Intelligent ground lock anomaly detection method and system based on generative adversarial network

Abstract

The present invention relates to the technical field of smart ground locks, and in particular to a smart ground lock anomaly detection method and system based on a generative adversarial network. The method collects ground lock design information of a target ground lock series, extracts model intrinsic features of each series model based on the ground lock design information, and obtains a model intrinsic feature set; cluster analysis is performed based on the model intrinsic feature set, and multiple virtual models are correspondingly defined. Real abnormal sample data corresponding to each virtual model is collected and output as a real training data set; the sample generation boundary of each virtual model is defined in combination with the ground lock design information, and a generation filter is configured based on the sample generation boundary; the real training data set and the generation filter are combined to construct and train a corresponding generation adversarial network for each virtual model, and perform anomaly detection of smart ground locks based on the generation adversarial network, thereby achieving the technical effect of improving the reliability and practicality of anomaly detection of smart ground locks.

Description

Technical Field

[0001] The present invention relates to the technical field of smart ground locks, and in particular to a smart ground lock anomaly detection method and system based on a generative adversarial network. Background Art

[0002] With the development of smart parking technology, smart ground locks, as a crucial piece of equipment for parking lot management, are attracting increasing attention for their functional stability and safety. Different models of smart ground locks differ in product structure, electrical parameters, and sensing devices, resulting in diverse manifestations of abnormal conditions. Traditional anomaly detection methods are often designed for a single model of ground lock and rely on large amounts of labeled anomaly sample data. However, when faced with a wide variety of models, dispersed samples, or insufficient sample size, detection accuracy and generalization capabilities are poor, making it difficult to adapt to the diverse needs of smart ground lock anomaly detection. Technical issues exist, such as poor reliability and practicality of smart ground lock anomaly detection. Summary of the Invention

[0003] The present invention aims to solve the technical problems of poor reliability and practicality of intelligent ground lock anomaly detection in the prior art by providing an intelligent ground lock anomaly detection method and system based on a generative adversarial network.

[0004] The technical solution of the present invention to solve the above technical problems is as follows:

[0005] In the first aspect, the present invention provides a method for detecting anomalies of smart ground locks based on a generative adversarial network, comprising: collecting ground lock design information of a target ground lock series, and extracting the model intrinsic features of each series model according to the ground lock design information to obtain a model intrinsic feature set; performing cluster analysis based on the model intrinsic feature set, defining multiple virtual models accordingly, and collecting real abnormal sample data corresponding to each of the virtual models, and outputting them as a real training data set; defining the sample generation boundary of each of the virtual models in combination with the ground lock design information, and configuring a generation filter based on the sample generation boundary; combining the real training data set and the generation filter, respectively constructing and training a corresponding generation adversarial network for each virtual model, and performing anomaly detection of smart ground locks based on the generation adversarial network.

[0006] Optionally, the ground lock design information of the target ground lock series is collected, and the model intrinsic features of each series of models are extracted based on the ground lock design information to obtain a model intrinsic feature set, including: traversing the target ground lock series to obtain the ground lock design information of each series of models, wherein the ground lock design information includes product structure information, electrical parameter configuration and sensing device specifications; based on preset structural feature extraction rules, extracting the feature fields of the product structure information and vectorizing them to obtain a structural feature vector; performing standardized conversion on the electrical parameter configuration and the sensing device specifications, fusing the standardized conversion results and correspondingly defining a parameter feature vector; splicing the structural feature vector and the parameter feature vector to obtain a model structural feature vector, and iteratively obtaining the model structural feature vector of each series of models, and storing them as the model intrinsic feature set.

[0007] Among them, cluster analysis is performed based on the model intrinsic feature set, multiple virtual models are defined accordingly, and real abnormal sample data corresponding to each of the virtual models is collected and output as a real training data set, including: performing density-based unsupervised cluster analysis on the model intrinsic feature set to obtain multiple clusters; defining each cluster as a virtual model, and establishing a model mapping table according to the cluster boundaries; based on the model mapping table, traversing the series models corresponding to each of the virtual models, collecting historical abnormal status data and standardizing it, and outputting the structured data as the real training data set.

[0008] Among them, in combination with the ground lock design information, the sample generation boundary of each virtual model is defined, and a generation filter is configured based on the sample generation boundary, including: defining the structural freedom boundary according to the product structure information corresponding to the virtual model; extracting the electrical performance indicator boundary according to the electrical parameter configuration corresponding to the virtual model; determining the sensor perception indicator boundary according to the sensing device specification corresponding to the virtual model; using the structural freedom boundary, the electrical performance indicator boundary and the sensor perception indicator boundary as the sample generation boundary, correspondingly executing interval mapping and constructing a threshold detection function to form a generation filtering rule; configuring the generation filter for boundary detection according to the generation filtering rule.

[0009] Among them, in combination with the real training data set and the generation filter, a corresponding generative adversarial network is constructed and trained for each virtual model, including: taking each virtual model as a target, constructing a sample generator and anomaly discriminator with an adversarial structure, and connecting the generation filter to the output end of the sample generator; based on the real training data set, the sample generator is subjected to feature modeling of abnormal sample data; through adversarial training, the sample generator and the anomaly discriminator are iteratively optimized until the preset iteration constraints are met, and the generative adversarial network containing the sample generator and the anomaly discriminator is output.

[0010] Among them, the abnormality detection of the smart floor lock based on the generative adversarial network includes: matching the generative adversarial network corresponding to the series model of the target smart floor lock according to the model mapping table, and extracting the corresponding abnormality discriminator; obtaining the real-time status data of the target smart floor lock, and synchronously inputting it into the generative filter and the abnormality discriminator for 2PASS abnormality detection.

