Self-learning method and system for reducing cable false alarm and related equipment

Abstract

The invention discloses a self-learning method and system for reducing cable false alarm and related equipment, and the method comprises the steps: obtaining a to-be-recognized real-time vibration event, and intercepting the real-time vibration characteristics of the real-time vibration event; inputting the real-time vibration characteristics into a dual-stage recognition model, and executing dual-stage forward propagation on the real-time vibration characteristics; wherein the dual-stage identification model is formed by connecting a first sub-model with frozen parameters in series with a second sub-model for packaging; the training set of the second sub-model is generated by splitting and labeling the training set of the first sub-model; triggering or inhibiting cable alarm response according to the dichotomy instruction of whether to give an alarm; according to the method, the positive and negative sample pairs are constructed by using the time sequence slice samples, the similarity characteristics of the construction vibration samples under the time sequence continuity and the difference characteristics of the non-construction samples are automatically learned, and compared with a mode of directly training a classification model on an original signal, the method can reduce the false alarm rate of the cable while shortening the training time.

Description

Technical Field

[0001] The present invention relates to the field of cable alarms, and specifically to a self-learning method, system and related equipment for reducing false alarms of cables. Background Art

[0002] Currently, in cable safety monitoring systems, vibration sensors are often buried to monitor the vibration of the surrounding environment of underground cable trenches to identify external damage risks such as construction intrusion. Traditional identification methods generally adopt a single-stage processing framework based on classification models, that is, directly input the vibration signal into the classification model and output a judgment result of whether to alarm. However, this type of method has the following technical deficiencies: First, as one-dimensional time-series data, the original vibration signal often contains a large amount of non-structural noise, such as traffic vibration, equipment resonance, wind and rain disturbance, etc. Directly inputting it into the classification model for training is difficult to extract stable discriminative features, affecting the recognition accuracy.

[0003] Secondly, existing models generally ignore the evolutionary structure of vibration signals in the time dimension. In actual scenarios, construction behaviors have obvious persistence, and their vibration patterns have consistent characteristics across time segments, while non-construction behaviors are mostly instantaneous interferences with local mutations. Without modeling and processing, it is difficult for the classification model to fully capture such time-series change laws and is prone to false alarms. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention provides a self-learning method, system and related equipment for reducing false alarms of cables, and solves the technical problems raised in the background art through contrast learning in the previous stage and classification recognition in the later stage.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A self-learning method for reducing false alarms of cables, comprising the following steps: S1. Obtain a real-time vibration event to be recognized and intercept the real-time vibration characteristics of the real-time vibration event; S2. Input the real-time vibration characteristics into a two-stage recognition model and perform two-stage forward propagation on the real-time vibration characteristics; The two-stage forward propagation includes: S2-1. Receive the real-time vibration characteristics as the input of the first sub-model and output the feature representation of the real-time vibration characteristics; S2-2. Receive the feature representation of the real-time vibration characteristics as the input of the second sub-model and output a binary classification instruction of whether to alarm; Among them, the two-stage recognition model is encapsulated by connecting a first sub-model with frozen parameters in series to a second sub-model; the first sub-model is obtained by iterative training of a contrastive learning model, the second sub-model is obtained by iterative training of a policy discrimination model, and the training set of the second sub-model is generated by splitting and annotating the training set of the first sub-model.

[0006] S3. Trigger or inhibit the cable alarm response according to the binary classification instruction of whether to alarm.

[0007] In some specific embodiments, the iterative training steps of the first sub-model include: A1. Obtain a contrastive learning sample training set; Among them, the contrastive learning sample training set contains a number of contrastive learning samples, each contrastive learning sample is represented in the structure of a positive and negative sample pair, and each positive and negative sample pair is represented as a historical vibration event in a quadruple structure. A2. Select an initialized contrastive learning model; A3. Input the positive and negative sample pairs of the contrastive learning sample training set into the initialized contrastive learning model, and iteratively update the model parameters until the contrastive learning model meets the predefined convergence condition; A4. When the contrastive learning model meets the convergence condition, freeze the current model parameters to obtain the first sub-model.

[0008] In some specific embodiments, obtaining the contrastive learning sample training set includes: A1-1. Select a number of time windows on the historical time axis; at least two independent and different historical vibration events occur within each time window; the different historical vibration events are respectively construction vibration events and non-construction vibration events; A1-3. Mark the historical vibration events within the time window on the spatial map; wherein, the spatial map is used to represent the occurrence location of the historical vibration events in the geographical space; A1-4. Cut the time window based on a preset time granularity, and cut the historical vibration events into vibration samples of K consecutive time segments; A1-5. According to the vibration samples of K consecutive time segments within the time window, construct K groups of positive and negative sample pairs within each time window; A1-6. Take the central time stamps of several time windows as the time sequence numbers, obtain the groups of positive and negative sample pairs, and construct a contrastive learning sample training set with time sequence numbers.

