, and the guide rail is mounted on a frame along the railway track. This frame forms the working area identified by the open wagon side pillar reinforcement seat 3. Figure 1 As shown. The steps of the method are as follows, and the principle framework is as follows. Figure 3 As shown:

[0065] Step 1, LiDAR data acquisition:

[0066] Establish a global coordinate system OXYZ and ensure that the radar coordinate system is aligned with it, such as... Figure 1 As shown, the origin O of the coordinate system is set at one end of the open wagon entering the work area, the X-axis is the direction of travel of the open wagon, the Y-axis is the left and right direction of the open wagon, and the Z-axis is the height direction of the open wagon. The hardware of the constructed lidar monitoring system includes a guide rail platform that moves along the X-axis and is installed on the upper part of the work area frame, and a 2D line lidar installed on the guide rail platform.

[0067] Data Acquisition Process: After the open vehicle stops in the work area, the guide rail platform drives the radar to move at a constant speed along the X-axis. In each acquisition cycle, it outputs a YZ plane depth data to form a three-dimensional point cloud matrix.

[0068] Step 2, Feature dataset construction:

[0069] LiDAR point cloud data is unstructured data and needs to be transformed into structured feature vectors through feature engineering to be compatible with support vector machine models. This method adopts a "local window + multi-dimensional feature" extraction strategy, balancing target feature capture with model computation efficiency.

[0070] Sub-step 201, Sliding window settings:

[0071] Considering that the side column reinforcement seats are distributed along the X-axis and have a certain width, with the X-axis as the sliding direction, the window parameters are set as follows:

[0072] Window size: 5 (includes point cloud data of the current X position and the two adjacent X positions before and after it);

[0073] Window step size: 1 (one window is taken for every X position moved to ensure feature continuity);

[0074] Window data format: Each window contains 5 sets of point cloud data, and each set contains all points in the YZ plane at the X position;

[0075] Sub-step 202, multi-dimensional feature extraction:

[0076] A 15-dimensional feature vector is extracted from each window, covering three categories of features: geometric, statistical, and location, as shown in the table below:

[0077] Table 1 Multidimensional Feature Extraction Table

[0078]

[0079] Referring to Table 1, the specific composition of the feature vector can be expressed as follows:

[0080]

[0081] Step 3, Labeling System Design:

[0082] Labeling is the core step in assigning a "feature-target" correspondence to a dataset. A scientific labeling system needs to be established and implemented through standardized processes to ensure the accuracy of the labels.

[0083] Sub-step 301, Category label:

[0084] definition: , where 1 indicates that the window contains the side column reinforcement seat, and 0 indicates that it does not contain (normal side wall, noise, frozen coal, etc.).

[0085] Judgment criteria: If all three of the following conditions are met, it is marked as 1:

[0086] Vertical transition segment length: Match the width feature, for example, set the length to: 190-210mm.

[0087] Maximum Z mutation value: Matching height features, for example, setting the height to 290-310mm.

[0088] Average reflection intensity: This value reflects the energy of the laser echo. The measured average reflection intensity of frozen and sticky coal is 30-80, and the average reflection intensity of the side column reinforcement seat is 120-180. Therefore, the average reflection intensity is set to 120-180. Under the premise of excluding low reflective objects such as frozen and sticky coal, the characteristics of high reflective objects such as side column reinforcement seats are identified.

[0089] Sub-step 302, standard result storage format:

[0090] The annotation results of each sample are stored in association with its feature vector. The complete sample format is as follows: ,in, .

[0091] Step 4, Support Vector Machine classification model construction and training:

[0092] Support Vector Machine (SVM) is a strongly non-linear classification algorithm with the advantages of "relying solely on support vectors and having strong generalization ability", making it suitable for small-sample, high-dimensional feature classification scenarios.

[0093] Sub-step 401, Dataset:

[0094] Input: Training dataset ,in This represents the sample size of the dataset.

[0095] Output: Separating hyperplane and classification decision function.

[0096] Sub-step 402, SVM optimization problem construction:

[0097] Select an appropriate kernel function and penalty function Construct and solve the convex quadratic programming problem:

[0098]

[0099] in: Using Lagrange multipliers, the optimal solution is obtained by solving the above convex quadratic programming problem. Choose the Gaussian kernel function:

[0100]

[0101] in: This is a hyperparameter.