[0011] Among them, the real-time status data of the target smart lock is obtained and synchronously input into the generation filter and the abnormality discriminator for 2PASS abnormality detection, including: performing boundary detection on the real-time status data according to the generation filter, if any item in the real-time status data fails, then the abnormality detection result is output as the first type of abnormality, wherein the first type of abnormality includes at least sensing abnormality and transmission abnormality; if the real-time status data all pass the generation filter, then the real-time status data is input into the abnormality discriminator, and the corresponding abnormality detection result is the second type of abnormality, wherein the second type of abnormality includes at least abnormality code and abnormal position.

[0012] In a second aspect, the present invention provides an intelligent ground lock anomaly detection system based on a generative adversarial network, comprising:

[0013] An intrinsic feature collection module is used to collect the ground lock design information of the target ground lock series, and extract the model intrinsic features of each series model according to the ground lock design information to obtain a model intrinsic feature set;

[0014] The intrinsic feature clustering module is used to perform cluster analysis based on the model intrinsic feature set, define multiple virtual models, and collect real abnormal sample data corresponding to each virtual model, and output it as a real training data set;

[0015] a generation boundary configuration module, configured to define a sample generation boundary of each virtual model in combination with the ground lock design information, and configure a generation filter based on the sample generation boundary;

[0016] The ground lock anomaly detection module is used to combine the real training data set with the generated filter to construct and train a corresponding generative adversarial network for each virtual model, and perform anomaly detection of the smart ground lock based on the generative adversarial network.

[0017] By implementing the present invention, it is possible to collect the design information of the target ground lock series, extract the model intrinsic features of each model in the series based on the ground lock design information, obtain the model intrinsic feature set, extract key information such as product structure, electrical parameters and sensing device and convert it into feature vectors, and convert the physical properties and technical parameters of the ground locks into a computable and analyzable data form, providing standardized basic data for subsequent cluster analysis and model training, ensuring that the features of different models of ground locks are comparable and consistent;

[0018] By implementing the present invention, cluster analysis can be performed based on the model's intrinsic feature set, corresponding to the definition of multiple virtual models, and correspondingly collecting real abnormal sample data corresponding to each of the virtual models, outputting them as a real training data set, thus solving the problem of sample dispersion caused by the large number of models. By collecting abnormal samples corresponding to virtual models, abnormal data of similar models can be integrated, expanding the training sample volume, improving the data reliability and generalization ability of subsequent model training, and at the same time establishing a model mapping table to facilitate the subsequent rapid matching of corresponding models and samples.

[0019] By implementing the present invention, it is possible to define the sample generation boundary of each virtual model in combination with the ground lock design information, and configure a generation filter based on the sample generation boundary. The definition of the sample generation boundary is based on the actual design parameters of the ground lock, such as structural degrees of freedom, electrical performance, and sensor indicators, to ensure that the generated samples meet the physical and technical limitations of the ground lock. The generation filter can filter out abnormal samples that do not meet the boundary conditions, ensure the validity and rationality of the subsequent generation of adversarial network training data, and avoid the interference of invalid samples on model performance.

[0020] By implementing the present invention, it is possible to combine the real training data set with the generated filter to construct and train a corresponding generative adversarial network for each virtual model, and perform anomaly detection of smart ground locks based on the generative adversarial network. The generative adversarial network can learn the characteristic distribution of abnormal samples through adversarial training of the generator and the discriminator, improve the ability to recognize abnormal states, and construct a model separately for each virtual model, which can adapt to the characteristics of different types of ground locks and improve the accuracy of detection.

[0021] In summary, by implementing the present invention, the technical effect of improving the reliability and practicality of abnormality detection of smart ground locks can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1A schematic diagram of a flow chart of a method for detecting anomalies of intelligent ground locks based on a generative adversarial network provided by the present invention;

[0023] Figure 2 This is a structural diagram of an intelligent ground lock anomaly detection system based on a generative adversarial network provided by the present invention.

[0024] In the accompanying drawings, the components represented by the reference numerals are as follows:

[0025] Intrinsic feature collection module 11, intrinsic feature clustering module 12, boundary configuration generation module 13, ground lock anomaly detection module 14. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0027] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Therefore, features specified as "first" or "second" may explicitly or implicitly include one or more features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0028] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or explanation". Any embodiment of the present invention described as "for example" is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed herein.

[0029] Example 1, as Figure 1 As shown, an embodiment of the present invention provides a method for detecting abnormalities of intelligent ground locks based on a generative adversarial network, comprising:

[0030] S100: collecting design information of a target ground lock series, and extracting model intrinsic features of each model in the series based on the ground lock design information to obtain a model intrinsic feature set;

[0031] S200: performing cluster analysis based on the model intrinsic feature set, defining multiple virtual models, and collecting real abnormal sample data corresponding to each virtual model, and outputting the data as a real training data set;

[0032] S300: Defining a sample generation boundary for each virtual model in combination with the ground lock design information, and configuring a generation filter based on the sample generation boundary;

[0033] S400: Combining the real training data set and the generated filter, respectively constructing and training a corresponding generative adversarial network for each virtual model, and performing anomaly detection of the smart ground lock based on the generative adversarial network.