[0009] In some specific embodiments, according to the vibration samples of K consecutive time segments within the time window, constructing K groups of positive and negative sample pairs within each time window includes: A1-5-1. Mark the spatial center of each historical vibration event within the time window on the spatial map, where the spatial center corresponds to the main earthquake source location or the center of the ground-breaking behavior of the historical vibration event; A1-5-2. Set a collection radius with each spatial center as the collection center; the minimum value of the collection radius is defined as: beyond the regional boundary of the collection radius, the response intensity of the collection device to the vibration event is lower than the recognizable minimum threshold; A1-5-3. Within the collection radius, extract from the time window: K consecutive construction vibration samples corresponding to the construction vibration event; K consecutive non-construction vibration samples corresponding to the non-construction vibration event; A1-5-4. Perform data perturbation enhancement on the K consecutive construction vibration samples to obtain K construction vibration enhanced samples; A1-5-5. For each time segment : Pair the construction vibration samples with the construction vibration enhanced samples to obtain positive sample pairs; Pair the construction vibration samples with the non-construction vibration samples to obtain negative sample pairs; A1-5-6. Repeat A1-5-1 to A1-5-5, and construct the K time segments within each time window into positive sample pairs and negative sample pairs in sequence to generate K groups of positive and negative sample pairs.

[0010] In some specific embodiments, input the contrast learning samples of the contrast learning sample training set into the initialized contrast learning model, and iteratively update the model parameters until the contrast learning model meets the predefined convergence conditions, including: A3-1. Sequentially select K groups of positive and negative sample pairs from the contrast learning sample training set based on the time window as the current training batch; A3-2. Input the K groups of positive and negative sample pairs of the current training batch into the initialized contrast learning model, perform forward propagation, and output K four-tuple prediction vectors of the current training batch; among them, the four-tuple prediction vector includes a positive prediction vector pair and a negative prediction vector pair; A3-3. Calculate the contrast learning loss of the batch corresponding to the time window based on the predefined contrast learning loss function; The loss function is used to maximize the similarity between positive sample prediction vector pairs and minimize the similarity between negative sample prediction vector pairs; A3-4. Determine whether the batch corresponding to the time window meets the predefined model convergence conditions, and the convergence conditions include but are not limited to: the loss value is lower than the preset threshold, the change range of the loss is less than the set range, or the number of training rounds reaches the preset upper limit; A3-5. If the convergence condition is not met, perform backpropagation based on the contrastive learning loss to calculate the model gradients, and update the model parameters using the model gradients. Subsequently, continue to execute steps A3-1 to A3-4; A3-6. If the convergence condition is met, stop the model training and freeze the current parameters to obtain the first sub-model.

[0011] In some specific embodiments, the iterative training steps of the second sub-model include: B1. Construct a policy discrimination sample training set according to the contrastive learning sample training set; B2. Select an initialized policy discrimination model; B3. Input the feature-label pairs of the policy discrimination sample training set into the initialized policy discrimination model, and iteratively update the model parameters until the policy discrimination model meets the predefined convergence condition.

[0012] B4. When the policy discrimination model meets the convergence condition, freeze the current model parameters to obtain the second sub-model.

[0013] In some specific embodiments, obtaining the policy discrimination sample training set according to the contrastive learning sample training set includes: B1-1. Extract two independent and different historical vibration events within the time window according to the time sequence number of the contrastive learning sample training set; B1-2. Input the two independent and different historical vibration events into the first sub-model respectively to extract the vibration features of different historical vibration events; B1-3. After marking the historical vibration events within the time window on the spatial map, perform label annotation on the two independent and different historical vibration events; The policy of label annotation is: if it corresponds to a construction vibration event, label it as alarm; otherwise, label it as no alarm; B1-4. Combine each vibration feature with the corresponding label to construct two independent and different feature-label pairs within the corresponding time sequence number; B1-5. Repeat B1-1 to B1-4 to construct several policy discrimination sample training sets with time sequence numbers; Among them, the policy discrimination sample training set contains several policy discrimination samples, each policy discrimination sample is represented in the structure of a feature-label pair, and each time number corresponds to two independent and different feature-label pairs.

[0014] In some specific embodiments, inputting the feature-label pairs of the policy discrimination sample training set into the initialized policy discrimination model and iteratively updating the model parameters until the policy discrimination model meets the predefined convergence condition includes: B3-1. Extract the feature-label pairs corresponding to the time window for the current batch from the policy discrimination training set; B2-2. Input the feature-label pairs corresponding to the time window for the current batch into the policy discrimination model; B2-3. Predict the alarm probability based on the vibration features of the feature-label pairs; B2-4. Convert the alarm probability into a binary classification prediction value; B2-5. Calculate the classification loss between the binary classification prediction value and the sample label, and perform backpropagation with the goal of minimizing the classification loss to update the model parameters; B2-6. Determine whether the current training round meets the predefined convergence condition; B2-7. If the convergence condition is met, stop the model training, freeze the current parameters, and obtain the second sub-model.

