General repeated alarm suppression method for power transmission line monitoring system based on time sequence sample analysis

Abstract

The invention discloses a time sequence sample analysis-based universal repeated alarm suppression method for a power transmission line monitoring system, and the method can analyze and judge the change condition of an abnormal region through the combination of the feature map analysis of a deep learning algorithm and time sequence sample modeling, thereby determining whether to continue to give an alarm or not. According to the invention, through time sequence analysis and thermodynamic diagram comparison, repeated alarm can be effectively inhibited. According to the method, weighted combination is carried out on the gradient map of back propagation of the neural network and the output layer feature map, the activation response thermodynamic diagram is generated, and therefore the area which contributes to the abnormal output result most in the image is accurately recognized. Through global average pooling and a ReLU activation function, the system can accurately position an abnormal region. According to the technical means, the accuracy of abnormal region identification is improved, and a reliable basis is provided for subsequent alarm suppression.

Description

Technical Field

[0001] The present invention belongs to the technical field of monitoring and alarming, and particularly relates to a general repeated alarm suppression method for a transmission line monitoring system based on time-series sample analysis. Background Art

[0002] With the intelligent development of the power system, the monitoring and maintenance of transmission lines have become an important link to ensure the safety of power supply. Traditional transmission line monitoring relies on manual inspections and real-time data collection of equipment, which improves the safety of transmission lines to a certain extent. In recent years, with the development of AI technology, more and more visual intelligent recognition technologies have been applied to the monitoring of transmission lines. However, due to the large number of monitoring devices, huge amounts of data, and wide distribution, the difficulty of information processing and analysis has increased accordingly.

[0003] In the current transmission line monitoring system, the alarm system often issues alarm signals based on the judgment of abnormalities by AI vision algorithms, and most of these alarm signals are triggered through threshold setting or preset rules. However, AI vision analysis will introduce many repeated alarms and false alarms. When there is no obvious change in the target monitoring environment, alarms will continue to be generated, and these repeated alarms bring a great workload to the staff.

[0004] At the same time, there are already some technologies for reducing false alarms, but most of these technologies focus on single data analysis means, such as false alarm suppression based on statistical methods. These technologies rely on the specific form of the monitoring target itself. For example, for the intrusion of floating objects on the transmission line, the construction of tower cranes on the line, and the growth of plants under the line, the system needs to formulate corresponding repeated alarm suppression strategies for different monitoring targets. That is, the alarm suppression strategies for each monitoring content are independent, which will obviously increase the additional monitoring operation cost. Summary of the Invention

[0005] To solve the problems existing in the prior art, the present invention proposes a general repeated alarm suppression method for a transmission line monitoring system based on time-series sample analysis.

[0006] The technical solution of the present invention is as follows: A general repeated alarm suppression method for a transmission line monitoring system based on time-series sample analysis, the transmission line monitoring system is provided with a monitoring mode, and the monitoring mode uses a deep learning network to output a feature map group of a transmission channel. The method includes: Step 1) Set an alarm mode in the transmission line monitoring system. When an abnormality is detected in the monitoring mode, trigger the alarm mode and give a first alarm; Step 2) In the alarm mode, obtain a feature map group that represents that the confidence of the channels representing the abnormal category in the deep learning network is greater than the confidence of the channels representing the normal category; Step 3) Based on the feature map group, construct an activation response heat map; Step 4) Collect the activation response heat maps at multiple moments to form a time series sample; Step 5) Analyze the time series sample, obtain the difference value between adjacent activation response heat maps in the time series sample, and suppress repeated alarms or continuous alarms based on the difference value in combination with a given temperature threshold and area threshold; Step 6) When the anomaly disappears and exceeds the time threshold, automatically exit the alarm mode and resume the monitoring mode.

[0007] Further, the specific method of step 2) is as follows: The deep learning network includes a classification and object detection network and a segmentation network. The output channel of the classification and object detection network is a one-dimensional vector V(1×c), where c in the vector V represents the corresponding category; the output channel of the segmentation network is a three-dimensional matrix map A(n×m×c), and c in the matrix represents the corresponding category confidence; Step 2.1) Obtain the feature map group finally output by the deep learning network , where w represents the width of the feature map, h represents the height of the feature map, and c represents the number of channels of the feature map; Step 2.2) Select the channel with the highest confidence in the category channels c in step 2.1) as the category classification. During alarm, this category is the abnormal category .