[0034] In step S100 of the embodiment of the present application, the design information of the target ground lock series is collected, and the model intrinsic features of each series model are extracted based on the ground lock design information to obtain a model intrinsic feature set, including:

[0035] Traversing the target ground lock series to obtain ground lock design information for each series model, wherein the ground lock design information includes product structure information, electrical parameter configuration, and sensing device specifications;

[0036] Based on a preset structural feature extraction rule, the characteristic fields of the product structure information are extracted and vectorized to obtain a structural feature vector;

[0037] Performing standardized conversion on the electrical parameter configuration and the sensing device specifications, fusing the standardized conversion results and correspondingly defining parameter feature vectors;

[0038] The structural feature vector and the parameter feature vector are concatenated to obtain a model structural feature vector, and the model structural feature vector of each series of models is iteratively obtained and stored as the model intrinsic feature set.

[0039] Step S100 of the embodiment of the present application is the basic data processing link of the entire anomaly detection method. The core purpose is to convert the physical properties and technical parameters of smart locks of different series and models into unified and computable feature data, providing standardized input for subsequent clustering analysis, model training, etc.

[0040] First, it is necessary to traverse the target ground lock series to obtain the ground lock design information of each series model, wherein the ground lock design information includes product structure information, electrical parameter configuration and sensing device specifications, specifically including: product structure information, such as mechanical structure, size parameters, etc.; electrical parameter configuration, such as operating voltage, standby current, power, etc.; sensing device specifications, such as sensor type, detection range, accuracy, etc.

[0041] For example, the product structure information of model A ground lock is: material is aluminum alloy, size is 30cm×20cm×15cm, and the lifting structure is an electric push rod; the electrical parameter configuration is: working voltage 12V, power 50W, standby current 0.5A; the sensing device specifications are: infrared sensor, detection distance 0-5m, accuracy ±2cm.

[0042] The product structure information of model B ground lock is as follows: material is stainless steel, size is 40cm×30cm×20cm, and the lifting structure is a hydraulic rod; the electrical parameters are configured as working voltage 24V, power 100W, and standby current 1.0A; the sensing device specifications are ultrasonic sensors, detection distance 0-10m, and accuracy ±5cm.

[0043] The product structure information of the C model ground lock is: material is aluminum alloy, size is 35cm×25cm×18cm, and the lifting structure is an electric push rod; the electrical parameter configuration is: working voltage 12V, power 60W, standby current 0.6A; the sensing device specifications are: infrared sensor, detection distance 0-6m, accuracy ±3cm.

[0044] Next, based on preset structural feature extraction rules, the characteristic fields of the product structure information are extracted and vectorized to obtain a structural feature vector. For example, the feature extraction rules may include the characteristic fields including material, length, width, height, and lift structure type, with aluminum alloy = 1 and stainless steel = 2 for the material; electric push rod = 1 and hydraulic rod = 2 for the lift structure; and dimensional parameters are directly taken as numerical values ​​in cm. The structural feature vector of model A ground lock is [1, 30, 20, 15, 1]. The same method can be used to obtain structural feature vectors for multiple models of ground locks, for example, the structural feature vector of model B is [2, 40, 30, 20, 2], and the structural feature vector of model C is [1, 35, 25, 18, 1].

[0045] Furthermore, it is necessary to perform standardized conversion on the electrical parameter configuration and the sensing device specifications, fuse the standardized conversion results and define the parameter feature vector accordingly.

[0046] The standardization of electrical parameter configuration and the specifications of the sensing device can be specifically as follows: working voltage: uniformly coded as "12V=1, 24V=2"; power is normalized to the range of 0-1, and the specific normalization method is: normalized value = (current power - maximum power) / (maximum power - minimum power). For example, for three types of ground locks with power of 50W, 60W, and 100W, if the minimum value is 50W, the normalized value is (50-50) / (100-50) = 0, and similarly. 100W is normalized to 1, and 60W is normalized to 0.2. Based on the same logic as power normalization, the standby current is normalized to the range of 0-1. Model A 0.5A is normalized to 0, Model B 1.0A is normalized to 1, and Model C 0.6A is normalized to 0.2. Based on the same logic as power normalization, the detection distance is normalized to the range of 0-1. Model A 5m is normalized to 0, Model B 10m is normalized to 1, and Model C 6m is normalized to 0.2. For sensor type, infrared = 1, ultrasonic = 2. Based on the same logic as power normalization, the accuracy is normalized after taking the absolute value. Model A 2cm is normalized to 0, Model B 5cm is normalized to 1, and Model C 3cm is normalized to 0.33.

[0047] Based on the above steps, it can be obtained that the parameter feature vector of model A ground lock is [1, 0, 0, 1, 0, 0], the parameter feature vector of model B ground lock is [2, 1, 1, 2, 1, 1], and the parameter feature vector of model C ground lock is [1, 0.2, 0.2, 1, 0.2, 0.33]. Then, the structural feature vector and the parameter feature vector are spliced ​​to obtain the model structural feature vector, and the model structural feature vector of each series of models is iteratively obtained and stored as the model intrinsic feature set.

[0048] For example, the structural feature vectors and parameter feature vectors of each model of ground lock are concatenated to obtain the complete feature vector of each model: Model A [1, 30, 20, 15, 1, 1, 0, 0, 1, 0, 0], Model B [2, 40, 30, 20, 2, 2, 1, 1, 2, 1, 1], Model C [1, 35, 25, 18, 1, 1, 0.2, 0.2, 1, 0.2, 0.33].

[0049] In step S200 of the embodiment of the present application, cluster analysis is performed based on the model intrinsic feature set, multiple virtual models are defined accordingly, and real abnormal sample data corresponding to each virtual model is collected and output as a real training data set, including:

[0050] Performing density-based unsupervised cluster analysis on the model intrinsic feature set to obtain multiple clusters;

[0051] Define each cluster as a virtual model, and establish a model mapping table based on the cluster boundaries;

[0052] Based on the model mapping table, the series models corresponding to each virtual model are traversed, historical abnormal state data is collected and standardized, and the structured output is the real training data set.