[0015] Compared with the prior art, a self-learning method for reducing false alarms of cables in the present invention introduces a contrastive learning structure, constructs positive and negative sample pairs using time-series slice samples, and automatically learns the similarity features of construction vibration samples under time-series continuity and the difference features from non-construction samples, thereby training a first sub-model with stronger feature expression ability and improving the discriminative ability of the model for samples with fuzzy boundaries; further, using the vibration features obtained through contrastive learning as input again, a lightweight policy discrimination model is constructed to complete the efficient classification and discrimination task of whether to alarm. Compared with the method of directly training a classification model on the original signal, the present invention can shorten the training time and reduce the false alarm rate of cables.

[0016] In a second aspect, the present invention provides a self-learning system for reducing false alarms of cables, including: A real-time vibration feature acquisition module, configured to acquire real-time vibration events to be recognized and intercept the real-time vibration features of the real-time vibration events; A two-stage forward recognition module, configured to input the real-time vibration features into a two-stage recognition model and perform two-stage forward propagation on the real-time vibration features; The two-stage forward propagation includes: Receiving the real-time vibration features as the input of the first sub-model and outputting the feature representation of the real-time vibration features; Receiving the feature representation of the real-time vibration features as the input of the second sub-model and outputting a binary classification instruction of whether to alarm; Wherein, the two-stage recognition model is composed of a first sub-model with frozen parameters connected in series with a second sub-model; the first sub-model is obtained through iterative training of a contrastive learning model, the second sub-model is obtained through iterative training of a policy discrimination model, and the training set of the second sub-model is generated by splitting and annotating the training set of the first sub-model; A cable alarm response module is used to trigger or suppress a cable alarm response according to the binary classification instruction of whether to alarm.

[0017] Thirdly, the present invention provides an electronic device, which includes a memory and a processor. The memory stores at least one computer-executable instruction, and the processor is configured to run the computer-executable instruction. When the computer-executable instruction is run by the processor, it is used to implement a self-learning method for reducing cable false alarms described in the first aspect.

[0018] Compared with the prior art, the beneficial effects of a self-learning system and an electronic device for reducing cable false alarms according to the present invention are the same as those of a self-learning method for reducing cable false alarms described above, so they will not be elaborated here. Description of the Drawings

[0019] Figure 1 is a flowchart of a self-learning method for reducing cable false alarms according to the present invention; Figure 2 is an iterative schematic diagram of the first sub-model according to the present invention; Figure 3 is a structural block diagram of a self-learning system for reducing cable false alarms according to the present invention; Figure 4 is a diagram of the electronic device according to the present invention. Detailed Embodiments

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0021] Please refer to Figure 1 , the present invention provides a self-learning method for reducing cable false alarms, including the following steps: S1. Obtain a real-time vibration event to be recognized, and intercept the real-time vibration characteristics of the real-time vibration event; S2. Input the real-time vibration characteristics into a two-stage recognition model, and perform two-stage forward propagation on the real-time vibration characteristics; The two-stage forward propagation includes: S2-1. Receive the real-time vibration characteristics as the input of the first sub-model, and output the feature representation of the real-time vibration characteristics; S2-2. Receive the feature representation of the real-time vibration characteristics as the input of the second sub-model, and output a binary classification instruction of whether to alarm; Among them, the two-stage recognition model is encapsulated by connecting the first sub-model with frozen parameters in series to the second sub-model; the first sub-model is obtained through iterative training of a contrastive learning model, the second sub-model is obtained through iterative training of a policy discrimination model, and the training set of the second sub-model is generated by splitting and annotating the training set of the first sub-model.

[0022] S3. Trigger or inhibit the cable alarm response according to the binary classification instruction of whether to alarm.

[0023] In this embodiment, real-time vibration events are continuously collected by vibration sensors buried near the cable trench. By setting a fixed sampling frequency, the real-time vibration waveform is intercepted and preprocessed to extract vibration characteristics in the time domain, frequency domain or time-frequency domain.

[0024] Among them, the first sub-model adopts a contrastive learning structure. Through the construction of positive and negative sample pairs, it automatically learns the discriminative features between construction vibrations and non-construction vibrations, and obtains a more expressive vibration feature representation; while the second sub-model performs lightweight classification training based on the vibration features frozen and output by the first sub-model, and only needs to complete the discriminative learning of whether to alarm, significantly reducing the training difficulty and the cost of parameter tuning. By connecting the two-stage sub-models in series and encapsulating them into a unified two-stage recognition model, not only the recognition accuracy is improved, but also the false alarm information caused by non-construction factors such as environmental resonance, equipment operation, and weather changes can be effectively filtered.

