np.nonzero, np.argwhere, np.where, np.squeeze用法小结

import numpy as np import pywt import matplotlib.pyplot as plt import scipy.io as scio from sklearn.cluster import DBSCAN from sklearn.manifold import TSNE from sklearn.preprocessing import StandardScaler from scipy.stats import entropy from sklearn.svm import OneClassSVM import joblib def sliding_window(sequence, window_size, step_size): """ 对一维序列进行滑窗切分。 参数： sequence (array-like): 输入的一维序列。 window_size (int): 滑窗的窗口大小。 step_size (int): 滑窗的步长。 返回： np.ndarray: 切分后的窗口数组，每一行是一个窗口。 """ sequence = np.asarray(sequence) n = len(sequence) windows = [ sequence[i:i + window_size] for i in range(0, n - window_size + 1, step_size) ] return np.array(windows) if __name__ == "__main__": # 1. 载入数据 sliding_length, sliding_step = 16384, 16384 dataFile = r'.\subei GGZQ-DIS-G11-001-05.mat' dataTemp = scio.loadmat(dataFile) dataSet = np.array(dataTemp['mm']) dataSet = np.squeeze(dataSet) print(dataSet.shape) signals = sliding_window(dataSet, sliding_length, sliding_step) print(signals.shape) feature_matrix = [] for ii in range(signals.shape[0]): # 小波特征 coeffs = pywt.swt(signals[ii, :], wavelet='db1', level=4, trim_approx=True, norm=True) wavelet_features = np.var(np.array(coeffs)[1:, :], axis=1) # 新增的4个特征 window = signals[ii, :] # 差值 diff = window[:-1] - window[1:] diff_var = np.var(diff) # 极值 extremes = np.max(window) - np.min(window) # 均值 mean_value = np.mean(window) # 交叉熵 hist, _ = np.histogram(window, bins=10) cross_entropy = entropy(hist / hist.sum()) new_features = [diff_var, extremes, mean_value, cross_entropy] combined_features = np.concatenate((wavelet_features, new_features)) feature_matrix.append(combined_features) X = feature_matrix = np.vstack(feature_matrix) # 2. 特征标准化 scaler = StandardScaler() X_scaled = scaler.fit_transform(X) # 3.3 应用 DBSCAN dbscan = DBSCAN(eps=2, min_samples=5) labels = dbscan.fit_predict(X_scaled) # 4. 失真异常定位 anomaly_indices = np.squeeze(np.array(np.nonzero(labels))) anomaly_indices_start = [] anomaly_indices_end = [] for ii in range(anomaly_indices.shape[0]): anomaly_indices_start.append(anomaly_indices[ii] * sliding_step) anomaly_indices_end.append(anomaly_indices[ii] * sliding_step + sliding_length) anomaly_indices_start = np.squeeze(np.array(anomaly_indices_start)) anomaly_indices_end = np.squeeze(np.array(anomaly_indices_end)) # 5. 使用 t-SNE 降维 # 获取样本数量 n_samples = X_scaled.shape[0] # 确保 perplexity 小于样本数量 perplexity = min(30, n_samples - 1) tsne = TSNE(n_components=2, perplexity=perplexity) X_tsne = tsne.fit_transform(X_scaled) # 6. 使用 OC-SVM 进行训练 # 提取 DBSCAN 识别的正常数据（排除噪声点） normal_indices = np.where(labels != -1)[0] X_normal = X_scaled[normal_indices] oc_svm = OneClassSVM(nu=0.05, kernel="rbf", gamma=0.1) oc_svm.fit(X_normal) # 使用正常数据训练 OC-SVM # 7. 保存 OC-SVM 模型和标准化器 joblib.dump(oc_svm, 'oc_svm_model2.pkl') joblib.dump(scaler, 'scaler.pkl') # 保存标准化器 # 8. 可视化 t-SNE 降维聚类结果及失真定位结果 plt.figure(figsize=(8, 6)) plt.scatter(X_tsne[:, 0], X_tsne[:, 1], c=labels, cmap='viridis', s=50) plt.title("t-SNE Visualization based on Clustering Results") plt.xlabel("Feature 1") plt.ylabel("Feature 2") plt.colorbar(label="Cluster Label") plt.show() plt.figure(11) plt.plot(dataSet) print(anomaly_indices.shape[0]) for ii in range(anomaly_indices.shape[0]): plt.axvspan(anomaly_indices_start[ii], anomaly_indices_end[ii], facecolor='red', alpha=0.3) plt.show() 在我的代码中我通过对输入数据每个窗口进行8维特征提取，用OC-SVM训练得到了一个决策边界oc_svm_model2.pkl文件，单这个边界是通过8维特征训练得到的，我现在想用T-SNE对这个边界进行降维为2维度，以得到可视化的决策边界，请你整理思路并为我输出可行的代码，输出的代码需要你先运行直到不报错为止

将下列numpy代码转化为纯pytorch代码： w = np.maximum(0.0, xx2 - xx1 + 1) h = np.maximum(0.0, yy2 - yy1 + 1) inter = w * h ovr = inter / (areas[i] + areas[order[1:]] - inter) inds = np.where(ovr <= thresh)[0] order = order[inds + 1]