[0008] Further, the specific method of step 3) is as follows: Take the abnormal category Take the gradient of the feature map group : ; ; Obtain the weights of the c channels in the feature map group through global average pooling; ; Use the weights of the c channels to weight the feature map , and then pass through the ReLU activation function to obtain the activation response heat map k(w, h) of the abnormal category on the feature map group. The activation response heat map k(w, h) is used to locate the region that determines the final abnormal category classification, that is, the abnormal occurrence region , .

[0009] Further, the weights of c channels in the feature map group are calculated as follows: ; where ; represents the size of the feature map, ; represents the value corresponding to the i-th row and j-th column on the feature map.

[0010] The specific method of step 4) includes: Step 4.1) The transmission line monitoring system obtains the activation response heat maps corresponding to multiple moments in the alarm mode, denoted as , where t is the moment; Step 4.2) Sort the activation response heat maps at multiple moments in chronological order to obtain the time-series sample .

[0011] The specific method of step 5) includes: Step 5.1) Denote each activation response heat map in the time-series sample as , , ,..., ; , , ,..., represent the activation response heat maps corresponding to the 1st, 2nd, 3rd,..., nth moments respectively; take the absolute value difference of the activation response heat maps corresponding to two adjacent moments ; Step 5.2) Given a temperature threshold T, if the value of each coordinate point is less than T, it means that there is no obvious change in the activation response heat maps at two moments, and subsequent alarms are suppressed; if there exists a coordinate point with a value greater than T, it indicates that there is a local change in the activation response heat maps at two moments, and the abnormal position is changing or spreading; Step 5.3) Given an area threshold A for the region, in step 5.2 when there exists a coordinate point with a value greater than T, further judge whether the pixel area is greater than A. If it is greater than A, it indicates that the abnormal region is undergoing a large change, and continuous alarm is issued. If it is less than A, it indicates that there is a slight change or no change in the abnormal region, indicating that the abnormal situation has not spread, and subsequent alarms are suppressed.

[0012] The expression for the absolute value difference between the activation response heat maps corresponding to the t-th moment and the (t - 1)-th moment is: ; Among them, and are the activation response heatmaps corresponding to the t-th moment and the (t - 1)-th moment respectively, .

[0013] The specific method of step 6) includes: Record the difference between the current moment and the moment when an anomaly was last detected. If the difference exceeds the time threshold, automatically exit the alarm mode and resume the monitoring mode.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a general repeated alarm suppression method for a transmission line monitoring system based on time-series sample analysis. By uniformly processing the AI algorithm outputs of different monitoring targets, it avoids designing independent suppression strategies for each monitoring target. This is different from the prior art, which usually relies on formulating separate suppression strategies for each type of anomaly, resulting in additional operating costs. The present invention only needs a set of alarm suppression strategies to handle multiple anomaly situations. The generality of the present invention significantly reduces the operating costs of the monitoring system and improves the intelligence level of the system.

[0015] By combining the feature map analysis of the deep learning algorithm and the time-series sample modeling, the present invention can analyze and judge the changes in the anomaly area, and then decide whether to continue the alarm. Compared with the traditional technology that only relies on threshold setting or rule triggering, the present invention effectively suppresses repeated alarms through time-series analysis and heatmap comparison. This method significantly reduces ineffective alarms and improves the work efficiency of the staff.

[0016] The present invention uses the gradient map of the neural network backpropagation and the feature map of the output layer for weighted combination to generate the activation response heatmap, so as to accurately identify the area in the image that contributes the most to the abnormal output result. Through global average pooling and the ReLU activation function, the system can accurately locate the area where the anomaly occurs. This technical means improves the accuracy of anomaly area recognition and provides a reliable basis for subsequent alarm suppression.

[0017] According to the changes in the anomaly area, the present invention dynamically decides whether to continue the alarm by setting a temperature threshold and an area threshold. This flexible alarm suppression mechanism can better adapt to different monitoring scenarios compared with the traditional fixed-threshold alarm system, improves the response accuracy and reliability of the system, and avoids the frequent occurrence of ineffective alarms.