[0102] Sub-step 403, Classification Decision Tree:

[0103] choose A component Meet the conditions ,calculate:

[0104]

[0105] in, This represents the optimal intercept of the hyperplane.

[0106] The corresponding SVM is a Gaussian radial basis function classifier. In this case, the classification decision function is:

[0107]

[0108] in: It is a mathematical symbol function.

[0109] Step 5, Building the Online Learning Framework: This includes the following sub-steps:

[0110] Sub-step 501, Model accuracy evaluation and update trigger:

[0111] Periodically use new calibration data as the test dataset. The test set contains 20-50 data entries. When an error occurs in the prediction result, the model is updated.

[0112] Sub-step 502, Implementation of the forgetting strategy: The forgetting strategy includes the following sub-steps:

[0113] Step 5021, Adding New Samples: Adding incorrectly predicted samples... The samples are transformed into new feature samples, labeled with the true labels, and then used as samples to be added.

[0114] Step 5022, Similarity Calculation: Calculate the similarity between the sample to be added and the current training set. Kernel function vectors for all samples:

[0115]

[0116] The larger the kernel function value, the higher the sample similarity;

[0117] Step 5023, Sample Set Update: Select The sample corresponding to the maximum value of the kernel function is from the current training dataset. Remove from the middle; remove new feature samples Add to training dataset In this process, the size of the training set is kept stable at a preset threshold to ensure computational efficiency and sample diversity;

[0118] Sub-step 503, iterative model training:

[0119] Retrain the SVM model using the updated training set.

[0120] The SVM model is retrained based on the updated training set, specifically by executing sub-steps 401 to 403 in step 4.

[0121] Step 6, Application: Use the trained model to perform actual recognition tasks, and continuously use the newly generated data from the tasks to continue training.

[0122] Application Example 1:

[0123] This embodiment was conducted in the frozen coal unloading workshop of a northern railway freight hub, focusing on the identification test of side pillar reinforcement seats for C80 type coal open wagons. The operating environment temperature was -15℃ to 5℃, with interference factors such as frozen coal adhesion and frost on the wagon surface. The test objective was to achieve early warning when the crushing device was 300mm away from the reinforcement seat, with an obstacle avoidance response time requirement of ≤50ms.

[0124] 1. Setup of the LiDAR data acquisition system:

[0125] (1) Hardware configuration:

[0126] LiDAR: The SICK LMS511-20100 2D line LiDAR is used, with a ranging range of 0.5-20m, ranging accuracy of ±2mm, scanning frequency of 50Hz, angular resolution of 0.1°, and output data format of Ethernet / IP.

[0127] Mobile platform: effective travel distance 18m, positioning accuracy ±1mm, moving speed 0.5m / s;

[0128] Coordinate system definition: The origin O is the front end of the carriage, the X-axis is along the length of the carriage, the Y-axis is along the width of the carriage, and the Z-axis is perpendicular to the ground.

[0129] (2) Data collection process:

[0130] After the open-top vehicle enters the work area and comes to a stop, the lidar moves along the X-axis, collecting one frame of YZ plane data every 20ms; each frame contains 2801 points, and the data frame range is Y∈[-1.4m,1.4m], Z∈[0.1-3.2m]; the original frame data is in the format of " "Formatted storage, approximately 1200 frames of data are collected per carriage;"

[0131] 2. Feature Dataset Construction Process

[0132] (1) Sliding window processing:

[0133] Window size: 5, including the current X position and two positions before and after it, namely X-20mm, X-10mm, X, X+10mm, X+20mm;

[0134] Window step size: 1, one window is taken every 10mm, and 1196 windows are generated in a single car;

[0135] Data cropping: Each window retains only the point cloud of the distribution area of ​​the side column reinforcement seats to reduce invalid calculations;

[0136] (2) Example of 15-dimensional feature vector extraction (taking a window at X=4.5m as an example):

[0137] Table 2 Examples of 15-dimensional feature vectors (X=4.5m)

[0138]

[0139] 3. Data annotation implementation

[0140] Category label: This window is labeled as 1 (including reinforcement seat) because the vertical transition section length is 203mm (within the range of 190-210mm), the maximum Z-curve value is 298mm (within the range of 290-310mm), and the average reflection intensity is 156 (within the range of 120-180).