[0053] In step S200 of the embodiment of the present application, this step is to integrate similar models through cluster analysis on the basis of the model intrinsic feature set, so as to solve the problem of scattered abnormal samples and insufficient sample size of multiple models of ground locks, and provide high-quality unified training data for subsequent model training.

[0054] First, it is necessary to perform density-based unsupervised clustering analysis on the model's intrinsic feature set to obtain multiple clusters. Specifically, an unsupervised clustering algorithm such as DBSCAN (density-based spatial clustering application) can be used to analyze the feature vectors in the model's intrinsic feature set. That is, the distance between each feature vector is calculated, such as the Euclidean distance, and vectors with similar density, that is, similar features, are clustered into one category.

[0055] For example, Model A (aluminum alloy, electric push rod, infrared sensor) has more similar features to Model C (aluminum alloy, electric push rod, infrared sensor), and is therefore clustered into one cluster. Model B (stainless steel, hydraulic rod, ultrasonic sensor) has significantly different features and is therefore clustered into its own cluster. Ultimately, two clusters are obtained.

[0056] Then, each cluster needs to be defined as a virtual model. For example, cluster 1 (A, C) is defined as virtual model X; cluster 2 (B) is defined as virtual model Y.

[0057] Based on the cluster boundaries, such as the maximum and minimum value ranges of the feature vectors, a model mapping table between virtual models and actual series models is established. For example, virtual model X corresponds to models A and C, and virtual model Y corresponds to model B.

[0058] Finally, real abnormal sample data is collected and standardized, and the real training data set is output.

[0059] Based on the model mapping table, traverse the A and C models corresponding to the virtual model X and collect historical abnormal status data. For example, the historical abnormal status data of model A includes sensor false alarms and lift jams. The historical abnormal status data of model C includes voltage instability and detection deviation.

[0060] The collected abnormal data is standardized, such as by unifying the data format, converting units, and normalizing values. The standardized abnormal data is then stored in a structured form, such as a table, containing the abnormality type, characteristic parameters, timestamp, etc., to form the real training data of virtual model X. Similarly, the real training data of virtual model Y is obtained, and finally merged into a complete real training data set.

[0061] Through the above process, the originally scattered abnormal samples of models A, B, and C are integrated into training data for virtual models X and Y. This not only retains the common characteristics of similar models, but also expands the number of single-class samples, laying a data foundation for subsequent training of generative adversarial networks.

[0062] In step S300 of the embodiment of the present application, the sample generation boundary of each virtual model is defined in combination with the ground lock design information, and a filter is generated based on the sample generation boundary configuration, including:

[0063] Defining a structural freedom degree boundary according to the product structure information corresponding to the virtual model;

[0064] Extracting electrical performance indicator boundaries according to the electrical parameter configuration corresponding to the virtual model;

[0065] Determining a sensor perception indicator boundary according to the sensing device specifications corresponding to the virtual model;

[0066] Taking the structural freedom degree boundary, the electrical performance index boundary and the sensor perception index boundary as the sample generation boundary, correspondingly performing interval mapping and constructing a threshold detection function to form a generation filtering rule;

[0067] The generation filter for boundary detection is configured according to the generation filter rule.

[0068] In the embodiment of the present application, the core purpose of step S300 is to define a reasonable range for sample generation for each virtual model, and by configuring the generation filter, ensure that the samples involved in the subsequent generation adversarial network training process conform to the physical characteristics and technical limitations of the ground lock, avoid invalid or unreasonable samples interfering with model training, thereby improving the accuracy and reliability of anomaly detection.

[0069] First, based on the product structure information corresponding to the virtual model, it is necessary to determine the reasonable range of the structural level, namely the structural freedom boundary. For example, for virtual model X, models A and C are both made of aluminum alloy, and the lifting structure is an electric push rod. Then, it is necessary to extract key parameters from the product structure information, such as the size range of model A: 30cm×20cm×15cm, model C: 35cm×25cm×18cm; the lifting stroke of model A: 0-15cm; model C: 0-18cm. Next, the structural freedom boundary is defined, for example: length 30-35cm, width 20-25cm, height 15-18cm, and lifting stroke 0-18cm. The structural freedom boundary is set by taking the maximum range of similar models.

[0070] For the virtual model Y that only includes the B model, the structural freedom boundaries can be based on a small range of fluctuation values ​​according to its basic structural information. For example, the structural freedom boundaries of the virtual model Y are: length 40cm±2cm, width 30cm±2cm, height 20cm±2cm, and the hydraulic rod extension range is 0-20cm.

[0071] Based on the same logic, the structural freedom boundaries of multiple virtual models are obtained.

[0072] Furthermore, it is necessary to configure the electrical parameters corresponding to the virtual model. Specifically, small fluctuations can be allowed within the maximum and minimum ranges of the electrical performance of each model of ground lock contained in the virtual model to obtain the electrical performance indicator boundaries. For example, the electrical parameter configurations of the two models of ground locks in virtual model X are: Model A has an operating voltage of 12V and a power of 50W; Model C has an operating voltage of 12V and a power of 60W. The electrical performance indicator boundaries that can be extracted are operating voltage of 11-13V, power of 45-65W, and standby current of 0.4-0.7A. For virtual model Y, the electrical performance indicator boundaries can be: operating voltage of 22-26V, power of 90-110W, and standby current of 0.8-1.2A.

