CN116434487A - Distributed and self-adaptive threshold adjustment method and DAS system applying same - Google Patents
Distributed and self-adaptive threshold adjustment method and DAS system applying same Download PDFInfo
- Publication number
- CN116434487A CN116434487A CN202211581002.5A CN202211581002A CN116434487A CN 116434487 A CN116434487 A CN 116434487A CN 202211581002 A CN202211581002 A CN 202211581002A CN 116434487 A CN116434487 A CN 116434487A
- Authority
- CN
- China
- Prior art keywords
- alarm
- score
- recall
- precision
- updating
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 64
- 239000013307 optical fiber Substances 0.000 claims description 22
- 230000003044 adaptive effect Effects 0.000 claims description 13
- 238000004364 calculation method Methods 0.000 claims description 8
- 238000003062 neural network model Methods 0.000 claims description 8
- 235000007164 Oryza sativa Nutrition 0.000 claims description 3
- 235000009566 rice Nutrition 0.000 claims description 3
- 240000007594 Oryza sativa Species 0.000 claims 1
- 230000006870 function Effects 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 5
- 238000003491 array Methods 0.000 description 4
- 230000003287 optical effect Effects 0.000 description 4
- 230000006399 behavior Effects 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 3
- 238000012790 confirmation Methods 0.000 description 3
- 230000001066 destructive effect Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 241000209094 Oryza Species 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000009412 basement excavation Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- CLOMYZFHNHFSIQ-UHFFFAOYSA-N clonixin Chemical compound CC1=C(Cl)C=CC=C1NC1=NC=CC=C1C(O)=O CLOMYZFHNHFSIQ-UHFFFAOYSA-N 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 239000002689 soil Substances 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
- G01H9/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves by using radiation-sensitive means, e.g. optical means
- G01H9/004—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves by using radiation-sensitive means, e.g. optical means using fibre optic sensors
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B21/00—Alarms responsive to a single specified undesired or abnormal condition and not otherwise provided for
- G08B21/18—Status alarms
- G08B21/182—Level alarms, e.g. alarms responsive to variables exceeding a threshold
Abstract
The invention discloses a distributed self-adaptive threshold adjustment method and a DAS system applying the method, wherein the method comprises the following steps: determining the capacities of an accuracy scoring array, a recall scoring array and a confidence threshold array when the system is started and initializing; the user confirms and submits alarm feedback information; the previous precision and recall rate scores of the alarm occurrence position are obtained from the precision score array and the recall rate score array in an index way; updating the precision and recall rate scores of the alarm occurrence position and each position in the influence range according to the alarm feedback information and updating the confidence threshold value in the alarm influence range by using the acquired previous precision and recall rate scores; the accuracy of the completed update, recall score, and confidence threshold are stored in a disk of the DAS system. The invention can dynamically update the precision, recall rate score and confidence threshold of each position according to the actual condition of the alarm position fed back by the user so as to balance the precision and recall rate of the system and optimize the comprehensive index of the system.
Description
Technical Field
The invention relates to the technical field of distributed optical fiber vibration sensor mode identification, in particular to a distributed and self-adaptive threshold adjustment method and a DAS system applying the method.
Background
The basic principle of the DAS system is that an optical pulse is emitted into an optical fiber, then the optical pulse is received along a rayleigh scattering (Rayleigh scattering) signal on the optical fiber path, the external vibration can cause fluctuation of a reflected signal, and the vibration everywhere along the optical fiber path can be detected by detecting the reflected signal. DAS system has been widely used in perimeter security and pipeline early warning fields at present. The DAS system software recognizes and detects potentially destructive behaviors such as excavation, construction and the like according to the vibration signals, and if the events are detected, an alarm is sent to remind a user.
According to the principle, the effect of external vibration on the optical fiber is very complex, and in addition, in the application of pipeline security, the optical cable is usually buried underground, a vibration signal between a signal to be detected (such as an excavator) and the optical cable, and the propagation process of the vibration signal in soil are very difficult to attempt to directly model from the signal to restore the target characteristics.
Currently, deep learning techniques have been applied to DAS systems for pattern recognition. The deep neural network (DNN, deep Neural Networks) has strong capability of extracting target signal characteristics from big data, and algorithms such as convolutional neural network (Convolutional Neural Networks), time recurrent neural network (Recurrent Neural Networks), YOLO (you only look once) and the like can be applied to the DAS system, and better alarm accuracy is obtained than the traditional algorithm.