[0141] Sample storage format:

[0142] [1256,89.2,1005,298,1.8,156,12.3,82,0.87,3,12,4500,203,85,15,1] (15-dimensional features + category label);

[0143] 4. SVM model training process:

[0144] (1) Dataset partitioning

[0145] Training set: Data from 20 C80 open wagons with different service years were collected, generating 23,920 samples (1,196 samples / wagon × 20), which were divided into a training set (19,136 samples) and a validation set (4,784 samples) at an 8:2 ratio.

[0146] Test set: 5 additional sections of open-top vehicle data that were not used in training, totaling 5980 samples;

[0147] (2) Model parameter optimization:

[0148] Kernel function: Choose the Gaussian radial basis kernel function;

[0149] Parameter optimization: Optimize the parameters on the validation set using a grid search method to obtain the optimal parameter combination. ;

[0150] Cross-validation: The average accuracy of 5-fold cross-validation was 93.7%, confirming the final model;

[0151] Decision function generation: Solving the dual problem yields 2856 support vectors (representing 14.9% of the training set). The decision function is then constructed.

[0152]

[0153] Among them, the bias term Lagrange multipliers It is distributed in the range of 0.002-0.987.

[0154] Application Example 2:

[0155] 1. Example of running an online learning framework:

[0156] (1) Real-time recognition process:

[0157] After the new open vehicle enters the work area, the lidar collects data and extracts features in real time; each window feature vector is generated and input into the SVM model to obtain the prediction result; when three consecutive windows predict 1, it is identified as a side pillar reinforcement seat, the position coordinates are output and the obstacle avoidance signal is triggered.

[0158] (2) Model update triggering instance:

[0159] In Section 37, the test open wagon suffered severe wear on its side pillar reinforcements, resulting in a top plane length of only 60mm. This caused five windows to be predicted incorrectly, with the actual value being 1 and the predicted value being 0. The online update mechanism was triggered, transforming these five erroneous samples into new feature samples, labeled as 1, because the transition surface features were preserved.

[0160] (3) Implementation of forgetting strategies:

[0161] Calculate the kernel function vector of the new sample and the 19136 samples in the training set. It was found that sample number 1856 had the largest kernel function value compared to the new sample. The sample was identified as the most similar, and the 1856th sample was removed. A new sample was added, the training set size remained unchanged, the model was retrained, the support vectors were updated to 2858, and the accuracy was restored to 92.5% after the decision function parameters were adjusted.

[0162] 2. Implementation Effectiveness Evaluation:

[0163] Recognition performance: In 50 test open wagons, the average recognition accuracy was 92.3%, with 96.5% accuracy for new wagons (service <1 year) and 88.7% accuracy for older wagons (service >5 years).

[0164] Real-time performance: Single-window feature extraction takes 7.2ms, SVM prediction takes 3.8ms, and the total processing latency is 11ms, which is less than the requirement of 50ms.

[0165] Robustness: With 30% frozen coal coverage, the recognition accuracy remains at 85.6%, and the accuracy decreases by ≤2% when the sensor drifts by ±5mm.

[0166] Online update results: After 10 online updates, the model's accuracy in identifying wear-resistant reinforcement seats improved from the initial 76.3% to 89.1%.

[0167] This embodiment verifies the effectiveness of the method in actual industrial scenarios. Through precise acquisition by lidar, multi-dimensional feature extraction, SVM classification, and online learning mechanism, it achieves real-time and accurate identification of side column reinforcement seats, meeting the obstacle avoidance requirements of crushing equipment.