[0073] Furthermore, it is necessary to determine the sensor perception indicator boundary based on the sensor device specifications corresponding to the virtual model, and in specific fields, determine the sensor perception indicator boundary based on the maximum value of the sensor device specifications corresponding to the virtual model.

[0074] For example, for a virtual model X, models A and C are both infrared sensors. Model A has a detection range of 0-5m and an accuracy of ±2cm; model C has a detection range of 0-6m and an accuracy of ±3cm. The sensor perception performance boundaries for virtual model X can be: detection range 0-6m, accuracy error ≤ ±3cm, and response time ≤ 0.5s. The response time can be determined based on the sensor type.

[0075] For virtual model Y, model B is an ultrasonic sensor with a detection distance of 0-10m and an accuracy of ±5cm. The sensing perception indicator boundaries of virtual model Y can be: detection distance 0-10m, accuracy error ≤±5cm, and response time ≤0.8s.

[0076] Finally, using the structural freedom boundaries, electrical performance indicator boundaries, and sensor perception indicator boundaries as the sample generation boundaries, we perform interval mapping and construct a threshold detection function to form the generation filtering rules. To generate the sample generation boundaries, we also need to convert each boundary into a specific interval range, such as the operating voltage ∈ [11V, 13V] format.

[0077] Then, a threshold detection function is constructed based on the interval range to generate filtering rules. For example, the filtering rules for virtual model X may include: if the detected "length = 36cm" exceeds the range of 30-35cm, then the sample is considered invalid; if the "operating voltage = 14V" exceeds the range of 11-13V, then the sample is considered invalid; if the sensor accuracy error = 4cm exceeds the range of ≤±3cm, then the sample is considered invalid.

[0078] Based on the generated filtering rules, a generation filter is configured. The generation filter can be understood as a parameter verification module, which is used to automatically filter out invalid data that exceeds the boundary during subsequent sample generation or detection. The specific filtering rule configuration method for this generation filter is described above. The specific configuration method is existing in the art and will not be repeated here.

[0079] In step S400 of the embodiment of the present application, the real training data set and the generative filter are combined to construct and train a corresponding generative adversarial network for each virtual model, including:

[0080] Taking each of the virtual models as a target, respectively constructing a sample generator and an anomaly discriminator with an adversarial structure, and connecting the generation filter to the output end of the sample generator;

[0081] Performing feature modeling of abnormal sample data on the sample generator based on the real training data set;

[0082] The sample generator and the anomaly discriminator are iteratively optimized through adversarial training until preset iteration constraints are met, and the generative adversarial network including the sample generator and the anomaly discriminator is output.

[0083] In the embodiment of the present application, the core purpose of step S400 is to construct a dedicated generative adversarial network (GAN) for each virtual model. Through adversarial training between the sample generator and the anomaly discriminator, the model learns the characteristic patterns of abnormal samples of the virtual model ground lock. At the same time, the rationality of the generated samples is constrained by the generation filter, and finally a model is obtained that can accurately identify the abnormal state of the corresponding virtual model ground lock.

[0084] Taking virtual model X as an example, we first need to build a sample generator and anomaly detector with an adversarial structure and connect the generated filter to the output of the sample generator. The sample generator is used to generate simulated abnormal sample data, such as feature vectors of sensor false alarms, lifting and lowering delays, etc. The anomaly detector is built to determine whether the input sample is a real abnormal sample or a fake sample generated by the sample generator.

[0085] Next, it is necessary to perform feature modeling of the abnormal sample data of the sample generator based on the real training data set. Specifically, it is necessary to extract abnormal sample features from the real training data set of virtual model X, such as the signal fluctuation characteristics of the infrared sensor of model A when it falsely alarms, and the current change characteristics of the lifting and lowering jams of model C.

[0086] Then, the sample generator and the anomaly discriminator are iteratively optimized through adversarial training until preset iteration constraints are met, and the generative adversarial network including the sample generator and the anomaly discriminator is output.

[0087] Optionally, in the construction of the generative adversarial network, the structure of the sample generator includes an input layer, an output layer, and three hidden layers. Among them, the input layer is used to receive a 100-dimensional random noise vector; hidden layer 1 contains 128 neurons, uses a ReLU activation function, and is equipped with a batch normalization layer; hidden layer 2 contains 256 neurons, uses a ReLU activation function, and is equipped with a batch normalization layer; hidden layer 3 contains 512 neurons, uses a ReLU activation function, and is equipped with a batch normalization layer; the number of neurons in the output layer is consistent with the dimension of the model structure feature vector, such as the dimension of the vector after splicing in step S100, uses a Tanh activation function, and the output end is connected to generate a filter. The structure of the anomaly discriminator includes an input layer, an output layer, and three hidden layers.

[0088] Among them, the input layer is used to receive the feature vector with the same output dimension as the sample generator; the hidden layer 1 contains 512 neurons, uses the LeakyReLU activation function with a slope of 0.2, and is equipped with a dropout layer with a dropout rate of 0.3; the hidden layer 2 contains 256 neurons, uses the LeakyReLU activation function with a slope of 0.2, and is equipped with a dropout layer with a dropout rate of 0.3; the hidden layer 3 contains 128 neurons, uses the LeakyReLU activation function with a slope of 0.2; the output layer contains 1 neuron, uses the Sigmoid activation function, and the probability of the output sample being a true abnormal sample is between 0 and 1.

[0089] For the hyperparameter settings of the generative adversarial network, the Adam optimizer was used for both the sample generator and the anomaly discriminator. The learning rates for the sample generator and the anomaly discriminator were 0.0002, 0.0002, 0.5 for momentum parameters β1 and 0.999 for β2, and 32 for batch size. The binary cross-entropy loss function was used for both. The initial training epochs were set to 200, which could be adjusted dynamically based on convergence.