[0018] The present invention divides the monitoring system into a monitoring mode and an alarm mode. The alarm mode, as an independent technical module, can be applied to any monitoring system in a plug-and-play manner. The separation of the monitoring mode and the alarm mode enables the system to be unaffected by the alarm suppression technology during normal operation, and to quickly switch to the alarm mode when an anomaly is detected. This technical means enhances the flexibility and scalability of the system, facilitating subsequent function expansion and maintenance.

[0019] The present invention realizes the automatic exit of the alarm mode by continuously analyzing the changes in the abnormal area and combining the time threshold judgment. When the abnormal situation is handled or disappears, the system will automatically return to the monitoring mode, avoiding unnecessary continuous alarms. This technical means optimizes the alarm management process and improves the automation level and operation efficiency of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a principle block diagram of a general repeated alarm suppression method for a transmission line monitoring system based on time series sample analysis; Figure 2 is a flow schematic diagram of a general repeated alarm suppression method for a transmission line monitoring system based on time series sample analysis. DETAILED DESCRIPTION OF THE INVENTION

[0021] The present invention will be further clarified below with reference to the drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art fall within the scope defined by the appended claims of this application.

[0022] Embodiment: A general repeated alarm suppression method for a transmission line monitoring system based on time series sample analysis according to the present invention. The transmission line monitoring system is provided with a monitoring mode. The monitoring mode uses a deep learning network to output a feature map group of the transmission channel, as Figure 1 and Figure 2 shown. The method includes: Step 1) Set an alarm mode in the transmission line monitoring system. In the monitoring mode, obtain the inspection map to be inspected for the transmission line, and judge the inspection map based on the inspection model. If no anomaly is found, the inspection map is judged as a normal picture. If an anomaly is detected, the inspection map is judged as an abnormal picture and the alarm mode is triggered and the first alarm is sent to the operation personnel; Step 2) In the alarm mode, obtain a feature map group indicating that the confidence of the channel representing the abnormal category in the deep learning network is greater than the confidence of the channel representing the normal category; Step 3) Based on the feature map group, construct an activation response heat map for locating the abnormal occurrence area through backpropagation extraction and weighted processing; Step 4) Collect activation response heat maps at multiple moments over time to form time series samples; Step 5) Analyze the time series samples, obtain the difference value between adjacent activation response heat maps in the time series samples, and suppress repeated alarms or continuous alarms based on the difference value and the given temperature threshold and area threshold; Step 6) When the anomaly disappears and exceeds the time threshold, the alarm mode is automatically exited and the monitoring mode is restored.

[0023] The present invention solves the problem of repeated alarms in the power transmission line monitoring system caused by no obvious changes in the environment by modeling the feature map of the output layer of the AI algorithm in time series, combined with the analysis of the activation response heat map. Specifically, the system determines whether the abnormal area has changed by analyzing the activation response heat map of the previous and next time series. If the response area has not changed significantly, repeated alarms are suppressed; if the response area changes significantly, the alarm continues. This technical means effectively reduces the frequency of invalid alarms and reduces the workload of staff.

[0024] Furthermore, the specific method of step 2) in this example is as follows: Step 2.1) The alarm has been triggered in step 1. The source of the alarm is that the monitoring algorithm has found an anomaly. At the output level of the neural network, the confidence of the channel representing the abnormal category is greater than the confidence of the channel representing the normal category. The deep learning network includes classification and target detection networks and segmentation networks. For classification and target detection networks, the output channel is a one-dimensional vector V (1×c), where c in vector V represents the corresponding category; for segmentation networks, the output channel is a three-dimensional matrix A (n×m×c), where c in the matrix represents the corresponding category confidence; Step 2.2) Obtain the feature map group of the final output of the deep learning network , where w represents the width of the feature map, h represents the height of the feature map, and c represents the number of channels of the feature map; Step 2.3) Select the channel with the highest confidence in the category channel c in step 2.1) as the category to be classified. When an alarm is issued, this category is the abnormal category. .