The final output of DNN is mostly the value of the (0, 1) interval obtained by Sigmoid or Softmax function, representing the concept of probability, confidence (confidence is used herein). After the DAS application DNN performs signal identification, if explicit warning is needed, a confidence coefficient threshold value needs to be set, and when the output confidence coefficient exceeds the set threshold value, the DAS application DNN indicates that the target is detected and generates the warning. However, the confidence level of the neural network output is not actually and physically significant, so that the threshold value is set relatively subjectively, and the threshold value is usually a global fixed value or a partition fixed value and is not the threshold value of each position, so that the threshold value of each position on the optical fiber cannot be set properly
Disclosure of Invention
In view of the above, the present invention provides a distributed, adaptive threshold adjustment method and a DAS system applying the same, which are used for solving the above-mentioned problems in the prior art.
A distributed and self-adaptive threshold adjustment method specifically comprises the following steps:
s1, determining the capacities of an accuracy scoring array, a recall scoring array and a confidence threshold array when the system is started according to the spatial resolution of optical fibers used by a DAS system and the total length of optical fiber lines, and initializing;
s2, a user confirms an alarm sent by the DAS based on the deep neural network model and submits alarm feedback information;
s3, according to the position coordinates of the alarm occurrence position, respectively acquiring the previous precision score and recall score corresponding to the position coordinate index from the precision score array and the recall score array;
s4, updating the precision score and the recall score of each position in the alarm occurrence position and the influence range thereof and updating the confidence coefficient threshold value in the alarm influence range according to the alarm feedback information and a set updating method by utilizing the acquired previous precision score and recall score;
and S5, storing the accuracy score, recall score and confidence threshold for finishing updating in a disk of the DAS system.
Preferably, the alarm feedback information includes correct alarm, false alarm and missing alarm,
when the alarm feedback information is correct alarm or false alarm, updating the precision scores of the alarm occurrence position and each position in the influence range according to the corresponding updating method by utilizing the acquired previous precision scores;
when the alarm feedback information is correct alarm or missing alarm, the obtained previous recall rate score is utilized to update the recall rate score of each alarm occurrence position and each position in the influence range according to the corresponding updating method.
Preferably, the specific steps for updating the precision scores of the alarm occurrence positions and the positions in the influence range are as follows:
firstly, calculating the precision expectation of the alarm occurrence position according to the precision score;
then, calculating the precision score updating quantity of the alarm occurrence position according to the precision expectation, and updating the precision score of the position according to the calculated precision score updating quantity;
and then, diffusing the precision score updating quantity to each position in the influence range of the alarm event by utilizing a Gaussian kernel function, and updating the precision score of each position in the influence range of the alarm event.
Preferably, the precision expectation calculation method:
the accuracy score update amount at the alert occurrence position x 0:
ΔP＝K*(WP-EP(x0))
the updating method of the precision score at the alarm occurrence position x0 comprises the following steps:
P(x0)←P(x0)+K*(WP-EP(x0))
the updating method of the precision scores of all the positions in the influence range of the alarm event comprises the following steps:
P(x)←P(x)+ΔP*Gk(x-x0)，x∈[x0-3σ，x0+3σ]
wherein P (x 0) is a previous accuracy score at alert occurrence location x 0; EP (x 0) is the precision expectation at the alert occurrence position x 0; wp=1 when the alarm feedback information is a correct alarm, wp=0 when the alarm feedback information is a false alarm; k is a scoring coefficient; gk is a gaussian kernel function and σ is the variance of the gaussian distribution.
Preferably, the specific steps for updating the recall score of each position in the alarm occurrence position and the influence range thereof are as follows:
firstly, calculating recall rate expectations at alarm occurrence positions according to recall rate scores;
then, calculating the recall score updating amount of the alarm occurrence position according to the recall expectation, and updating the recall score of the position according to the calculated recall score updating amount;
and then, diffusing the recall score updating quantity to each position in the influence range of the alarm event by using a Gaussian kernel function, and updating the recall score of each position in the influence range of the alarm event.