[0090] The training samples are collected based on the model mapping table, and the historical abnormal status data of the series models corresponding to each virtual model is collected. This data is then standardized and structured to form a real training dataset. This data includes information related to product structure, such as the duration of lift jams and lock body deformation deviations; electrical parameters such as voltage fluctuations and power overload values; sensor false alarm readings and signal interruption duration related to sensing devices; and abnormality type labels such as "sensor false alarm" and abnormal location labels such as "infrared module." After standardization, this data becomes the real training dataset.

[0091] At the same time, the samples generated by the sample generator and passed through the generated filter are combined. The number of samples in the real training dataset corresponding to each virtual model is not less than 5,000, and the ratio of the number of samples generated by the sample generator to the number of samples in the real training dataset is 1:1.

[0092] The specific training process is as follows: the sample generator generates a batch of false anomaly samples. After filtering through the generation filter, samples that exceed the X boundary are eliminated. The samples are mixed with some real anomaly samples in the real training data set and input into the anomaly discriminator; the anomaly discriminator outputs the "realness" judgment of each sample. The judgment result is a probability between 0 and 1, where 1 is a real anomaly and 0 is a false anomaly. The discrimination error is calculated using the cross-entropy loss function.

[0093] The sample generator adjusts its parameters based on the error feedback of the anomaly discriminator, with the goal of making the generated fake samples more difficult for the discriminator to detect, that is, making the anomaly discriminator's judgment on fake samples close to 1; the anomaly discriminator also adjusts its parameters based on the error, with the goal of more accurately distinguishing real samples from fake samples, with the judgment on real samples close to 1 and on fake samples close to 0.

[0094] Repeat the above adversarial process. When the accuracy of the anomaly detector stabilizes at a preset threshold, such as 50%, it indicates that the samples generated by the sample generator are close to the real ones, and training is stopped. The generative adversarial network containing the sample generator and the anomaly detector is obtained. Through the above steps, a corresponding generative adversarial network is configured for each virtual model.

[0095] In step S400 of the embodiment of the present application, anomaly detection of the smart ground lock is performed based on the generative adversarial network, including:

[0096] Matching the generative adversarial network corresponding to the series model of the target smart lock according to the model mapping table, and extracting the corresponding anomaly discriminator;

[0097] The real-time status data of the target intelligent ground lock is obtained and synchronously input into the generation filter and the abnormality discriminator for 2PASS abnormality detection.

[0098] In the embodiment of the present application, the core purpose of this step is to apply the trained generative adversarial network (GAN) to the actual anomaly detection scenario of smart ground locks, and quickly locate the dedicated model corresponding to the target ground lock through the model mapping table to ensure the accuracy and pertinence of the detection.

[0099] Among them, the generative adversarial network corresponding to the series model of the target smart floor lock is matched according to the model mapping table, and the corresponding abnormal discriminator is extracted. First, it is necessary to obtain the model information of the target floor lock, that is, to obtain its series model, such as model A, model B, etc., from the identification or system record of the smart floor lock.

[0100] Then, the virtual model to which the model belongs is searched based on the model mapping table established in step S200. For example, if the target lock is model A, the mapping table shows that it belongs to virtual model X, so the generative adversarial network corresponding to virtual model X is matched; if the target lock is model B, which corresponds to virtual model Y, the generative adversarial network corresponding to virtual model Y is matched.

[0101] Then, a trained anomaly discriminator is extracted separately from the matched generative adversarial network, such as the discriminator optimized through adversarial training in step S400. This discriminator has learned the characteristic patterns of normal and abnormal states of the corresponding virtual model ground lock and can be directly used for anomaly detection of the target ground lock.

[0102] In step S400 of the embodiment of the present application, the real-time status data of the target smart lock is obtained and synchronously input into the generation filter and the abnormality discriminator for 2PASS abnormality detection, including:

[0103] Performing boundary detection on the real-time status data according to the generated filter, and outputting an abnormality detection result as a first-category abnormality if any item in the real-time status data fails to pass the test, wherein the first-category abnormality includes at least a sensing abnormality and a transmission abnormality;

[0104] If the real-time status data all pass the generation filter, the real-time status data is input into the abnormality discriminator, and the corresponding abnormality detection result obtained is a second type of abnormality, wherein the second type of abnormality at least includes an abnormality code and an abnormality position.

[0105] In the embodiment of the present application, the core purpose of this step is to use the 2PASS abnormality detection mechanism to identify the abnormal status of the smart ground lock in a hierarchical and comprehensive manner, so as to quickly screen out basic abnormalities that obviously exceed the physical or technical boundaries, and accurately identify complex abnormalities that meet the boundaries but have potential problems, thereby improving the comprehensiveness and accuracy of abnormality detection.

[0106] Specifically, taking the target ground lock belonging to virtual model X as model A as an example, during the specific implementation process, it is first necessary to implement the first PASS, that is, to perform boundary detection on the real-time status data according to the generated filter. First, it is necessary to collect the real-time operating data of model A ground lock, including structural status, such as lifting height; electrical parameters, such as current voltage; sensor data, such as infrared sensor detection distance, etc. The real-time data is input into the generated filter corresponding to the virtual model X configured in step S300. The generated filter performs item-by-item detection based on the preset sample generation boundaries, such as the sample generation boundaries: structural freedom boundary: lifting height 0-18cm; electrical performance index boundary: working voltage 11~13V; sensor perception index boundary: detection distance 0-6m.

