用MATLAB打造AI异常检测的“黑神话·钟馗，聊聊工业场景的实战经验 ... ... ... ...

Matlab
%% 激光雷达点云雨天异常检测
clear all; clc; close all;
rng(2025);
% 1.1 护栏
n_guard = 3000;
pc_guard = [linspace(0, 50, n_guard)', ones(n_guard, 1)*3, linspace(0, 1.2, n_guard)'];
color_guard = repmat([1,0,0], n_guard, 1);
% 1.2 路灯
pc_light = [];
color_light = [];
light_pos = [10,5,6; 25,6,5; 40,4.5,7];
for i = 1:3
single_light = [repmat(light_pos(i,1), 50, 1), repmat(light_pos(i,2), 50, 1), linspace(0, light_pos(i,3), 50)'];
pc_light = [pc_light; single_light];
color_light = [color_light; repmat([0,1,0], 50, 1)];
end
% 1.3 雨滴噪声
n_noise = 8000;
pc_noise = [rand(n_noise, 1)*50, rand(n_noise, 1)*13-5, rand(n_noise, 1)*8];
color_noise = repmat([0.5,0.5,0.5], n_noise, 1);
% 1.4 合并点云
pc_raw = [pc_guard; pc_light; pc_noise];
color_raw = [color_guard; color_light; color_noise];
total_core = size(pc_guard,1) + size(pc_light,1);

matlab
% 激光雷达点云异常检测
function lidar_anomaly_detection()
fprintf('=== 激光雷达点云异常检测 ===\n');
fprintf('版本: 基础MATLAB兼容版\n');
fprintf('日期: %s\n', datestr(now));
fprintf('====================================\n');
% 生成或加载数据
fprintf('1. 准备数据...\n');
[normal_features, rain_features] = prepare_data();
% 数据预处理
fprintf('2. 数据预处理...\n');
[X_train, X_test, mu, sigma] = preprocess_data(normal_features, rain_features);
% 使用PCA进行异常检测
fprintf('3. 训练PCA异常检测模型...\n');
[pca_model, train_recon_error, test_recon_error] = train_pca_anomaly_detection(X_train, X_test);
% 评估模型
fprintf('4. 模型评估...\n');
evaluate_model(X_train, X_test, train_recon_error, test_recon_error, pca_model);
% 可视化结果
fprintf('5. 生成可视化结果...\n');
create_visualizations(X_train, X_test, train_recon_error, test_recon_error, pca_model);
fprintf('处理完成！\n');
end

matlab
% 计算正常/雨天数据的重构误差
normal_recon = predict(autoenc, normal_features);
normal_err = mean((normal_recon - normal_features).^2, 2);
rain_recon = predict(autoenc, rain_features);
rain_err = mean((rain_recon - rain_features).^2, 2);% 设阈值：取正常误差的95分位数（平衡漏检和误检）
threshold = prctile(normal_err, 95);% 画误差分布直方图
figure('Name', '激光雷达点云异常检测结果');
histogram(normal_err, 'BinWidth', 0.01, 'DisplayName', '正常点云'); 
hold on;histogram(rain_err, 'BinWidth', 0.01, 'DisplayName', '雨天遮挡点云');
xline(threshold, 'r--', ['异常阈值：', num2str(threshold, '%.4f')]);
legend; title('重构误差分布');
saveas(gcf, 'lidar_anomaly_result.png');

matlab
% 从Simulink导出正常数据（转速、扭矩）
normal_data = [ship_data.speed, ship_data.torque];
% 训练孤立森林：100棵树，异常比例设5%
iforest = IsolationForest.fit(normal_data, 'NumTrees', 100, 'Contamination', 0.05);
% 模拟异常数据：第400-500步转速降30%
anomaly_data = normal_data;
anomaly_data(400:500, 1) = anomaly_data(400:500, 1) * 0.7;
% 检测异常并生成结果GIF
[is_anomaly, anomaly_score] = predict(iforest, [normal_data; anomaly_data]);
createAnomalyGIF([normal_data; anomaly_data], is_anomaly, 'ship_anomaly.gif');