Preferably, the recall expectation calculation method comprises the following steps:
recall score update amount at alert occurrence location x 0:
ΔR＝K*(WR-ER(x0))
the method for updating the recall score at the alarm occurrence position x0 comprises the following steps:
R(x0)←R(x0)+K*(WR-ER(x0))
the method for updating the recall rate scores of all the positions in the influence range of the alarm event comprises the following steps:
R(x)←R(x)+ΔR*Gk(x-x0)，x∈[x0-3σ，x0+3σ]
wherein R (x 0) is a previous recall score at alert occurrence location x 0; ER (x 0) is the recall expectation at alert occurrence location x 0; when the alarm feedback information is correct alarm, WR=1, and when the alarm feedback information is missing alarm, WR=0; k is a scoring coefficient; σ is the variance of the gaussian distribution.
Preferably, the updating method of the confidence threshold value in the alarm influence range in step S4 is as follows:
T(x)←T(x)+step*(ER(x)-EP(x))，x∈[x0-3σ，x0+3σ]；
where T (x) is the confidence threshold for the x-position, EP (x) is the precision expectation for the x-position in the alert impact range, ER (x) is the recall expectation for the x-position in the alert impact range, step is the threshold update coefficient for controlling the update rate, and σ is the variance of the gaussian distribution.
Preferably, the updating method of the confidence threshold value in the alarm influence range in step S4 is as follows:
T(x)←T(x)+step*(ER(x)-α*EP(x))，x∈[x0-3σ，x0+3σ]
where EP (x) is the precision expectation of the x-position in the alert influence range, ER (x) is the recall expectation of the x-position in the alert influence range, step is the threshold update coefficient for controlling the update speed, σ is the variance of the gaussian distribution, and α is the balance coefficient of the DAS system when dynamically adjusting between the two directions of precision and recall.
Preferably, the capacities of the precision score array, recall score array, and confidence threshold array:
wherein, L is the total length of the optical fiber line, and spatial_resolution is the spatial resolution of the optical fiber used by the DAS system, in units of: and (5) rice.
A DAS system dynamically updates the accuracy value, recall and confidence threshold of a deep neural network model by using the distributed and adaptive threshold adjustment method.
The beneficial effects of the invention are as follows:
1. the DAS system utilizes the deep neural network model to process the vibration signals so as to identify and detect the potential destructive behavior, can evaluate the precision, recall rate and confidence threshold of each position on the optical fiber, and can dynamically update the precision score, recall rate score and confidence threshold of each position according to the actual condition of the alarm position fed back by the user so as to balance the precision and recall rate and optimize the comprehensive index of the system.
2. When the method of the invention is utilized, the deep neural network model of the DAS system can properly improve the confidence coefficient threshold value for the region with high recall rate and low precision, thereby obtaining better precision, otherwise, the confidence coefficient threshold value is reduced to improve the recall rate, thereby realizing a distributed and self-adaptive threshold value, obtaining balanced precision and recall rate at each position, simultaneously setting the balance coefficient adjusted between the precision and the recall rate, and optimizing the comprehensive performance of the system.
3. The method of the invention does not need priori information, such as information of site environment, equipment parameters and the like, from the standpoint of statistics. And updating the warning precision and recall rate scores according to the correct warning, false positive and missing report fed back by the user, wherein only limited feedback information is used for evaluation and updating, and complete warning verification is not needed, namely the correctness of each warning is verified and each missing report is found.
4. The algorithm is light, simple and efficient, and is suitable for embedded systems and edge calculation.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic diagram of an array of stored precision scores P, recall scores R, and thresholds T, and a Tupdate.
Fig. 2 is a schematic diagram of the update mechanism of P and R.
Fig. 3 is a flow chart of the method of the present invention.
Detailed Description
For a better understanding of the technical solution of the present invention, the following detailed description of the embodiments of the present invention refers to the accompanying drawings.
It should be understood that the described embodiments are merely some, but not all, embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The present application is described in further detail below by way of specific embodiments and with reference to the accompanying drawings.
The invention provides a distributed and self-adaptive threshold adjustment method, wherein the DAS system utilizes a deep neural network model to process vibration signals so as to identify and detect potential destructive behaviors, and the method can evaluate the precision, recall rate and confidence threshold of each position on an optical fiber, dynamically update the precision score, recall rate score and confidence threshold of each position so as to balance the precision and recall rate and optimize the comprehensive index of the system.