[0107] If any data exceeds the sample generation boundary, such as the sensor detection distance shows 7m, which is beyond the range of 0-6m, then "first type abnormality" will be directly output;

[0108] The first type of anomalies includes sensing anomalies, such as sensor detection values ​​out of range, and transmission anomalies, such as erroneous values ​​appearing in data transmission.

[0109] Furthermore, a second PASS is required, that is, if the real-time status data all pass the generation filter, the real-time status data is input to the abnormality discriminator, and the corresponding abnormality detection result obtained is a second type of abnormality.

[0110] For example, if all real-time status data meets the boundary conditions for generating a filter, such as a voltage of 12V, a detection distance of 4m, and a lift height of 10cm, all within the range, the data is processed into a feature vector. The feature vector is then input into the anomaly discriminator corresponding to the virtual model X. The anomaly discriminator analyzes the learned anomaly feature distribution, such as the "current change pattern of lift stalls" and the "signal characteristics of sensor false alarms." The anomaly discriminator then outputs the anomaly probability and specific anomaly type, and the structured output is a second-category anomaly.

[0111] The second type of abnormality includes abnormal code and abnormal location. The abnormal code is such as "E01-lifting motor failure" and the abnormal location is such as "infrared sensor module".

[0112] Example 2, as Figure 2 As shown, based on the same inventive concept as the smart ground lock anomaly detection method based on a generative adversarial network provided in Example 1, an embodiment of the present invention further provides a smart ground lock anomaly detection system based on a generative adversarial network, including:

[0113] The intrinsic feature collection module 11 is used to collect the design information of the target ground lock series, and extract the model intrinsic features of each model in the series according to the ground lock design information to obtain a model intrinsic feature set;

[0114] The intrinsic feature clustering module 12 is used to perform cluster analysis based on the model intrinsic feature set, define multiple virtual models, and collect real abnormal sample data corresponding to each virtual model, and output it as a real training data set;

[0115] A generation boundary configuration module 13 is used to define a sample generation boundary of each virtual model in combination with the ground lock design information, and configure a generation filter based on the sample generation boundary;

[0116] The ground lock anomaly detection module 14 is used to combine the real training data set and the generated filter to construct and train a corresponding generative adversarial network for each virtual model, and perform anomaly detection of the smart ground lock based on the generative adversarial network.

[0117] Furthermore, the intrinsic feature acquisition module 11 includes the following execution steps:

[0118] Traversing the target ground lock series to obtain ground lock design information for each series model, wherein the ground lock design information includes product structure information, electrical parameter configuration, and sensing device specifications;

[0119] Based on a preset structural feature extraction rule, the characteristic fields of the product structure information are extracted and vectorized to obtain a structural feature vector;

[0120] Performing standardized conversion on the electrical parameter configuration and the sensing device specifications, fusing the standardized conversion results and correspondingly defining parameter feature vectors;

[0121] The structural feature vector and the parameter feature vector are concatenated to obtain a model structural feature vector, and the model structural feature vector of each series of models is iteratively obtained and stored as the model intrinsic feature set.

[0122] Furthermore, the intrinsic feature clustering module 12 includes the following execution steps:

[0123] Performing density-based unsupervised cluster analysis on the model intrinsic feature set to obtain multiple clusters;

[0124] Define each cluster as a virtual model, and establish a model mapping table based on the cluster boundaries;

[0125] Based on the model mapping table, the series models corresponding to each virtual model are traversed, historical abnormal state data is collected and standardized, and the structured output is the real training data set.

[0126] Furthermore, the boundary configuration generation module 13 includes the following execution steps:

[0127] Defining a structural freedom degree boundary according to the product structure information corresponding to the virtual model;

[0128] Extracting electrical performance indicator boundaries according to the electrical parameter configuration corresponding to the virtual model;

[0129] Determining a sensor perception indicator boundary according to the sensing device specifications corresponding to the virtual model;

[0130] Taking the structural freedom degree boundary, the electrical performance index boundary and the sensor perception index boundary as the sample generation boundary, correspondingly performing interval mapping and constructing a threshold detection function to form a generation filtering rule;

[0131] The generation filter for boundary detection is configured according to the generation filter rule.

[0132] Furthermore, the ground lock abnormality detection module 14 includes the following execution steps:

[0133] Taking each of the virtual models as a target, respectively constructing a sample generator and an anomaly discriminator with an adversarial structure, and connecting the generation filter to the output end of the sample generator;

[0134] Performing feature modeling of abnormal sample data on the sample generator based on the real training data set;

[0135] The sample generator and the anomaly discriminator are iteratively optimized through adversarial training until preset iteration constraints are met, and the generative adversarial network including the sample generator and the anomaly discriminator is output.

[0136] The abnormality detection of the smart ground lock based on the generative adversarial network includes:

[0137] Matching the generative adversarial network corresponding to the series model of the target smart lock according to the model mapping table, and extracting the corresponding anomaly discriminator;

[0138] The real-time status data of the target intelligent ground lock is obtained and synchronously input into the generation filter and the abnormality discriminator for 2PASS abnormality detection.

[0139] The real-time status data of the target intelligent ground lock is obtained and synchronously input into the generation filter and the abnormality discriminator for 2PASS abnormality detection, including:

[0140] Performing boundary detection on the real-time status data according to the generated filter, and outputting an abnormality detection result as a first-category abnormality if any item in the real-time status data fails to pass the test, wherein the first-category abnormality includes at least a sensing abnormality and a transmission abnormality;

[0141] If the real-time status data all pass the generation filter, the real-time status data is input into the abnormality discriminator, and the corresponding abnormality detection result obtained is a second type of abnormality, wherein the second type of abnormality at least includes an abnormality code and an abnormality position.