[0168] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred arrangement, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solution of the present invention (such as the type of carriage, the application of various formulas, the order of steps, etc.) without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for constructing an open wagon side post reinforcement seat recognition, characterized in that, The system used by the method comprises a laser radar arranged above a carriage of a railway and capable of reciprocally scanning the carriage from head to tail, the laser radar being connected with a digital processor, and the steps of the method are as follows: Step 1, laser radar data acquisition: A global coordinate system O-X-Y-Z is established to ensure that the radar coordinate system is aligned with the global coordinate system, wherein the coordinate system origin O is arranged at one end of the open wagon entering the working area, the X axis is the open wagon running direction, the Y axis is the left-right direction of the open wagon, and the Z axis is the height direction of the open wagon; The collection process is as follows: after the open wagon is stabilized in the working area, the guide rail platform drives the radar to move uniformly along the X axis, and a Y-Z plane depth data is output in each collection cycle to form a three-dimensional point cloud matrix; Step 2, feature data set construction: Substep 201, sliding window setting: taking the X axis as the sliding direction, the window parameters are set as follows: window size: 5; window step: 1; window data format: each window contains 5 groups of point cloud data, and each group contains all points of the Y-Z plane at the X position; Substep 202, multi-dimensional feature extraction: 15-dimensional feature vector is extracted from each window , covering three types of features: geometry, statistics, and location The geometric feature vector includes: β1, β2, β3, β4, β5; Wherein: β1 is the Z value mean, the extraction method is: the average value of all point Z coordinates in the window; β2 is the Z value variance, the extraction method is: the variance of all point Z coordinates in the window; β3 is the Y value median, the extraction method is: the median of all point Y coordinates in the window; β4 is the maximum Z mutation value, the extraction method is: the maximum value of the Z value difference between adjacent X positions; β5 is the flatness, the extraction method is: the deviation value of the plane fitted by the top point set; The statistical feature vector includes: β6, β7, β8, β9; Wherein: β6 is the reflection intensity mean, the extraction method is: the average value of all point reflection intensity in the window; β7 is the reflection intensity standard deviation; β8 is the point density, the extraction method is: the number of points in a unit volume; β9 is the effective point ratio, the extraction method is: the ratio of the number of effective points to the total number of points in the window; The position feature vector includes: β 10 , β 11 , β 12 , β 13 , β 14 , β 15 ; Where: β 10 is the X-axis deviation from the theoretical position, the extraction method is: the difference between the current window X coordinate and the theoretical side column reinforcement seat position; β 11 is the Z-axis deviation from the reference line, the extraction method is: the difference between the current window Z mean value and the reference line; β 12 is the window center X coordinate, the extraction method is: the average value of the window 5 X positions; β 13 is the vertical transition length, the extraction method is: the span of the point set with Z value mutation exceeding 50 mm in X axis; β 14 is the top plane length, the extraction method is: the span of the point set with Z value variance <5 mm² in X axis; β 15 is the reflection intensity gradient, the extraction method is: the difference between the front and rear X position reflection intensity in the window; Step 3, label system design: Substep 301, class label: Definitions: where 1 indicates that the window contains a jamb reinforcing seat within, and 0 indicates that it does not. Determination basis: if the following three conditions are met at the same time, it is marked as 1: Vertical transition segment length: matching width feature; Maximum Z mutation value: matching height feature; Reflection intensity mean: matching material feature; Substep 302, standard result storage format: The annotation result of each sample is stored in association with the feature vector, and a single sample The standard storage format of a single sample is composed of a feature vector and an annotation result, specifically , wherein x is the feature vector; is the annotation result. Step 4, support vector machine classification model construction and training: Substep 401, data set: Input: training dataset consists of samples, wherein, , is the sample capacity of the dataset, is the sample, is the feature vector of the sample, is the label of the sample; Output: separation hyperplane and classification decision function; Substep 402, vector machine optimization problem construction: Selecting an appropriate kernel function and a penalty function , constructing and solving a convex quadratic programming problem: wherein is the Lagrange multiplier corresponding to the sample, and the optimal solution is obtained by solving the above convex quadratic programming problem T is the transpose symbol; the kernel function is selected as a Gaussian kernel function: wherein is a hyperparameter; Substep 403, classification decision tree: selecting one component satisfying a condition , calculating wherein, represents the optimal intercept of the hyperplane; The corresponding vector machine is a Gaussian radial basis function classifier, in which case the classification decision function is: wherein is the mathematical sign function; Step 5, online learning framework construction: Substep 501, model accuracy evaluation and update triggering: Periodically use new calibration data as test data set Test set data capacity size is 20-50, when prediction result error occurs, trigger model update Substep 502, forgetting strategy implementation: In order to avoid the calculation amount increasing sharply due to the unlimited expansion of the training set size, a forgetting strategy based on kernel function similarity is introduced, which is divided into the following steps: Step 5021, new sample addition: samples with prediction errors are converted into new feature samples, labeled with real labels, and added as to-be-added samples; Step 5022, similarity calculation: calculate the kernel function vector of the sample to be added and all samples in the current training set ​ The larger the kernel function value, the higher the sample similarity; Step 5023, sample set update: select the sample corresponding to the maximum value of the kernel function from the current training data set ; remove; add the new feature sample to the training data set , keep the training set size stable at a preset threshold, and ensure the calculation efficiency and sample diversity Substep 503, model iterative training: retrain the SVM model based on the updated training set; Step 6, Application: Actual recognition operation is performed with the trained model, and the acquired data is continuously trained.

Images