基于卷积神经网络的拖拉机工况识别

农机工况识别在细化农机作业状态和帮助掌握区域污染物排放趋势方面有着重要的研究价值。基于拖拉机不同运行状态下的行驶速度、发动机转速以及实时油耗等时间序列,首次提出将图像识别方法引入到拖拉机工况识别中的思路,并分别应用参数优化的支持向量机与卷积神经网络对实际作业拖拉机工况进行研究。结果表明:(1)基于参数优化的支持向量机可以较好地实现样本点的工况识别且识别准确度达到99.851 9%,但无法实现农机工况的连续性识别,同时无法对农机工况转换阶段进行有效识别。(2)以拖拉机运行速度与发动机转速等信息构建样本图像来描述农机工况变化的数据表达,并在此基础上应用卷积神经网络可以有效实现农机工况的连续性识别,且识别准确率可以达到93.3%。本研究在农机工况识别方面具有一定参考价值,并为后续农机不同工况下区域污染物排放研究提供技术支持。     

镇江长江豚类省级自然保护区水域生态修复效果的监测与评价

自2006年以来,长江豚类极度濒危的生存状况得到广泛关注,在其自然保护区实施一系列生态修复工程,为镇江长江豚类省级自然保护区内的江豚提供良好的生存环境。于2015—2018年在和畅洲水域建设生态浮岛、人工鱼巢和放流底栖动物进行水域生态修复,对生态修复工程进行跟踪监测和效果评价。结果表明：生态浮岛的建设可提高区域内浮游植物和浮游动物的生物多样性和群落结构复杂程度;人工鱼巢对产黏性卵鱼类具有一定增殖效果,增殖种类主要为鲤、鲫、䱗和黄尾鲴;生态浮岛和人工鱼巢的建设对鱼类均有较好的聚集效果;底栖动物增殖放流显著增加了放流物种的丰度、生物量及保护区底栖动物群落生物量,增殖放流效果评价为中等。该自然保护区水域生态修复效果明显,为工程施工影响下水生生物保护区生态修复提供科学依据。     

基于BP神经网络的柴油发动机特性建模

建立连续的发动机燃油特性和调速特性数学模型作为液压机械无级变速器虚拟试验平台的动力源。根据虚拟平台对不同特性区域的精度需求对柴油发动机不同特性区域的试验数据进行不同的密度选取、乱序和归一化处理,采用单隐层BP神经网络对试验数据进行训练,对比不同隐层节点数网络的训练误差和测试误差,选取误差最小的网络,求解出网络的数学表达式。通过该方法以ISLe310柴油发动机为例建立燃油特性和调速特性的连续数学模型,这两个简单的数学表达式准确反映了发动机万有特性和外特性,连续模型避免了虚拟试验中出现信号的突变和奇异点。通过和经典的最小二乘法拟合得到的最优特性模型进行对比,其具有更小的误差、更强的泛化能力,能够更好地反映柴油发动机的相关特性。     

基于声振信号对称极坐标图像的苹果霉心病早期检测

为实现苹果早期霉心病较高精度的检测，该研究采用对称极坐标法（Symmetrized Dot Pattern，SDP）将苹果声振信号变换为雪花图，然后采用AlexNet、VGG16和ResNet50卷积神经网络以迁移学习方式深度挖掘SDP雪花图像的特征信息，将其输入到支持向量机（Support Vector Machine，SVM）分类器，对霉心程度≤7%的苹果进行检测。研究结果表明，当时间间隔系数为25和角度放大因子为50°时，健康果与早期霉心果声振信号的SDP图形状特征差异最大，在此条件下获取的SDP图经卷积神经网络AlexNet、VGG16和ResNet50提取特征并构建了不同核函数的SVM霉心果检测模型，在各类SVM模型中，ResNet50-SVM-gaus（高斯基）模型用相对较少的训练时间和参数量可取得训练集霉心果较高分类准确率，经超参数优化训练该模型对健康果和早期霉心果测试集不平衡样本（10∶1）的总体分类准确率达到96.97%，平均查准率、平均查全率、平均加权调和均值、Kappa系数和马修斯相关系数值分别为80.19%、90.36%、86.21%，82.54%和82.68%，该模型不仅对多数类的健康果保持较高分类准确率，而且对少数类的早期霉心果也具有较高判别能力。这些研究结果为声振法应用于果蔬内部病害的早期在线检测系统研发提供了技术支撑。     

基于原型网络的小样本禽蛋图像特征检测方法

机器视觉因具有检测速度快、稳定性高及成本低等优点，已发展成为禽蛋无损检测领域主流检测手段。使用该技术对禽蛋进行无损检测时，需要依赖大量禽蛋图像作为数据支撑才能取得较好的检测效果。由于养殖安全等限制，禽蛋图像数据的采集成本较高，针对该问题，提出了一种适应于小样本禽蛋图像检测的原型网络（Prototypical network）。该网络利用引入注意力机制的逆残差结构搭建的卷积神经网络将不同类别的禽蛋图像映射至嵌入空间，并利用欧氏距离度量测试禽蛋图像在嵌入空间的类别，从而完成禽蛋图像的分类。本文利用该网络分别验证了小样本条件下受精蛋与无精蛋、双黄蛋与单黄蛋及裂纹蛋与正常蛋的分类检测效果，其检测精度分别为95%、98%、88%。试验结果表明本文方法能够有效地解决禽蛋图像检测中样本不足的问题，为禽蛋图像无损检测研究提供了新的思路。     