The alarms sent by DAS systems are basically divided into 4 cases, respectively:
TP (True Positive) there is an actual alarm event, there is an alarm
FP (False Positive) there is no alarm, i.e. false alarm
TN (True Negative) actual no alarm event, no alarm
FN (False Negative) in practice, there is an alarm event, no alarm, i.e. no alarm
Precision (Precision) is defined as:
recall (Recall) is defined as:
in an ideal case, the precision and recall may both reach 100%, but in practice, both are usually mutually exclusive and constrained, and are usually evaluated using the comprehensive index (F-Measure), where the most common F1 score is:
it can be seen that in the practical case where the precision and recall are mutually exclusive, the precision and recall values can achieve a better F1 score at equilibrium (this is not strictly mathematically justified, since the precision and recall are theoretically uncorrelated and constrained, but conform to most practical cases). According to the definition of the precision and the recall rate, the statistics of the correct alarm TP, the false alarm FP and the missing alarm FN are needed to be obtained, so that the statistics (classical probability) of the precision and the recall rate can be obtained. In the actual situation, a few alarms cannot confirm whether the alarms are correct alarms, and the situation that no alarms are found and counted exists, so that three quantities TP, FP and FN may not be matched. It is therefore difficult in practice to derive accuracy and recall from the definition above, and it is more difficult for the distributed system to count the accuracy and recall for each location.
Therefore, the distributed, adaptive threshold adjustment method of the present invention applies an erlo ranking system (Elo rating system), evaluates the accuracy and recall of the DAS system using an accuracy score P (hereinafter P score) and a recall score R (hereinafter R score), and updates and determines the confidence threshold based on the scoring results. The DAS system of the present invention uses 3 arrays to store and represent these distributed values, P, R and threshold T, respectively, which 3 arrays load data at system initialization and maintain the system at run-time.
Specifically, the distributed and self-adaptive threshold adjustment method of the invention specifically comprises the following steps:
s1, determining the capacities of a P array, an R array and a T array when the system is started according to the spatial resolution of optical fibers used by the DAS system and the total length of optical fiber lines, and initializing. The capacity was calculated as follows:
wherein, L is the total length of the optical fiber line, and spatial_resolution is the spatial resolution of the optical fiber used by the DAS system, in units of: and (5) rice. The index of the x position is:
wherein idx is the index of the x position, and the value of the array at the x position can be found through the index. The spatial resolution of the optical fibers is different, and the array capacities of the three arrays are correspondingly adjusted. For DAS systems, if the DAS system adopts different spatial resolution settings, only the array capacity and the calculation index need to be adjusted, the complexity of the method is not increased, and the spatial resolution is assumed to be 1 meter to simplify the description.
S2, a user confirms an alarm sent by the DAS based on the deep neural network model and submits alarm feedback information.
The alarm feedback information comprises correct alarm, false alarm and missing alarm, namely, a user confirms whether the alarm is correct alarm, false alarm or missing alarm, and submits the confirmation result as alarm feedback information.
S3, according to the position coordinates of the alarm occurrence position, respectively acquiring the previous P score and the previous R score corresponding to the position coordinate index from the P, R array.
S4, updating the P and R of the alarm occurrence position and each position in the influence range of the alarm occurrence position and the confidence coefficient threshold value in the alarm influence range respectively by using the obtained previous P score and R score and according to the alarm feedback information and a set updating method.
