基于深度并行時序網(wǎng)絡的用戶側異常數(shù)據(jù)智能診斷

摘" 要： 在用戶側數(shù)據(jù)中，異常往往隱藏在復雜的時序關系中，傳統(tǒng)的時序分析方法在處理用戶側數(shù)據(jù)中復雜的時序關系時存在困難，特征提取難以捕獲關鍵特征，導致診斷精度低且易漏檢。為此，研究一種基于深度并行時序網(wǎng)絡的用戶側異常數(shù)據(jù)智能診斷方法。深度并行時序網(wǎng)絡分解層利用滑動窗口法分割用戶側數(shù)據(jù)，得到數(shù)個窗口序列。編碼層依據(jù)層疊時序卷積神經(jīng)網(wǎng)絡與長短期記憶（LSTM）網(wǎng)絡建立編碼器，提取各窗口序列的時空特征；解碼層通過引入時間注意力機制的門控循環(huán)單元建立解碼器，重構窗口序列的時空特征；推斷層依據(jù)重構特征計算異常分數(shù)，當異常分數(shù)大于設置閾值時，說明該窗口內(nèi)的用戶側數(shù)據(jù)為異常數(shù)據(jù)，即完成了用戶側異常數(shù)據(jù)的智能診斷。實驗結果表明，所提方法可有效提取用戶側數(shù)據(jù)特征，計算異常分數(shù)，并完成用戶側異常數(shù)據(jù)智能診斷。

關鍵詞： 深度并行時序網(wǎng)絡； 用戶側； 異常數(shù)據(jù)； 智能診斷； 滑動窗口； LSTM

中圖分類號： TN929?34； TP391" " " " " " " " " " 文獻標識碼： A" " " " " " " " " " "文章編號： 1004?373X（2025）04?0140?05

User?side abnormal data intelligent diagnosis based on deep parallel temporal network

ZHENG Yansong， LIAO Weiguo

（School of Artificial Intelligence， Zhujiang College， South China Agricultural University， Guangzhou 510980， China）

Abstract： In user?side data， anomalies are often hidden in complex temporal relationships. The traditional temporal analysis methods face difficulties in handling complex temporal relationships in user?side data， and feature extraction is difficult to capture key features， resulting in low diagnostic accuracy and easy missed detections. To this end， a user?side abnormal data intelligent diagnosis method based on deep parallel temporal networks is studied. In the deep parallel temporal network decomposition layer， the sliding window method is used to segment user?side data and obtain several window sequences. In the encoding layer， an encoder can be established based on stacked temporal convolutional neural networks and long short?term memory networks （LSTM） to extract the spatiotemporal features of each window sequence. In the decoding layer， a decoder is established and the spatiotemporal features of the window sequence are constructed by introducing a gated loop unit with temporal attention mechanism. In the inference layer， the anomaly score is calculated based on the reconstructed features. When the anomaly score exceeds the set threshold， it indicates that the user?side data in the window is abnormal data， completing the intelligent diagnosis of user?side anomaly data. The experimental results show that this method can effectively extract features from user?side data， calculate anomaly scores， and realize the intelligent diagnosis of user?side anomaly data.

Keywords： deep parallel temporal network; user side; abnormal data; intelligent diagnosis; sliding window; long short?term memory network

0" 引" 言

用戶側數(shù)據(jù)中蘊含著豐富的信息，對于企業(yè)的決策支持、業(yè)務優(yōu)化以及用戶服務提升等方面具有重要意義[1]。然而，在實際應用中，由于各種原因（如設備故障、網(wǎng)絡波動、用戶誤操作等）用戶側數(shù)據(jù)往往會出現(xiàn)異常[2]，這些異常數(shù)據(jù)不僅會影響數(shù)據(jù)的質(zhì)量和可靠性，還可能誤導決策，給企業(yè)帶來損失。用戶側數(shù)據(jù)中的異常往往隱藏在復雜的時序關系中，傳統(tǒng)的時序分析方法無法有效捕獲這些關系中的關鍵特征，在處理高維、非線性或動態(tài)變化的數(shù)據(jù)時容易導致診斷精度低和漏檢異常。……