[0025] In this embodiment, the iterative training steps of the first sub-model are as follows: A1. Obtain the contrastive learning sample training set; Among them, the contrastive learning sample training set contains a number of contrastive learning samples, each contrastive learning sample is characterized in the structure of positive and negative sample pairs, and each positive and negative sample pair is characterized as a historical vibration event in a quadruple structure; A2. Select the initialized contrastive learning model; A3. Input the positive and negative sample pairs of the contrastive learning sample training set into the initialized contrastive learning model, and iteratively update the model parameters until the contrastive learning model meets the predefined convergence condition; A4. When the contrastive learning model meets the convergence condition, freeze the current model parameters to obtain the first sub-model.

[0026] In this embodiment, the contrastive learning sample training set is constructed in the structure of positive and negative sample pairs. Without the need for a large number of manual labels, it can learn a discriminative vibration feature representation. Each group of quadruple sample pairs guides the model to gradually converge to a feature space with clustering characteristics by reflecting the relative relationship between construction vibration events and non-construction vibration events in the feature space. It avoids the strong dependence of traditional classification training on labeled samples and significantly improves the model's expression ability for complex vibration patterns.

[0027] Specifically in step A1, the sub-steps for obtaining the comparative learning sample training set are: A1-1. Select several time windows on the historical time axis; at least two independent and different historical vibration events occur in each time window; different historical vibration events are construction vibration events and non-construction vibration events; The time window is used to aggregate historical vibration events within a specified time period. The length of the time window can be determined based on the actual sampling frequency and the duration of the ground-breaking behavior.

[0028] A1-3. Marking the historical earthquake events within the time window on the spatial map; wherein the spatial map is used to represent the occurrence location of the historical earthquake events in the geographic space; The occurrence location specifically includes the sensor deployment coordinates, the estimated location of the vibration event source, or the location index in the geographic coordinate system. Specifically in this embodiment, the spatial map can be derived from the GNSS positioning system, the sensor layout map, or the geographic information system (GIS) platform, supporting two-dimensional coordinates, three-dimensional coordinates, or grid area definitions; and the annotation of the vibration event data can be completed by manual confirmation, event log push, or docking with the construction scheduling system to ensure the accuracy of positive and negative samples.

[0029] A1-4. Cut the time window based on the preset time granularity and cut the historical vibration event into vibration samples of K consecutive time segments; Among them, K is a positive integer, and each time segment is used to intercept the vibration data of the historical vibration event in the corresponding time interval to form K vibration samples with a time sequence; the K vibration samples constitute a vibration sample sequence with a time series structure, and different vibration samples have a natural time sequence relationship.

[0030] A1-5. Construct K groups of positive and negative sample pairs in each time window based on the vibration samples of K consecutive time segments in the time window; A1-6. Using the central timestamps of several time windows as time series numbers, obtain a group of positive and negative sample pairs, and construct a contrastive learning sample training set with time series numbers.

[0031] In this embodiment, in view of the natural properties of the continuity and time series characteristics of historical vibration events, it is proposed to intercept historical vibration events according to time windows, and further cut them into K continuous time segments based on time granularity, thereby forming K vibration sample sequences with a chronological relationship, which significantly enriches the structural information and time series characteristics of the training set. In addition, the introduction of spatial map information is used to limit the geographical distribution of samples, which improves the accuracy of geographical screening of positive and negative samples and enhances the true restoration ability of the sample construction stage. This embodiment provides more stable feature input support for the cable false alarm problem by integrating time series slicing and spatial positioning.

[0032] Furthermore, specifically in step A1-5, the sub-steps for constructing the K pairs of positive and negative samples are as follows: A1-5-1. Mark the spatial center of each historical vibration event within the time window on the spatial map, where the spatial center corresponds to the main earthquake source location or the center of the ground-breaking behavior of the historical vibration event; A1-5-2. Set the acquisition radius with each spatial center as the acquisition center; the minimum value of the acquisition radius is defined as: outside the boundary of the area beyond the acquisition radius, the response intensity of the acquisition device to the vibration event is lower than the recognizable minimum threshold; A1-5-3. Within the acquisition radius, extract from the time window: K consecutive construction vibration samples corresponding to the construction vibration events; K consecutive non-construction vibration samples corresponding to the non-construction vibration events; A1-5-4. Perform data perturbation enhancement on the K consecutive construction vibration samples to obtain K enhanced construction vibration samples; The perturbation enhancement operations may include, but are not limited to, methods such as additive noise, amplitude perturbation, time-domain cropping, filter transformation, convolution blur, etc., aiming to introduce minor changes to construct positive sample pairs while maintaining the overall semantic consistency of the vibration events.

[0033] A1-5-5. For each time segment : Pair the construction vibration samples with the enhanced construction vibration samples to obtain positive sample pairs; Pair the construction vibration samples with the non-construction vibration samples to obtain negative sample pairs; A1-5-6. Repeat steps A1-5-1 to A1-5-5, and construct the K time segments within each time window into positive and negative sample pairs in sequence to generate K pairs of positive and negative samples.