[0142] It should be noted that, in the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0143] Those skilled in the art will appreciate that embodiments of the present invention may provide methods, systems, or computer program products. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0144] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0145] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0146] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0147] Although preferred embodiments of the present invention have been described, additional changes and modifications to these embodiments may occur to those skilled in the art once the basic inventive concepts become known.

[0148] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. An intelligent ground lock anomaly detection method based on generative adversarial network is characterized by: include: Collecting ground lock design information of a target ground lock series, and extracting model intrinsic features of each model in the series based on the ground lock design information to obtain a model intrinsic feature set, wherein the ground lock design information includes product structure information, electrical parameter configuration, and sensing device specifications; Perform cluster analysis based on the model intrinsic feature set, define multiple virtual models accordingly, and collect real abnormal sample data corresponding to each virtual model, and output it as a real training data set; In combination with the ground lock design information, a sample generation boundary of each virtual model is defined, and a generation filter is configured based on the sample generation boundary; Combining the real training data set with the generative filter, respectively constructing and training a corresponding generative adversarial network for each virtual model, and performing anomaly detection of the smart ground lock based on the generative adversarial network; Among them, cluster analysis is performed based on the model intrinsic feature set, multiple virtual models are defined accordingly, and real abnormal sample data corresponding to each virtual model is collected and output as a real training data set, including: Performing density-based unsupervised cluster analysis on the model intrinsic feature set to obtain multiple clusters; Define each cluster as a virtual model, and establish a model mapping table based on the cluster boundaries; Based on the model mapping table, traverse the series models corresponding to each virtual model, collect and standardize historical abnormal status data, and output the structured data as the real training data set; Among them, the corresponding generative adversarial network is constructed and trained for each virtual model, including: Taking each of the virtual models as a target, respectively constructing a sample generator and an anomaly discriminator with an adversarial structure, and connecting the generation filter to the output end of the sample generator; Performing feature modeling of abnormal sample data on the sample generator based on the real training data set; Iteratively optimizing the sample generator and the anomaly discriminator through adversarial training until preset iteration constraints are satisfied, and outputting the generative adversarial network including the sample generator and the anomaly discriminator; The abnormality detection of the smart ground lock based on the generative adversarial network includes: Matching the generative adversarial network corresponding to the series model of the target smart lock according to the model mapping table, and extracting the corresponding anomaly discriminator; Acquire the real-time status data of the target intelligent ground lock and synchronously input it into the generation filter and the abnormality discriminator for 2PASS abnormality detection; Among them, 2PASS abnormality detection is performed, including: Performing boundary detection on the real-time status data according to the generated filter, and outputting an abnormality detection result as a first-category abnormality if any item in the real-time status data fails to pass the test, wherein the first-category abnormality includes at least a sensing abnormality and a transmission abnormality; If the real-time status data all pass the generation filter, the real-time status data is input into the abnormality discriminator, and the corresponding abnormality detection result obtained is a second type of abnormality, wherein the second type of abnormality at least includes an abnormality code and an abnormality position.

2. The method for detecting anomalies of intelligent ground locks based on generative adversarial networks according to claim 1, wherein: Collect the ground lock design information of the target ground lock series, and extract the model intrinsic features of each series model based on the ground lock design information to obtain the model intrinsic feature set, including: Traversing the target ground lock series to obtain ground lock design information of each series model; Based on a preset structural feature extraction rule, the characteristic fields of the product structure information are extracted and vectorized to obtain a structural feature vector; Performing standardized conversion on the electrical parameter configuration and the sensing device specifications, fusing the standardized conversion results and correspondingly defining parameter feature vectors; The structural feature vector and the parameter feature vector are concatenated to obtain a model structural feature vector, and the model structural feature vector of each series of models is iteratively obtained and stored as the model intrinsic feature set.

3. The method for detecting anomalies of intelligent ground locks based on generative adversarial networks according to claim 2, wherein: In combination with the ground lock design information, a sample generation boundary of each virtual model is defined, and a generation filter is configured based on the sample generation boundary, including: Defining a structural freedom degree boundary according to the product structure information corresponding to the virtual model; Extracting electrical performance indicator boundaries according to the electrical parameter configuration corresponding to the virtual model; Determining a sensor perception indicator boundary according to the sensing device specifications corresponding to the virtual model; Taking the structural freedom degree boundary, the electrical performance index boundary and the sensor perception index boundary as the sample generation boundary, correspondingly performing interval mapping and constructing a threshold detection function to form a generation filtering rule; The generation filter for boundary detection is configured according to the generation filter rule.

4. The intelligent ground lock anomaly detection system based on generative adversarial network is characterized by: The method for detecting anomalies of an intelligent ground lock based on a generative adversarial network according to any one of claims 1 to 3 comprises: An intrinsic feature collection module is used to collect the ground lock design information of the target ground lock series, and extract the model intrinsic features of each series model according to the ground lock design information to obtain a model intrinsic feature set; The intrinsic feature clustering module is used to perform cluster analysis based on the model intrinsic feature set, define multiple virtual models, and collect real abnormal sample data corresponding to each virtual model, and output it as a real training data set; a generation boundary configuration module, configured to define a sample generation boundary of each virtual model in combination with the ground lock design information, and configure a generation filter based on the sample generation boundary; The ground lock anomaly detection module is used to combine the real training data set with the generated filter to construct and train a corresponding generative adversarial network for each virtual model, and perform anomaly detection of the smart ground lock based on the generative adversarial network.

Images