The user confirms that the correct or incorrect alarm can cause the updating of the P score, and confirms that the correct alarm or the missing alarm can cause the updating of the R score, namely when the alarm feedback information is the correct alarm or the false alarm, the P of each position in the alarm occurrence position and the influence range is updated by utilizing the obtained previous P score according to the corresponding updating method, and the specific updating steps are as follows:
the calculation method of the precision expectation obtained according to P is as follows:
where P (x) is the previous P score at position x and EP (x) is the accuracy expectation at x that would have an impact on both the threshold update and the P score update;
then, the previous P (x 0) at the alarm occurrence position x0 is substituted into the above formula (1), so that the expected EP (x 0) of the precision of the position can be calculated, wherein P (x 0) is a value corresponding to the position coordinate obtained by indexing from the P array according to the position coordinate x0 of the expected EP (x 0);
then, the P update amount Δp of the alarm occurrence position x0 is calculated from the accuracy expectation:
ΔP＝K*(WP-EP(x0)) (2)
the P updating method of the x0 position comprises the following steps:
P(x0)←P(x0)+K*(WP-EP(x0)) (3)
WP determines the value according to the alarm feedback information confirmed by the user, namely when the alarm feedback information is correct alarm, WP=1, and when the alarm feedback information is false alarm, WP=0; k is a score coefficient, and when the score value is higher, the K value can reduce scores which avoid too much error loss at a high segmentation;
the updating only updates the P part of the x0 position, and then the precision updating quantity delta P is required to be diffused to each position in the influence range of the alarm event by utilizing a Gaussian kernel function, and the updating method comprises the following steps:
P(x)←P(x)+ΔP*Gk(x-x0)，x∈[x0-3σ，x0+3σ] (4)
where Gk is a gaussian kernel function and σ is the variance of the gaussian distribution, which is actually the range of influence of the alarm event, e.g., σ of the mechanical dig event takes 30 meters.
Thus, the alarm feedback information of the user at the x0 position enables the P component of the x0 nearby area [ x0-3 sigma, x0+3 sigma ] to be updated.
When the alarm feedback information is correct alarm or missing alarm, the obtained previous R score is utilized to update the R of each position in the alarm occurrence position and the influence range according to the corresponding updating method, and the specific updating steps are as follows:
the calculation method for obtaining the recall rate prediction according to R is as follows:
where R (x) is the previous R score at location x, ER (x) is the recall expectation at location x that would have an impact on both the threshold update and the R score update;
then, the previous R (x 0) at the alarm occurrence position x0 is substituted into the above formula (6), so that the recall rate expected ER (x 0) of the position can be calculated, wherein R (x 0) is a value corresponding to the position coordinate obtained by indexing from the R array according to the position coordinate x 0;
then, R update amount DeltaR of alarm occurrence position x0 is calculated according to the recall expectation,
ΔR＝K*(WR-ER(x0)) (7)
the R updating method of the x0 position comprises the following steps:
R(x0)←R(x0)+K*(WR-ER(x0)) (8)
the WR determines the value according to the alarm feedback information confirmed by the user, namely when the alarm feedback information is correct alarm, WR=1, and when the alarm feedback information is missing alarm, WR=0; k is a score coefficient, and when the score value is higher, the K value can reduce scores which avoid too much error loss at a high segmentation;
the above update only updates R at the x0 position, and then, the update amount Δr needs to be diffused to each position update method within the influence range of the alarm event by using the gaussian kernel function, which is as follows:
R(x)←R(x)+ΔR*Gk(x-x0)，x∈[x0-3σ，x0+3σ] (9)
wherein σ is the variance of the Gaussian distribution as described above
Thus, the alarm feedback information of the user at the x0 position allows the R component of the x0 nearby area [ x0-3 sigma, x0+3 sigma ] to be updated.
The confidence threshold T (x) of the alarm event influence area is updated according to the expectations of precision and recall rate, and the updating method comprises the following steps:
T(x)←T(x)+step*(ER(x)-EP(x)) (11)
where EP (x) is the precision expectation of the x-position within the alert influence range, ER (x) is the recall expectation of the x-position within the alert influence range, step is the threshold update coefficient for controlling the update rate.
When updating the confidence threshold according to formula (11), the system dynamically adjusts towards the direction that the recall is equal to the precision, but sometimes the user considers that the recall is different from the precision, for example, the recall is considered to be more important than the alarm precision, and the updating method of the confidence threshold T (x) is modified as follows:
T(x)←T(x)+step*(ER(x)-α*EP(x)) (12)
wherein α is a balance coefficient of the DAS system when dynamically adjusting between two directions of accuracy and recall, so that the balance direction of the system is adjusted, and if α > 1, the recall is considered to be more preferable than the accuracy. According to the method, the system dynamically updates the threshold value of the area near the feedback position according to the alarm feedback information submitted by the user during operation, so that the distributed and self-adaptive confidence threshold value setting of the whole line is obtained.
S5, storing the P score, the R score and the confidence threshold T which are updated in a disk of the DAS system.
The distributed, adaptive threshold adjustment method of the present invention is described in detail below by way of example.