[0034] In this embodiment, by calibrating the spatial center of each historical vibration event on the spatial map and combining the acquisition radius to impose spatial constraints on the sample extraction range, it is ensured that the extracted construction and non-construction vibration samples are geographically representative. The K consecutive vibration samples extracted based on the acquisition radius retain the dynamic evolution characteristics of the construction behavior in the time sequence dimension, while the K enhanced construction vibration samples constructed through data perturbation enhancement introduce minor perturbations while maintaining semantic consistency. By constructing positive and negative sample pairs within each time segment and generating K groups of sample pairings with a time sequence structure as a whole, not only the learning ability of the model for multi-period features is enhanced, but also the discrimination of interference patterns such as non-construction resonance and accidental perturbations is improved, providing an accurate input basis for false alarm identification.

[0035] Refer to Figure 2, in step A3, the sub-steps for iteratively updating the contrastive learning model are as follows: A3-1: Sequentially select K pairs of positive and negative samples from the contrastive learning sample training set based on a time window as the current training batch; A3-2: Input the K pairs of positive and negative samples in the current training batch into the initialized contrastive learning model, perform forward propagation, and output K quadruple prediction vectors for the current training batch; among them, the quadruple prediction vector includes a positive prediction vector pair and a negative prediction vector pair; A3-3: Calculate the contrastive learning loss for the batch corresponding to the time window based on a predefined contrastive learning loss function; The loss function is used to maximize the similarity between positive sample prediction vector pairs and minimize the similarity between negative sample prediction vector pairs; The contrastive learning loss function is:

[0036] where L represents the contrastive learning loss, T represents the number of time series numbers, that is, the total number of time windows, represents the t-th construction vibration sample, represents the t-th construction vibration enhanced sample, represents the t-th non-construction vibration sample, such as the vibration collected and recognized when a vehicle passes by; represents the cosine similarity between positive sample pairs, that is, the cosine similarity between the feature vectors of the construction vibration sample and the construction vibration enhanced sample, represents the cosine similarity between negative sample pairs, that is, the cosine similarity between the feature vectors of the construction vibration sample and the non-construction vibration sample.

[0037] Specifically, is used to maximize the feature similarity between positive sample pairs (construction vibration sample and construction vibration enhanced sample). When the similarity between the two is high, the cosine value approaches 1 and the loss value approaches 0. is used to minimize the similarity between negative sample pairs to ensure that the feature vectors of the construction vibration sample and the non-construction vibration sample are clearly separated in the feature space. This contrastive learning loss function is accumulated in all time series number dimensions, so that the two-stage recognition model can learn a stable time series feature mapping and strengthen the discrimination of different vibration types.

[0038] A3-4: Determine whether the batch corresponding to the time window meets the predefined model convergence conditions, and the convergence conditions include but are not limited to: the loss value is lower than a preset threshold, the loss change range is less than a set range, or the number of training epochs reaches a preset upper limit; A3-5. If the convergence condition is not met, perform backpropagation based on the contrastive learning loss to calculate the model gradients, and use the model gradients to update the model parameters, then continue to execute steps A3-1 to A3-4; A3-6. If the convergence condition is met, stop the model training, freeze the current parameters, and obtain the first sub-model.

[0039] In this embodiment, the contrastive learning model is iteratively updated with time windows as training batches. Each batch contains K pairs of positive and negative samples, and the contrastive learning loss function is combined to optimize the measurement of feature similarity. By maximizing the feature similarity between the construction vibration samples and their augmented samples and minimizing the feature similarity between them and the non-construction samples, the ability to effectively distinguish different types of vibration events in the feature space can be formed. At the same time, introducing the time series number as the aggregation dimension of the loss function helps to improve the sensitivity of the model to potential semantic changes in continuous vibration samples, thereby more stably identifying the feature differences between real ground-breaking behaviors and false alarm events.

[0040] In this embodiment, the iterative training of the second sub-model: B1. Construct a policy discrimination sample training set according to the contrastive learning sample training set; B2. Select an initialized policy discrimination model; B3. Input the feature-label pairs of the policy discrimination sample training set into the initialized policy discrimination model, and iteratively update the model parameters until the policy discrimination model meets the predefined convergence condition.

[0041] B4. When the policy discrimination model meets the convergence condition, freeze the current model parameters to obtain the second sub-model.

[0042] In this embodiment, the training of the second sub-model is based on the vibration features output by the first sub-model, constructs a policy discrimination sample training set containing the vibration features and their corresponding labels, and completes lightweight classification learning. Through iterative optimization, the policy discrimination model can accurately judge whether it is an alarm event according to the vibration features, so as to achieve efficient discrimination between construction and non-construction vibrations, and thus obtain the second sub-model.

