修改filter 参数

src/main/java/cn/mesalab/config/ApplicationConfig.java

@@ -44,8 +44,10 @@ public class ApplicationConfig {
     public static final String HBASE_ZOOKEEPER_CLIENT_PORT= ConfigUtils.getStringProperty("hbase.zookeeper.client.port");
 
 
-    public static final Double BASELINE_KALMAN_P = ConfigUtils.getDoubleProperty("baseline.kalman.p");
+    public static final Double BASELINE_KALMAN_Q = ConfigUtils.getDoubleProperty("baseline.kalman.q");
     public static final Double BASELINE_KALMAN_R = ConfigUtils.getDoubleProperty("baseline.kalman.r");
+    public static final Double BASELINE_KALMAN_P = ConfigUtils.getDoubleProperty("baseline.kalman.p");
+    public static final Double BASELINE_KALMAN_M = ConfigUtils.getDoubleProperty("baseline.kalman.m");
 
     public static final Integer BASELINE_GENERATE_BATCH_SIZE = ConfigUtils.getIntProperty("baseline.generate.batch.size");
     public static final Long DRUID_READ_BATCH_TIME_GRAD_HOUR = ConfigUtils.getLongProperty("druid.read.batch.time.grad.hour");

src/main/java/cn/mesalab/service/BaselineSingleThread.java

@@ -95,11 +95,13 @@ public class BaselineSingleThread extends Thread {
             Arrays.fill(baselineArr, (int)percentile);
             // KF
             baselineArr = baselineFunction(series);
+            System.out.println("高频IP:" + ip + " origin:" + series + "\n baseline:" + Arrays.toString(baselineArr));
         } else {
             // 判断周期性
             if (SeriesUtils.isPeriod(series)){
                 // KF
                 baselineArr = baselineFunction(series);
+                System.out.println("低频周期IP:" + ip + " origin:" + series + "\n baseline:" + Arrays.toString(baselineArr));
             } else {
                 // 百分位数
                 int ipPercentile = SeriesUtils.percentile(
@@ -107,6 +109,7 @@ public class BaselineSingleThread extends Thread {
                                 Integer.valueOf(i.get(ApplicationConfig.BASELINE_METRIC_TYPE).toString())).collect(Collectors.toList()),
                         ApplicationConfig.BASELINE_RATIONAL_PERCENTILE);
                 Arrays.fill(baselineArr, ipPercentile);
+                System.out.println("其他IP:" + ip + " origin:" + series + "\n baseline:" + Arrays.toString(baselineArr));
             }
         }
 

src/main/java/cn/mesalab/service/algorithm/KalmanFilter.java

@@ -21,6 +21,7 @@ public class KalmanFilter {
     private double mdelt;
     private double Gauss;
     private double kalmanGain;
+    private final static double Q = ApplicationConfig.BASELINE_KALMAN_Q;
     private final static double R = ApplicationConfig.BASELINE_KALMAN_R;
 
     public KalmanFilter() {
@@ -29,21 +30,23 @@ public class KalmanFilter {
 
     public void initial(){
         pdelt = ApplicationConfig.BASELINE_KALMAN_P;
+        mdelt = ApplicationConfig.BASELINE_KALMAN_M;
     }
 
     private ArrayList<Integer> smoothSeries;
     private ArrayList<Integer> forecastSeries;
 
     public Integer calSingleKalPoint(Integer oldValue, Integer value){
-        //第一个估计值
         predict = oldValue;
         current = value;
+        //高斯噪声方差
+        Gauss = Math.sqrt(pdelt * pdelt + mdelt * mdelt) + Q;
         //估计方差
-        kalmanGain = pdelt/(pdelt + R);
+        kalmanGain = Math.sqrt((Gauss * Gauss)/(Gauss * Gauss + pdelt * pdelt)) + R;
         //估计值
         estimate = (int) (kalmanGain * (current - predict) + predict);
         //新的估计方差
-        mdelt = (1-kalmanGain) * pdelt ;
+        mdelt = Math.sqrt((1-kalmanGain) * Gauss * Gauss);
 
         return estimate;
     }
@@ -68,6 +71,9 @@ public class KalmanFilter {
             }
             forecastSeries.add((int)sum/partitonNum);
         }
+
+        System.out.println("KF test: origin:" + historicalSeries + "\n smooth:" + smoothSeries + "\n baseline:" + forecastSeries);
+
     }
 
     public ArrayList<Integer> getSmoothSeries() {

src/main/resources/application.properties

@@ -47,8 +47,10 @@ baseline.historical.ratio.threshold=0.1
 baseline.historical.sparse.fill.percentile=0.95
 baseline.rational.percentile=0.95
 #Kalman Filter
-baseline.kalman.p=2
+baseline.kalman.q=0.00001
-baseline.kalman.r=4
+baseline.kalman.r=0.1
+baseline.kalman.p=1
+baseline.kalman.m=1
 
 # FOR TEST
 baseline.generate.batch.size=100