文本挖掘下的话语认知与构建——以人工智能和5G领域白宫新闻文本为例

[目的/意义]政府新闻文本是政策文本的表达形式之一。对政府新闻文本进行挖掘分析,能直观的展现表面动态,揭示对话事件的立场及其背后关系,能对情报分析工作、中国新型智库的发展建设,提供有效的助力。[方法/过程]文章利用TextRank算法进行核心句提取,并结合批评话语分析理论框架来做评估,以语料驱动从关键句、关键词两个宏观角度讨论政府决策态度,以词性从话语策略的微观角度总结美国官方对人工智能和5G领域的话语认知形象。[结果/结论]通过分析可以看出,美国政府对于人工智能和5G领域的态度将持续保持在“政府引领下的美国需优先”这一局面,面对发展中国家的科技竞争,增加政府政策性文本的指导无疑是最好的对策建议。     

基于卷积神经网络的玉米病害识别方法研究

为解决传统的玉米病害识别方法中特征提取主观性强及误识率高的问题,提出利用卷积神经网络对玉米病害进行识别。以玉米病害图像和健康图像共5种类别的玉米图像为研究对象,并采用LeNet模型进行试验。首先,按照8∶2的比例为每种玉米病害图像选择训练集和测试集。然后,通过试验组合和对比分析的方法比较不同卷积神经网络结构设置对准确率的影响,选出最佳参数。另外,选用Adam算法代替SGD算法来优化模型,通过指数衰减法调整学习率,将L2正则项添加到交叉熵函数中,并选择Dropout策略和ReLU激励函数。最后,确定了一个10层CNN网络结构。试验结果显示,玉米花叶病、灰斑病、锈病、叶斑病和玉米健康识别率分别为95.83%、90.57%、100%、93.75%、100%,平均识别率达96%,平均计算时间为0.15 s。经试验结果比较,该模型识别效果明显高于传统方法,为玉米病害的防治提供技术支持。     

Evidence of cormorant‐induced mortality,disparate migration strategies and repeatable circadian rhythm in the endangered North Sea houting (Coregonus oxyrinchus): A telemetry study mapping the postspawning migration

Life history theory predicts a trade‐off between migration and residency where migration is favoured when it infers elevated fitness. Although migration to more favourable environments may offer higher growth rates, migrants often experience increased mortality due to predation. Here, we investigated mortality and migration behaviour of the North Sea houting (Coregonus oxyrinchus), an anadromous salmonid endemic to the Wadden Sea. We used acoustic telemetry to map the migration of the only remaining indigenous population by applying stationary hydrophones combined with manual tracking. Data suggested a total mortality of 26%, with 30% of the total mortality attributed to predation by great cormorants (Phalacrocorax carbo sinensis), highlighting that North Sea houting conservation could be jeopardised by increased cormorant predation. Risk of cormorant predation was size‐dependent, with smaller fish suffering higher risk of predation. The study found North Sea houting to exhibit disparate migration strategies and identified a lentic area in the estuary as an important habitat. Two newly established artificial lakes within the river system significantly reduced the migration speeds, possibly indicating constrained navigation through the lakes. The migration into the Wadden Sea correlated with temperature perhaps indicating osmoregulatory constraints of sea entry. Unlike many salmonid species, migration occurred both day and night. Moreover, fish exhibited repeatable individual differences in diel activity patterns, suggesting that individuals differ consistently in their migratory activity throughout the 24‐hr period. Our study provides novel information on salmonid migration, which is crucial for the development of science‐based conservation strategies.     

A comparison of the effects of carbon dioxide and medical air for abdominal insufflation on respiratory parameters in xylazine-sedated sheep undergoing laparoscopic artificial insemination

AIMS: To determine if abdominal insufflation with medical air will improve oxygenation and ventilation parameters when compared to insufflation with CO₂ in xylazine-sedated sheep undergoing laparoscopic artificial insemination (AI).

METHODS: Forty-seven sheep underwent oestrus synchronisation and were fasted for 24 hours prior to laparoscopic AI. Each animal was randomised to receive either CO₂ or medical air for abdominal insufflation. An auricular arterial catheter was placed and utilised for serial blood sampling. Respiratory rates (RR) and arterial blood samples were collected at baseline, after xylazine (0.1?mg/kg I/V) sedation, 2 minutes after Trendelenburg positioning, 5 minutes after abdominal insufflation, and 10 minutes after being returned to a standing position. Blood samples were collected in heparinised syringes, stored on ice, and analysed for arterial pH, partial pressure of arterial O₂ (PaO₂), and CO₂ (PaCO₂). The number of ewes conceiving to AI was also determined.

RESULTS: Repeated measures ANOVA demonstrated temporal effects on RR, PaO₂, PaCO₂ and arterial pH during the laparoscopic AI procedure (p<0.001), but no difference between insufflation groups (p>0.01). No sheep experienced hypercapnia (PaCO₂>50?mmHg) or acidaemia (pH<7.35). Hypoxaemia (PaO₂<70?mmHg) was diagnosed during the procedure in 14/22 (64%) ewes in the CO₂ group compared with 8/23 (35%) ewes in the medical air group (p=0.053). Overall, 15/20 (75%) ewes in the CO₂ group conceived to AI compared with 16/22 (72.7%) in the medical air group (p=0.867).

CONCLUSIONS AND CLINICAL RELEVANCE: There were no statistical or clinical differences in RR, PaO₂, PaCO₂, pH, or conception to AI when comparing the effects of CO₂ and medical air as abdominal insufflation gases. None of the sheep experienced hypercapnia or acidaemic, yet 42% (19/45) of sheep developed clinical hypoxaemia, with a higher percentage of ewes in the CO₂ group developing hypoxaemia than in the medical air group. Based on the overall analysis, medical air could be utilised as a comparable alternative for abdominal insufflation during laparoscopic AI procedures.