[0043] Specifically in step B1, the steps for obtaining the policy discrimination sample training set are as follows: B1-1. According to the time series numbers of the contrastive learning sample training set, extract two independent and different historical vibration events within the time window; B1-2. Input the two independent and different historical vibration events into the first sub-model respectively to extract the vibration features of the different historical vibration events; B1-3. After marking the historical vibration events within the time window on the spatial map, perform label annotation on the two independent and different historical vibration events; The strategy for label annotation is as follows: If it corresponds to a construction vibration event, it is labeled as an alarm; otherwise, it is labeled as no alarm. B1-4. Combine each vibration feature with the corresponding label to construct two independent and different feature-label pairs within the corresponding time series number. B1-5. Repeat B1-1 to B1-4 to construct several strategy discrimination sample training sets with time series numbers. Among them, the strategy discrimination sample training set contains several strategy discrimination samples. Each strategy discrimination sample is represented in the structure of a feature-label pair, and each time number corresponds to two independent and different feature-label pairs.

[0044] In this embodiment, the time series numbers of the contrast learning sample training set are organized so that the historical vibration events extracted within each time window have clear time positioning. Further, these historical vibration events are input into the first sub-model to extract discriminative vibration features, and combined with the spatial map and the manual or system annotation rules to form accurate alarm labels, thus forming a strategy discrimination sample training set with time series numbers.

[0045] Specifically, in step B3, the steps for iteratively updating the model parameters are as follows: B3-1. Extract the feature-label pairs corresponding to the time window batch from the strategy discrimination training set. B2-2. Input the feature-label pairs corresponding to the time window batch into the strategy discrimination model. B2-3. Predict the alarm probability according to the vibration features of the feature-label pairs. B2-4. Convert the alarm probability into a binary classification prediction value. B2-5. Calculate the classification loss between the binary classification prediction value and the sample label, and perform backpropagation with the goal of minimizing the classification loss to update the model parameters. Specifically, the classification loss can adopt the cross-entropy loss function, the squared loss function, or other supervised learning loss functions that can measure the prediction error.

[0046] B2-6. Determine whether the current training round meets the predefined convergence condition. B2-7. If the convergence condition is met, stop the model training, freeze the current parameters, and obtain the second sub-model.

[0047] In this embodiment, the policy discrimination model is trained based on the vibration features extracted by the first sub-model, and the mapping relationship between the features and the labels is constructed through supervised learning. During the training process, the model predicts whether it is an alarm event caused by construction according to the structure of the feature-label pair, and continuously optimizes the model parameters through the classification loss function. When the model reaches stable convergence in multiple batches corresponding to time windows, the second sub-model for identifying and classifying vibrations can be obtained.

[0048] A self-learning method for reducing cable false alarms in this embodiment introduces a contrastive learning structure, uses time series slice samples to construct positive and negative sample pairs, and automatically learns the similarity features of construction vibration samples under time series continuity and the difference features of non-construction samples, so as to train the first sub-model with stronger feature expression ability and improve the model's discrimination ability for samples with fuzzy boundaries; further, using the vibration features obtained through contrastive learning as input, a lightweight policy discrimination model is constructed to complete the efficient classification and discrimination task of whether to alarm. Compared with the method of directly training a classification model on the original signal, the present invention can shorten the training time and reduce the false alarm rate of the cable.

[0049] The embodiment of the present invention also provides a self-learning system for reducing cable false alarms. This system is used to implement the above method embodiment, and the parts that have been described will not be repeated here. The following terms such as "module", "unit", "sub-unit", etc. can be a combination of software and / or hardware that can achieve a predetermined function. Although the system described in the following embodiments is preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0050] As Figure 3 shown, Figure 3 is a structural block diagram of a self-learning system for reducing cable false alarms according to the present invention. The system includes: A real-time vibration feature acquisition module, which is used to acquire real-time vibration events to be identified and intercept the real-time vibration features of the real-time vibration events; A two-stage forward recognition module, which is used to input the real-time vibration features into the two-stage recognition model and perform two-stage forward propagation on the real-time vibration features; The two-stage forward propagation includes: Receiving the real-time vibration features as the input of the first sub-model and outputting the feature representation of the real-time vibration features; Receiving the feature representation of the real-time vibration features as the input of the second sub-model and outputting a binary classification instruction of whether to alarm; Among them, the two-stage recognition model is encapsulated by connecting a first sub-model with frozen parameters in series to a second sub-model; the first sub-model is obtained through iterative training of a contrastive learning model, the second sub-model is obtained through iterative training of a policy discrimination model, and the training set of the second sub-model is generated by splitting and annotating the training set of the first sub-model. A cable alarm response module, configured to trigger or inhibit a cable alarm response according to a binary classification instruction of whether to alarm.

[0051] In the above system, by means of contrastive learning in the previous stage and classification recognition in the subsequent stage, the problem that existing cable alarm devices are prone to false alarms is solved.

