振动筛作为选煤厂的核心装备,其运行环境异常恶劣且多变,获取标记的状态数据耗费时间和精力,传统方法在振筛机的运动状态识别以及泛化性方面存在不足。针对这一问题,提出了一种基于自监督记忆自适应的振筛机状态识别方法。首先,引入最大熵原理和多种数据变换方式对原始信号进行多样化分布增强。借助编码器,将处理后的特征传递至自监督学习模块和记忆存储模块。在自监督学习模块中,分别对经过多种变换的数据进行分类以及对经过掩码处理的数据进行重构,以获得更加丰富的特征表达。与此同时,通过信息熵和相似度技术,在记忆块中选取最具代表性特征项,将其与编码器提取的特征进行自适应融合。最终使用状态分类器对振筛机的运动状态进行分类。实验验证结果表明,与其他先进方法相比,该方法显著提升了状态识别的准确率,F1分数超过对比方法约23.92%。

The vibrating screen machine, as a core equipment in coal preparation plants, operates under harsh and dynamic conditions. Acquiring labeled status data consumes considerable time and effort, and traditional methods exhibit shortcomings in recognizing the motion states of vibrating screens and in achieving generalization. In response to this issue, we propose a self-supervised memory-adaptive method for recognizing the state of vibrating screens. Firstly, we enhance the diversity of the original signals by applying the principle of maximum entropy and various data transformation techniques. Subsequently, utilizing an encoder, we transmit the processed features to both a self - supervised learning module and a memory storage module. Within the self - supervised learning module, data subjected to multiple transformations are classified, and data subjected to masked processing is reconstructed, yielding more enriched feature representations. Concurrently, by employing information entropy and similarity techniques, representative feature items are selected within memory blocks and adaptively fused with features extracted by the encoder. Ultimately, a state classifier is employed to classify the motion states of vibrating screens. The proposed method is validated on a vibrating screen dataset, demonstrating significant enhancement in state recognition accuracy compared to other advanced methods. Specifically, on the vibrating screen dataset, the F1 score of our method surpasses the comparative approach by approximately 23. 92%.