Assuming that the line length of the optical fiber used by the DAS system is 50km and the spatial resolution is 1m, the capacities of the P array, the R array and the confidence threshold T array are initialized to 5 ten thousand, and the P (x) and R (x) values of the x position can be obtained by indexing the corresponding arrays through the position coordinates x.
The DAS system starts up, P, R and confidence threshold array T loads data stored on disk at the time of previous use of the system to recover the previous state.
Initialization is required if the DAS system is first enabled, where P, R is all initialized to 0, T is all initialized to a conventional fixed threshold, and T is initialized to 0.5 in this method.
FIG. 1 is a schematic diagram of storing P, R and an array of threshold values T and updating T, wherein the values of the array P, R in the left diagram represent the P and R components of each position, and the expected EP and ER of each position can be obtained according to the formula as shown in the EP and ER graphs; the value of the T array in the right graph is the confidence threshold value of each position, the solid line in the graph is the T value before updating, and the dotted line is the T+dT value after updating; the updating range of T in the schematic diagram is up to 1000 meters, and the range of one update is usually not so large when in actual implementation.
Fig. 2 is a schematic diagram of an update mechanism of P and R, in which a user confirms and feeds back a correct alarm, an incorrect alarm, and a missed alarm, which may cause P, R to be updated, a G curve in the figure is a gaussian distribution update curve of an area near a confirmation point, a solid line in the right figure is a P, R curve before update, and a dotted line is a P, R curve after update.
Assuming that the user confirms that an alarm is a correct alarm at the 294m position, the update mode of P to the correct alarm is as follows, and the update amount of 294m position is calculated according to the above:
ΔP＝K*(WP-EP[294])
where the correct alert takes wp=1, the desired EP 294=0.56 is obtained from P294=41.9. In this method, the score is 32 for K at (-1000, 1000) and 16 for K at above this range. Thus taking k=32 brings the above equation to Δp=14. The updated values are then spread to the vicinity, σ=30, and the gaussian kernel Gk is calculated as described above. As shown in the upper right graph of fig. 2, the P values in this range are updated as follows:
P(x)←P(x)+14*Gk(x-294),x∈[204,384]
the R value can be updated in the same way when the alarm is correct, and the description is omitted.
Assuming that a missing report occurs in the 294m position confirmation by the user, the update mode of the missing report by R is as follows, and the update amount of the 294m position is calculated according to the above:
ΔR＝K*(WR-ER[294])
where the leak alarm takes wr=0, the desired ER 294=0.65 can be obtained from R294=107.5, and k=32 is brought into the above equation to obtain Δr= -20.8. With the same gaussian kernel Gk as described above, as shown in the lower right graph of fig. 2, the R values in this range are updated as follows:
R(x)←R(x)-20.8*Gk(x-294)，x∈[204，384]
the P value can be updated in the same way by false alarm, and the description is omitted. After the update of P and R is completed, the threshold T of the area needs to be updated, and the method takes the update rate step=0.3 and the balance coefficient α=1.1, calculates ER and EP of the area, and then updates T:
T(x)←T(x)+0.3*(ER(x)-1.1*EP(x))，x∈[204，384]
both P, R and T require disk storage after the update is completed to restore state when the system is restarted.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather to enable any modification, equivalent replacement, improvement or the like to be made within the spirit and principles of the invention.
Claims (10)
1. The distributed and self-adaptive threshold adjustment method is characterized by comprising the following steps of:
s1, determining the capacities of an accuracy scoring array, a recall scoring array and a confidence threshold array when the system is started according to the spatial resolution of optical fibers used by a DAS system and the total length of optical fiber lines, and initializing;
s2, a user confirms an alarm sent by the DAS based on the deep neural network model and submits alarm feedback information;
s3, according to the position coordinates of the alarm occurrence position, respectively acquiring the previous precision score and recall score corresponding to the position coordinate index from the precision score array and the recall score array;
s4, updating the precision score and the recall score of each position in the alarm occurrence position and the influence range thereof and updating the confidence coefficient threshold value in the alarm influence range according to the alarm feedback information and a set updating method by utilizing the acquired previous precision score and recall score;
and S5, storing the updated precision value, recall and confidence threshold in a disk of the DAS system.