[0052] The embodiments of the present invention further provide related devices for a self-learning method for reducing cable false alarms. The related devices are used to execute, apply, or run a self-learning method for reducing cable false alarms. Exemplarily, the related devices include an electronic device, such as Figure 4 As shown, the electronic device includes a memory 230 and a processor 210. The memory 230 stores at least one computer-executable instruction, and the processor 210 is configured to run the computer-executable instruction. When the computer-executable instruction is run by the processor 210, the above-mentioned self-learning method for reducing cable false alarms is implemented.

[0053] The electronic device may include a processor 210, a communication interface 220, a memory 230, and a communication bus 240. Among them, the processor 210, the communication interface 220, and the memory 230 communicate with each other through the communication bus 240. The processor 210 can call the logical instructions in the memory 230 to execute a self-learning method for reducing cable false alarms disclosed in this embodiment.

[0054] In addition, when the logical instructions in the above-mentioned memory 230 are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in the various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, RandomAccessMemory), a magnetic disk, or an optical disc that can store program codes.

[0055] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application.

Claims

1. A self-learning method for reducing cable false alarms, characterized in that: include: Acquire a real-time vibration event to be identified, and intercept the real-time vibration feature of the real-time vibration event; Inputting the real-time vibration features into a two-stage recognition model, and performing a two-stage forward propagation on the real-time vibration features; The two-stage forward propagation includes: receiving the real-time vibration feature as an input of the first sub-model, and outputting a feature representation of the real-time vibration feature; receiving the feature representation of the real-time vibration feature as an input of the second sub-model, and outputting a binary classification instruction of whether to alarm; The two-stage recognition model is composed of a first sub-model with frozen parameters connected in series with a second sub-model; the first sub-model is obtained by iterative training of a contrastive learning model, the second sub-model is obtained by iterative training of a strategy discrimination model, and the training set of the second sub-model is generated by splitting and annotating the training set of the first sub-model; The cable alarm response is triggered or suppressed according to the binary instruction of whether to alarm.

2. A self-learning method for reducing cable false alarms according to claim 1, characterized in that: The iterative training steps of the first sub-model include: A1. Obtain a sample training set for comparative learning; The contrastive learning sample training set includes a number of contrastive learning samples, each of which is represented by a positive and negative sample pair structure, and each positive and negative sample pair is represented as a historical shock event of a four-tuple structure; A2. Select an initialized contrastive learning model; A3, inputting the positive and negative sample pairs of the contrastive learning sample training set into the initialized contrastive learning model, and iteratively updating the model parameters until the contrastive learning model meets the predefined convergence conditions; A4. When the contrastive learning model satisfies the convergence condition, freeze the current model parameters to obtain the first sub-model.

3. A self-learning method for reducing cable false alarms according to claim 2, characterized in that: Obtain a comparative learning sample training set, including: A1-1. Select several time windows on the historical time axis; at least two independent and different historical vibration events occur in each of the time windows; the different historical vibration events are construction vibration events and non-construction vibration events; A1-3. Marking the historical earthquake events within the time window on a spatial map; wherein the spatial map is used to represent the occurrence locations of the historical earthquake events in the geographic space; A1-4. Cut the time window based on the preset time granularity and cut the historical vibration event into vibration samples of K consecutive time segments; A1-5. Construct K groups of positive and negative sample pairs in each time window based on the vibration samples of K consecutive time segments in the time window; A1-6. Using the central timestamps of several time windows as time series numbers, obtain a group of positive and negative sample pairs, and construct a contrastive learning sample training set with time series numbers.

4. A self-learning method for reducing cable false alarms according to claim 3, characterized in that: According to the vibration samples of K consecutive time segments in the time window, K groups of positive and negative sample pairs in each time window are constructed, including: A1-5-1. Mark the spatial center of each historical earthquake event within the time window on the spatial map, where the spatial center corresponds to the main epicenter position or the groundbreaking center of the historical earthquake event; A1-5-2. Set a collection radius with the center of each space as the collection circle center; the minimum value of the collection radius is limited to: outside the area boundary of the collection radius, the response intensity of the collection device to the vibration event is lower than the lowest identifiable threshold; A1-5-3. Extract from the time window within the collection radius: K consecutive construction vibration samples corresponding to the construction vibration event; The non-construction vibration event corresponds to K consecutive non-construction vibration samples; A1-5-4. Perform data disturbance enhancement on K consecutive construction vibration samples to obtain K construction vibration enhanced samples; A1-5-5. For each time segment : Pair the construction vibration samples with the construction vibration enhancement samples to obtain positive sample pairs; Pair the construction vibration samples with the non-construction vibration samples to obtain negative sample pairs; A1-5-6. Repeat A1-5-1 to A1-5-5, and construct the K time segments in each time window into positive sample pairs and negative sample pairs in turn to generate K groups of positive and negative sample pairs.