matlab
% 加载我们采集的盾构机数据（正常+轻微磨损+严重磨损）
load('shield_vibration_data.mat'); 
% 提取时频特征：时域（均方根、峰峰值）+频域（频谱重心）
feature_extractor = signalTimeFrequencyFeatureExtractor(...
    'SampleRate', 10000, ...
    'TimeFeatures', {'RMS', 'PeakToPeak'}, ...
    'FrequencyFeatures', {'SpectralCentroid'});
% 生成特征矩阵
normal_feat = extractFeatures(feature_extractor, normal_vib);
light_feat = extractFeatures(feature_extractor, light_wear_vib);
severe_feat = extractFeatures(feature_extractor, severe_wear_vib);
% 画振动信号对比图
figure('Name', '刀盘振动信号对比');
subplot(3,1,1); plot(normal_vib(1:5000)); title('正常振动');
subplot(3,1,2); plot(light_wear_vib(1:5000)); title('轻微磨损振动');
subplot(3,1,3); plot(severe_wear_vib(1:5000)); title('严重磨损振动');
saveas(gcf, 'shield_vib_compare.png');

Matlab
% 步骤 2：CNN + 解释生成组合模型训练
% 1. 训练CNN提取振动特征
cnn_layers = [
    sequenceInputLayer(size(normal_feat,2))
    convolution1dLayer(10, 32, 'Padding', 'same')
    batchNormalizationLayer()
    reluLayer()
    globalAveragePooling1dLayer()
    fullyConnectedLayer(3) % 3类：正常/轻微/严重
    softmaxLayer()
    classificationLayer()];
% 训练参数
options = trainingOptions('adam', ...
    'MaxEpochs', 50, ...
    'Plots', 'training-progress', ...
    'UseGPU', 'auto', ...
    'OutputFcn', @(info) saveTrainingGIF(info, 'shield_cnn_train.gif'));
% 准备训练数据
train_data = [normal_feat; light_feat; severe_feat];
train_labels = categorical([repmat("normal",size(normal_feat,1),1); ...
    repmat("light",size(light_feat,1),1); ...
    repmat("severe",size(severe_feat,1),1)]);
cnn_model = trainNetwork(train_data, train_labels, cnn_layers, options);
% 2. 使用增强解释生成
fprintf('训练完成，开始生成解释...\n');
% 用测试数据验证
test_feat = extractFeatures(feature_extractor, test_vib);
[pred_label, score] = classify(cnn_model, test_feat);
confidence = max(score);
% 生成解释
if exist('api_key', 'var') && ~isempty(api_key)
    % 使用外部API（需要提前设置api_key）
    prompt = "根据振动特征，解释为什么判定为" + string(pred_label) + "，100字内";
    explanation = callLLMAPI(prompt, api_key);
else
    % 使用本地解释生成
    explanation = generateRuleBasedExplanation(test_feat, pred_label, confidence);
end
% 可视化结果
figure('Name', '盾构机刀盘检测结果');
subplot(2,1,1); 
plot(test_vib(1:5000)); 
title("测试振动信号");
xlabel('采样点'); ylabel('振幅');
subplot(2,1,2); 
text(0.1, 0.5, explanation, 'FontSize', 12, 'VerticalAlignment', 'middle');
axis off;
title("AI解释结果");
saveas(gcf, 'shield_detection_result.png');
fprintf('检测完成！结果：%s (置信度: %.1f%%)\n', string(pred_label), confidence*100);
fprintf('解释: %s\n', explanation);

matlab
% 用测试数据检测
test_feat = extractFeatures(feature_extractor, test_vib);
[pred_label, score] = classify(cnn_model, test_feat);
% 让LLM生成解释
prompt = "根据振动特征，解释为什么判定为" + string(pred_label) + "，100字内";
explanation = generate(llm, prompt);
% 可视化结果
figure('Name', '盾构机刀盘检测结果');
subplot(2,1,1); plot(test_vib(1:5000)); title("测试振动信号");
subplot(2,1,2); text(0.1,0.5, explanation, 'FontSize',10); title("LLM解释");
saveas(gcf, 'shield_detection_result.png');