[0025] Furthermore, the specific method of step 3) in this example is as follows: By exception category For feature map group Finding the Gradient , this gradient is grouped with the feature map in dimension The gradient reflects the contribution of the feature layer to the final result: ; Obtain a set of feature maps through global average pooling Weights of c channels in ; Among them, represents the size of the feature map and serves as a weight scaling factor, ; represents the value corresponding to the i-th row and j-th column on the feature map; Using the weights of c channels, weight the feature map and then pass through the ReLU activation function to obtain the activation response heat map k(w, h) of the abnormal category on the set of feature maps. The activation response heat map k(w, h) is used to locate the region that determines the classification of the final abnormal category, that is, the abnormal occurrence region , .

[0026] Furthermore, the specific method of step 4) in this example includes: Step 4.1) As time goes by, the power transmission line monitoring system obtains the activation response heat maps corresponding to multiple moments in the alarm mode by using the methods of step 2) and step 3), denoted as , where t is the moment; these activation response heat maps reflect the changes in the spatial position of the abnormal regions considered by the algorithm; Step 4.2) Sort the activation response heat maps at multiple moments in chronological order to obtain a time series sample .

[0027] Furthermore, the specific method of step 5) in this example includes: Obtain the difference values between adjacent activation response heat maps in the time series sample, and combine the temperature threshold and the region threshold to dynamically determine whether to suppress repeated alarms: if the value of each coordinate point in the difference value is less than the temperature threshold, suppress subsequent alarms; otherwise, compare the area of the abnormal region formed by all coordinate points greater than or equal to the temperature threshold in the difference value with the region area threshold. If the area of the abnormal region exceeds the region area threshold, continue to alarm. Specifically Step 5.1) Denote each activation response heat map in the time series sample as , , ,...,[[]]END]] ; , , ,...,[[]]END]] respectively represent the activation response heatmaps corresponding to the 1st, 2nd, 3rd, …, nth moments; take the absolute value difference between the activation response heatmaps corresponding to two adjacent moments, and the absolute value difference between the activation response heatmaps corresponding to the tth moment and the (t - 1)th moment The expression is as follows: , where and respectively represent the activation response heatmaps corresponding to the tth moment and the (t - 1)th moment, ;

[0028] Step 5.2) Given a temperature threshold T, if the value of each coordinate point is less than T, it means that there is no obvious change in the activation response heatmaps at two moments, then subsequent alarms are suppressed; if there exists a coordinate point whose value is greater than T, it indicates that there is a local change in the activation response heatmaps at two moments, and the abnormal position is changing or spreading; it can be expressed as: , then the alarm is suppressed; , the abnormal area is spreading.

[0029] Step 5.3) Given an area threshold A of the region, in step 5.2 in the case that there exists a coordinate point whose value is greater than T, further determine whether the pixel area of the abnormal region surrounded by the coordinate points greater than T is greater than A. If it is greater than A, it indicates that the abnormal region is undergoing a large change, then continuous alarm is given; if it is less than A, it indicates that the abnormal region has a slight change or no change, indicating that the abnormal situation has not spread, then subsequent alarms are suppressed. It can be expressed as: , then the abnormality is spreading rapidly; , then the abnormal spread is slow and the alarm is suppressed.

[0030] Furthermore, the specific method of step 6) in this example includes: Record the difference between the current moment and the moment when the abnormality was last detected. If the difference exceeds the time threshold, it means (the abnormality has disappeared), and the alarm mode is automatically exited and the monitoring mode is restored.

[0031] As mentioned above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A general repeated alarm suppression method for a transmission line monitoring system based on time-series sample analysis. The transmission line monitoring system has a monitoring mode, and the monitoring mode uses a deep learning network to output a feature map group of a transmission channel. It is characterized in that, The method includes: Step 1) Set an alarm mode in the transmission line monitoring system. When an anomaly is detected in the monitoring mode, trigger the alarm mode and issue the first alarm. Step 2) In the alarm mode, obtain a set of feature maps where the confidence of the channels representing the anomaly category in the deep learning network is greater than that of the channels representing the normal category. Step 3) Based on the set of feature maps, construct an activation response heat map. Step 4) Collect the activation response heat maps at multiple moments to form a time-series sample. Step 5) Analyze the time-series sample to obtain the difference values between adjacent activation response heat maps in the time-series sample. Based on the difference values and in combination with a given temperature threshold and area threshold, suppress repeated alarms or continuous alarms. Step 6) When the anomaly disappears and exceeds the time threshold, automatically exit the alarm mode and resume the monitoring mode.