5. A self-learning method for reducing cable false alarms according to claim 2, characterized in that: The contrastive learning samples of the contrastive learning sample training set are input into the initialized contrastive learning model, and the model parameters are iteratively updated until the contrastive learning model meets the predefined convergence conditions, including: A3-1. Select K groups of positive and negative sample pairs from the contrastive learning sample training set in sequence based on the time window as the current training batch; A3-2, inputting the K groups of positive and negative sample pairs of the current training batch into the initialized contrastive learning model, performing forward propagation, and outputting K quadruple prediction vectors of the current training batch; wherein the quadruple prediction vectors include positive prediction vector pairs and negative prediction vector pairs; A3-3. Based on the predefined contrastive learning loss function, calculate the contrastive learning loss of the batch corresponding to the time window; The loss function is used to maximize the similarity between the positive sample prediction vector pairs and minimize the similarity between the negative sample prediction vector pairs; A3-4. Determine whether the batch corresponding to the time window meets the predefined model convergence conditions, which include but are not limited to: the loss value is lower than a preset threshold, the loss change amplitude is less than a set range, or the number of training rounds reaches a preset upper limit; A3-5. If the convergence condition is not met, back propagation is performed based on contrastive learning loss to calculate the model gradient, and the model parameters are updated using the model gradient, and then steps A3-1 to A3-4 are continued; A3-6. If the convergence condition is met, stop model training and freeze the current parameters to obtain the first sub-model.

6. A self-learning method for reducing cable false alarms according to claim 4, characterized in that: The iterative training steps in the second sub-model include: B1. Construct a strategy discrimination sample training set based on the comparative learning sample training set; B2. Select an initialized strategy discrimination model; B3. Input the feature label pairs of the strategy discrimination sample training set into the initialized strategy discrimination model, and iteratively update the model parameters until the strategy discrimination model meets the predefined convergence conditions; B4. When the strategy discrimination model meets the convergence condition, the current model parameters are frozen to obtain the second sub-model.

7. A self-learning method for reducing cable false alarms according to claim 4, characterized in that: According to the contrastive learning sample training set, the strategy discrimination sample training set is obtained, including: B1-1, extract two independent and different historical shock events within the time window according to the time sequence number of the comparative learning sample training set; B1-2, inputting the two independent and different historical shock events into the first sub-model respectively, and extracting the shock characteristics of the different historical shock events; B1-3, after marking the historical earthquake events within the time window on the spatial map, labeling two independent and different historical earthquake events; The label marking strategy is: if it corresponds to a construction vibration event, it is marked as an alarm; otherwise, it is marked as no alarm; B1-4, combining each vibration feature with the corresponding label to construct two independent and different feature label pairs within the corresponding time series number; B1-5. Repeat B1-1 to B1-4 to construct several strategy discrimination sample training sets with time series numbers; Among them, the strategy discrimination sample training set contains several strategy discrimination samples, each strategy discrimination sample is represented by the structure of feature label pairs, and each time number corresponds to two independent and different feature label pairs.

8. A self-learning method for reducing cable false alarms according to claim 4, characterized in that: The feature label pairs of the strategy discrimination sample training set are input into the initialized strategy discrimination model, and the model parameters are iteratively updated until the strategy discrimination model meets the predefined convergence conditions, including: B3-1. Extracting feature label pairs of batches corresponding to the time window from the strategy discrimination training set; B2-2. Input the feature label pairs of the batch corresponding to the time window into the strategy discrimination model; B2-3, predict the alarm probability based on the vibration characteristics of the feature label pair; B2-4, convert the alarm probability into a binary prediction value; B2-5. Calculate the classification loss between the binary classification prediction value and the sample label, and perform back propagation to update the model parameters with the goal of minimizing the classification loss; B2-6, determine whether the current training round meets the predefined convergence conditions; B2-7. If the convergence condition is met, stop model training and freeze the current parameters to obtain the second sub-model.

9. A self-learning system for reducing cable false alarms, characterized in that: The system comprises: A real-time vibration feature acquisition module, used to acquire a real-time vibration event to be identified and intercept the real-time vibration feature of the real-time vibration event; A two-stage forward recognition module, used for inputting the real-time vibration features into a two-stage recognition model and performing two-stage forward propagation on the real-time vibration features; The two-stage forward propagation includes: receiving the real-time vibration feature as an input of the first sub-model, and outputting a feature representation of the real-time vibration feature; receiving the feature representation of the real-time vibration feature as an input of the second sub-model, and outputting a binary classification instruction of whether to alarm; The two-stage recognition model is composed of a first sub-model with frozen parameters connected in series with a second sub-model; the first sub-model is obtained by iterative training of a contrastive learning model, the second sub-model is obtained by iterative training of a strategy discrimination model, and the training set of the second sub-model is generated by splitting and annotating the training set of the first sub-model; The cable alarm response module is used to trigger or suppress the cable alarm response according to the binary instruction of whether to alarm.

10. An electronic device, characterized in that: It includes a memory and a processor, the memory stores at least one computer executable instruction, the processor is configured to run the computer executable instruction, and when the computer executable instruction is run by the processor, a self-learning method for reducing cable false alarms as described in any one of claims 1 to 8 is implemented.

Images