2. The distributed, adaptive threshold adjustment method of claim 1, wherein the alarm feedback information includes correct alarms, false alarms and false alarms,
when the alarm feedback information is correct alarm or false alarm, updating the precision scores of the alarm occurrence position and each position in the influence range according to the corresponding updating method by utilizing the acquired previous precision scores;
when the alarm feedback information is correct alarm or missing alarm, the obtained previous recall rate score is utilized to update the recall rate score of each alarm occurrence position and each position in the influence range according to the corresponding updating method.
3. The method for adjusting the distributed and adaptive threshold according to claim 2, wherein the specific step of updating the precision scores of the alarm occurrence location and each location within the influence range thereof is as follows:
firstly, calculating the precision expectation of the alarm occurrence position according to the precision score;
then, calculating the precision score updating quantity of the alarm occurrence position according to the precision expectation, and updating the precision score of the position according to the calculated precision score updating quantity;
and then, diffusing the precision score updating quantity to each position in the influence range of the alarm event by utilizing a Gaussian kernel function, and updating the precision score of each position in the influence range of the alarm event.
4. A distributed, adaptive threshold adjustment method according to claim 3, characterized by a precision expectation calculation method:
the accuracy score update amount at the alert occurrence position x 0:
ΔP＝K*(WP-EP(x0))
the updating method of the precision score at the alarm occurrence position x0 comprises the following steps:
P(x0)←P(x0)+K*(WP-EP(x0))
the updating method of the precision scores of all the positions in the influence range of the alarm event comprises the following steps:
P(x)←P(x)+ΔP*Gk(x-x0),x∈[x0-3σ,x0+3σ]
wherein P (x 0) is a previous accuracy score at alert occurrence location x 0; EP (x 0) is the precision expectation at the alert occurrence position x 0; wp=1 when the alarm feedback information is a correct alarm, wp=0 when the alarm feedback information is a false alarm; k is a scoring coefficient; gk is a gaussian kernel function and σ is the variance of the gaussian distribution.
5. The method for adjusting the distributed and adaptive threshold according to claim 2, wherein the specific step of updating the recall score of each of the alarm occurrence locations and the influence ranges thereof is as follows:
firstly, calculating recall rate expectations at alarm occurrence positions according to recall rate scores;
then, calculating the recall score updating amount of the alarm occurrence position according to the recall expectation, and updating the recall score of the position according to the calculated recall score updating amount;
and then, diffusing the recall score updating quantity to each position in the influence range of the alarm event by using a Gaussian kernel function, and updating the recall score of each position in the influence range of the alarm event.
6. The distributed, adaptive thresholding method of claim 5, characterized by a recall expectation calculation method:
recall score update amount at alert occurrence location x 0:
ΔR＝K*(WR-ER(x0))
the method for updating the recall score at the alarm occurrence position x0 comprises the following steps:
R(x0)←R(x0)+K*(WR-ER(x0))
the method for updating the recall rate scores of all the positions in the influence range of the alarm event comprises the following steps:
R(x)←R(x)+ΔR*Gk(x-x0),x∈[x0-3σ,x0+3σ]
wherein R (x 0) is a previous recall score at alert occurrence location x 0; ER (x 0) is the recall expectation at alert occurrence location x 0; when the alarm feedback information is correct alarm, WR=1, and when the alarm feedback information is missing alarm, WR=0; k is a scoring coefficient; σ is the variance of the gaussian distribution.
7. The method for adjusting a distributed and adaptive threshold according to claim 1, wherein the method for updating the confidence threshold in the alert influence range in step S4 is as follows:
T(x)←T(x)+step*(ER(x)-EP(x)),x∈[x0-3σ,x0+3σ]
where T (x) is the confidence threshold for the x-position, EP (x) is the precision expectation for the x-position in the alert impact range, ER (x) is the recall expectation for the x-position in the alert impact range, step is the threshold update coefficient for controlling the update rate, and σ is the variance of the gaussian distribution.
8. The method for adjusting a distributed and adaptive threshold according to claim 1, wherein the method for updating the confidence threshold in the alert influence range in step S4 is as follows:
T(x)←T(x)+step*(ER(x)-α*EP(x)),x∈[x0-3σ,x0+3σ]
where EP (x) is the precision expectation of the x-position in the alert influence range, ER (x) is the recall expectation of the x-position in the alert influence range, step is the threshold update coefficient for controlling the update speed, σ is the variance of the gaussian distribution, and α is the balance coefficient of the DAS system when dynamically adjusting between the two directions of precision and recall.