2. The method for suppressing general repeated alarms of a transmission line monitoring system based on time-series sample analysis according to claim 1, characterized in that, The specific method of Step 2) is as follows: Step 2.1) The deep learning network includes a classification and object detection network and a segmentation network. The output channel of the classification and object detection network is a one-dimensional vector V (1×c), where c in the vector V represents the corresponding category. The output channel of the segmentation network is a three-dimensional matrix map A (n×m×c), and c in the matrix represents the corresponding category confidence. Step 2.2) Obtain the feature map group finally output by the deep learning network , where w represents the width of the feature map, h represents the height of the feature map, and c represents the number of channels of the feature map; Step 2.3) Select the channel with the highest confidence in the category channel c in Step 2.1) as the category classification. In the case of an alarm, this category is the abnormal category .

3. The method for suppressing general repeated alarms of a transmission line monitoring system based on time series sample analysis according to claim 2, wherein The specific method of Step 3) is as follows: With the abnormal category For the feature map group Calculate the gradient : ; Obtain a set of feature maps through global average pooling The weights of c channels in ; Using the weights of c channels , weight the feature map , and then pass through the ReLU activation function to obtain the activation response heatmap k(w, h) of the abnormal class on the group of feature maps. The activation response heatmap k(w, h) is used to locate the region that determines the final abnormal class classification, that is, the abnormal occurrence region , .

4. The method for suppressing general repeated alarms of a transmission line monitoring system based on time-series sample analysis according to claim 3, characterized in that The group of feature maps The weights of c channels in are calculated as follows: ; where represents the size of the feature map, ; represents the value corresponding to the i-th row and j-th column on the feature map.

5. The general repeated alarm suppression method for a transmission line monitoring system based on timing sample analysis according to claim 4, wherein The specific method of Step 4) includes: Step 4.1) The transmission line monitoring system obtains the activation response heat maps corresponding to multiple moments in the alarm mode, denoted as , where t represents the moment; Step 4.2) Sort the activation response heatmaps at multiple moments in chronological order to obtain a time-series sample .

6. The general repeated alarm suppression method for the transmission line monitoring system based on timing sample analysis according to claim 5, wherein The specific method of Step 5) includes: Step 5.1) Respectively, for each activation response heat map in the timing sample is denoted as , , ,..., ; , , ,..., respectively represent the activation response heat maps corresponding to the 1st, 2nd, 3rd,..., nth moments; take the absolute value difference of the activation response heat maps corresponding to two adjacent moments ; Step 5.2) Given a temperature threshold T, if the value of each coordinate point is less than T, it means that there is no obvious change in the activation response heat maps at two moments, and subsequent alarms are suppressed; if there exists a coordinate point whose value is greater than T, it indicates that there are local changes in the activation response heat maps at two moments, and the abnormal position is changing or spreading; Step 5.3) Given the area threshold A of the region, in step 5.2), when there are coordinate points with values greater than T, then determine whether the pixel area is greater than A. If it is greater than A, it indicates that the abnormal area is undergoing significant changes, and continuous alarms are issued. If it is less than A, it indicates that the abnormal area has minor changes or no changes, indicating that the abnormal situation has not spread, and subsequent alarms are suppressed.

7. The general repeated alarm suppression method for the transmission line monitoring system based on time series sample analysis according to claim 6, characterized in that, The absolute value difference between the activation response heatmaps corresponding to the t-th moment and the (t-1)-th moment is expressed as: ； Among them, and are the activation response heatmaps corresponding to the t-th moment and the (t - 1)-th moment respectively, .

8. The method for suppressing general repeated alarms of a transmission line monitoring system based on time-series sample analysis according to claim 1, wherein The specific method of Step 6) includes: Record the difference between the current moment and the moment when the anomaly was last detected. If the difference exceeds the time threshold, automatically exit the alarm mode and resume the monitoring mode.

Images