9. The distributed, adaptive threshold adjustment method of claim 1, wherein the capacities of the precision score array, recall score array, and confidence threshold array:
wherein c is the group capacity, L is the total length of the optical fiber line, and spatial_resolution is the spatial resolution of the optical fiber used by the DAS system, in units of: and (5) rice.
10. A DAS system, characterized in that the DAS system dynamically updates the accuracy score, recall score, and confidence threshold of its deep neural network model using the distributed, adaptive threshold adjustment method of any of claims 1-9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211581002.5A CN116434487A (en) | 2022-12-09 | 2022-12-09 | Distributed and self-adaptive threshold adjustment method and DAS system applying same |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211581002.5A CN116434487A (en) | 2022-12-09 | 2022-12-09 | Distributed and self-adaptive threshold adjustment method and DAS system applying same |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116434487A true CN116434487A (en) | 2023-07-14 |
Family
ID=87093165
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211581002.5A Pending CN116434487A (en) | 2022-12-09 | 2022-12-09 | Distributed and self-adaptive threshold adjustment method and DAS system applying same |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116434487A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117056584A (en) * | 2023-10-08 | 2023-11-14 | 杭州海康威视数字技术股份有限公司 | Information system abnormal change monitoring method and equipment based on dynamic similarity threshold |
-
2022
- 2022-12-09 CN CN202211581002.5A patent/CN116434487A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117056584A (en) * | 2023-10-08 | 2023-11-14 | 杭州海康威视数字技术股份有限公司 | Information system abnormal change monitoring method and equipment based on dynamic similarity threshold |
| CN117056584B (en) * | 2023-10-08 | 2024-01-16 | 杭州海康威视数字技术股份有限公司 | Information system abnormal change monitoring method and equipment based on dynamic similarity threshold |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chuang et al. | Robust TSK fuzzy modeling for function approximation with outliers | |
| Chuang et al. | The annealing robust backpropagation (ARBP) learning algorithm | |
| US5479576A (en) | Neural network learning system inferring an input-output relationship from a set of given input and output samples | |
| CN107808122B (en) | Target tracking method and device | |
| CN109067773B (en) | Vehicle-mounted CAN network intrusion detection method and system based on neural network | |
| US5373456A (en) | Expert system for assessing accuracy of models of physical phenomena and for selecting alternate models in the presence of noise | |
| US9262672B2 (en) | Pattern recognition apparatus and pattern recognition method that reduce effects on recognition accuracy, and storage medium | |
| US7747084B2 (en) | Methods and apparatus for target discrimination using observation vector weighting | |
| US9336373B2 (en) | User biometric pattern learning and prediction | |
| CN116434487A (en) | Distributed and self-adaptive threshold adjustment method and DAS system applying same | |
| CN111523143B (en) | Method and device for clustering private data of multiple parties | |
| US20230237309A1 (en) | Normalization in deep convolutional neural networks | |
| CN109936568A (en) | A kind of preventing malicious attack sensor data acquisition method based on Recognition with Recurrent Neural Network | |
| CN116645396A (en) | Track determination method, track determination device, computer-readable storage medium and electronic device | |
| Pryor et al. | Evaluation of a User Authentication Schema Using Behavioral Biometrics and Machine Learning | |
| Khalid et al. | Exploiting vulnerabilities in deep neural networks: Adversarial and fault-injection attacks | |
| US5001631A (en) | Cellular network assignment processor using randomly triggered adaptive cell thresholds | |
| JP2022537977A (en) | Apparatus and method for lattice point enumeration | |
| Dihua et al. | Adaptive KLD sampling based Monte Carlo localization | |
| CN113392857B (en) | Target detection method, device and equipment terminal based on yolo network | |
| CN112633299B (en) | Target detection method, network, device, terminal equipment and storage medium | |
| US20240135150A1 (en) | Topology-augmented system for ai-model mismatch | |
| US20220284261A1 (en) | Training-support-based machine learning classification and regression augmentation | |
| Stindl et al. | Stochastic declustering of earthquakes with the spatiotemporal renewal ETAS model | |
| CN113657612A (en) | Cooperative game theory-based detection method and system for backdoor attacks in federal learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |