Risk identification model training method and device, storage medium and processor

Abstract

The invention discloses a risk identification model training method and device, a storage medium and a processor. In the scheme, a training sample set is acquired; the training sample set comprises multivariate time series data corresponding to a plurality of Internet of Things devices; inputting the training sample set into a conditional embedding representation learning network in the initial risk identification model, and converting the multivariate time sequence data into a space-time structure matrix by the conditional embedding representation learning network; inputting the space-time structure matrix into a conditional diffusion neural network in the initial risk identification model, and extracting target feature data from the space-time structure matrix by the conditional diffusion neural network; the target feature data comprises key features required for identifying risks existing in the state change process of the Internet of Things equipment; and performing model training on the initial risk identification model based on the space-time structure matrix and the target feature data to obtain a risk identification model. Through the scheme, the risk identification model with higher risk identification accuracy can be obtained through training.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a training method, device, storage medium and processor for a risk identification model. Background Art

[0002] With the rapid development of Internet of Things technology, the application of Internet of Things (IoT) data in insurance risk management is becoming more and more extensive, for example, it is widely used in risk assessment, accident prevention and many other aspects.

[0003] Most of the IoT data exists in the form of multivariate time series. When constructing risk identification models, existing technologies usually directly extract feature data from multivariate time series data and use the feature data to train the model. However, since multivariate time series data usually has complex patterns, high noise interference, data sparsity, and semantic diversity, it is difficult to capture effective feature data during feature extraction of multivariate time series data, resulting in a low accuracy rate in risk identification by the risk identification model after training. Summary of the invention

[0004] Based on the above problems, the present application provides a risk identification model training method, device, storage medium and processor, the purpose of which is to improve the accuracy of risk identification by the risk identification model.

[0005] The embodiments of the present application disclose the following technical solutions:

[0006] In a first aspect of the present application, a method for training a risk identification model is provided, the method comprising:

[0007] A training sample set for training an initial risk identification model is obtained; the training sample set includes multivariate time series data corresponding to multiple IoT devices; the multivariate time series data is time series data of multiple indicator dimensions within a preset time period;

[0008] The training sample set is input into the conditional embedding representation learning network in the initial risk identification model, and the conditional embedding representation learning network transforms the multivariate time series data into a spatiotemporal structure matrix; the spatiotemporal structure matrix is ​​used to characterize the state change process of IoT devices in the time dimension and the space dimension;

[0009] The spatiotemporal structure matrix is ​​input into the conditional diffusion neural network in the initial risk identification model, and the conditional diffusion neural network extracts target feature data from the spatiotemporal structure matrix; the target feature data includes key features required to identify the risks existing in the state change process of the IoT device;

[0010] The initial risk identification model is trained based on the spatiotemporal structure matrix and target feature data to obtain a risk identification model; the risk identification model is used to perform risk identification on the IoT device to be identified; the type of the IoT device to be identified is the same as the type of multiple IoT devices.

[0011] In an optional implementation, the target feature data is extracted from the spatiotemporal structure matrix by a conditional diffusion neural network, including:

[0012] The conditional diffusion neural network performs diffusion denoising on the spatiotemporal structure matrix to generate denoised data; the diffusion denoising process is used to remove abnormal data in the spatiotemporal structure matrix;

[0013] The denoised data is processed by a conditional diffusion neural network to reduce the dimension of the data to obtain the reduced-dimensional data;

[0014] The conditional diffusion neural network is used to extract target feature data from the dimensionality reduced data.

[0015] In an optional implementation, the training sample set is input into the conditional embedding representation learning network in the initial risk identification model, and the conditional embedding representation learning network transforms the multivariate time series data into a spatiotemporal structure matrix, including:

[0016] The conditional embedding representation learning network extracts conditional feature data from the multivariate time series data; the conditional feature data includes: multiple moments in a preset time period, multiple indicator dimensions, and the data value of each indicator dimension corresponding to each moment;

[0017] The time variation relationship corresponding to the multivariate time series data is determined by the conditional embedding representation learning network; the time variation relationship is the variation trend of the time series data of multiple indicator dimensions in the time dimension;

[0018] The spatial variation relationship corresponding to the multivariate time series data is determined by the conditional embedding representation learning network; the spatial variation relationship is the variation trend of the time series data of multiple indicator dimensions in the spatial dimension;

[0019] The conditional embedding representation learning network constructs a spatiotemporal structure matrix based on conditional feature data, temporal variation relations, and spatial variation relations.

[0020] In an optional implementation, the initial risk identification model is trained based on the spatiotemporal structure matrix and the target feature data to obtain a risk identification model, including:

[0021] The conditional feature data and the target feature data are input into the attention fusion network in the initial risk identification model, and the attention fusion network extracts multiple key features from the conditional feature data and the target feature data;

[0022] The attention fusion network performs weighted processing on each key feature based on the attention mechanism to obtain multiple weighted key features;

[0023] The attention fusion network performs feature fusion on multiple weighted key features to obtain feature enhanced data;

[0024] The initial risk identification model is trained based on the feature enhancement data to obtain a risk identification model.

[0025] In an optional implementation, during the process of training the initial risk identification model, the risk identification model training method further includes:

[0026] Obtain a first spatiotemporal structure matrix or a second spatiotemporal structure matrix; the first spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at the first moment; the second spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at the second moment; the second moment is later than the first moment, and the second moment is separated from the first moment by a preset time length;

[0027] Acquire first target feature data and second target feature data; the first target feature data is target feature data extracted from the first spatiotemporal structure matrix by the conditional diffusion neural network, and the second target feature data is target feature data extracted from the second spatiotemporal structure matrix by the conditional diffusion neural network;

[0028] Based on the first target feature data and the second target feature data, and at least one of the first space-time structure matrix and the second space-time structure matrix, the parameters of the conditional diffusion neural network are updated to obtain an updated conditional diffusion neural network; the updated conditional diffusion neural network is used to extract target feature data from a third space-time structure matrix, the third space-time structure matrix is ​​the space-time structure matrix output by the conditional embedding representation learning network at a third moment, the third moment and the second moment are separated by a preset time length, and the third moment is later than the second moment.

[0029] In an optional implementation, updating the parameters of the conditional diffusion neural network based on the first target feature data and the second target feature data and at least one of the first spatiotemporal structure matrix and the second spatiotemporal structure matrix to obtain an updated conditional diffusion neural network includes:

[0030] Constructing a positive sample pair based on the first target feature data and the second target feature data;

[0031] Constructing a negative sample pair based on the first target feature data and the first spatiotemporal structure matrix, or constructing a negative sample pair based on the second target feature data and the second spatiotemporal structure matrix;

[0032] The parameters of the conditional diffusion neural network are updated based on the positive sample pairs and the negative sample pairs to obtain an updated conditional diffusion neural network.

[0033] In a second aspect of the present application, a risk identification model training device is provided, the device comprising:

[0034] An acquisition module is used to acquire a training sample set for training an initial risk identification model; the training sample set includes multivariate time series data corresponding to multiple IoT devices; the multivariate time series data is time series data of multiple indicator dimensions within a preset time period;

[0035] The data conversion module is used to input the training sample set into the conditional embedding representation learning network in the initial risk identification model, and the conditional embedding representation learning network converts the multivariate time series data into a spatiotemporal structure matrix; the spatiotemporal structure matrix is ​​used to characterize the state change process of the IoT device in the time dimension and the space dimension;

[0036] A feature extraction module is used to input the spatiotemporal structure matrix into the conditional diffusion neural network in the initial risk identification model, and the conditional diffusion neural network extracts target feature data from the spatiotemporal structure matrix; the target feature data includes key features required to identify the risks existing in the IoT device during the state change process;

[0037] A training module is used to train an initial risk identification model based on a spatiotemporal structure matrix and target feature data to obtain a risk identification model; the risk identification model is used to identify risks of an IoT device to be identified; the type of the IoT device to be identified is the same as the type of multiple IoT devices.

[0038] Optionally, the feature extraction module includes:

[0039] A diffusion denoising processing unit is used to perform diffusion denoising on the spatiotemporal structure matrix by a conditional diffusion neural network to generate denoised data; the diffusion denoising process is used to remove abnormal data in the spatiotemporal structure matrix;

[0040] A dimension reduction processing unit is used to perform dimension reduction processing on the denoised data by a conditional diffusion neural network to obtain dimension-reduced data;

[0041] The feature extraction unit is used to extract target feature data from the dimensionally reduced data by a conditional diffusion neural network.

[0042] In a third aspect of the present application, a computer-readable storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, the training method of the risk identification model introduced in the first aspect above is implemented.

[0043] In a fourth aspect of the present application, a processor is provided for running a computer program, and when the computer program is running, the training method of the risk identification model introduced in the first aspect is executed.

[0044] Compared with the prior art, this application has the following beneficial effects:

[0045] In the technical solution of the present application, a training sample set for training an initial risk identification model is first obtained, wherein the training sample set includes multivariate time series data corresponding to multiple IoT devices; the multivariate time series data is time series data of multiple indicator dimensions within a preset time period; then the training sample set is input into the conditional embedding representation learning network in the initial risk identification model, and the conditional embedding representation learning network converts the multivariate time series data into a spatiotemporal structure matrix, so that the state change process of the IoT device in the time dimension and the space dimension is effectively reflected through the spatiotemporal structure matrix; then the spatiotemporal structure matrix is ​​input into the conditional diffusion neural network in the initial risk identification model, and the conditional diffusion neural network extracts the key features (i.e., target feature data) required to effectively capture the risks existing in the state change process of the IoT device from the spatiotemporal structure matrix, thereby capturing effective target feature data from the multivariate time series data; thus, based on the spatiotemporal structure matrix converted from the multivariate time series data and the effective target feature data extracted from the spatiotemporal structure matrix, a risk identification model with a higher risk identification accuracy can be trained, and then the risk identification model that completes the model training can accurately identify the risks existing in the IoT device to be identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0047] Figure 1 A flowchart of a risk identification model training method provided in an embodiment of the present application;

[0048] Figure 2 A flowchart of another risk identification model training method provided in an embodiment of the present application;

[0049] Figure 3 A flowchart of a process of constructing a space-time structure matrix provided in an embodiment of the present application;

[0050] Figure 4A flowchart of a target feature data extraction process provided in an embodiment of the present application;

[0051] Figure 5 A flowchart of a model training process provided in an embodiment of the present application;

[0052] Figure 6 A flow chart of a parameter updating process of a conditional diffusion neural network provided in an embodiment of the present application;

[0053] Figure 7 A schematic diagram of the structure of a risk identification model training device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0054] As described above, most of the current IoT data exists in the form of multivariate time series. When building a risk identification model, the existing technology usually directly extracts feature data from the multivariate time series data and uses the feature data for model training. However, since multivariate time series data usually has complex patterns, high noise interference, data sparsity, and semantic diversity, it is difficult to capture effective feature data during feature extraction of multivariate time series data, resulting in a low accuracy rate in risk identification by the risk identification model after training.

[0055] After research, the inventor proposed inputting the training sample set into the conditional embedding representation learning network in the initial risk identification model, and the conditional embedding representation learning network converts the multivariate time series data into a spatiotemporal structure matrix, so that the state change process of the Internet of Things devices in the time dimension and the space dimension can be effectively reflected through the spatiotemporal structure matrix; then the spatiotemporal structure matrix is ​​input into the conditional diffusion neural network in the initial risk identification model, and the conditional diffusion neural network extracts the key features (i.e., target feature data) required to effectively capture the risks existing in the state change process of the Internet of Things devices from the spatiotemporal structure matrix, thereby achieving the capture of effective target feature data from the multivariate time series data; therefore, based on the spatiotemporal structure matrix converted from the multivariate time series data and the effective target feature data extracted from the spatiotemporal structure matrix, a risk identification model with a higher risk identification accuracy can be trained, and then the risk identification model that completes the model training can accurately identify the risks existing in the Internet of Things devices to be identified.

[0056] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0057] It should be noted that the training of the risk identification model provided in the embodiment of the present application can be implemented through a computer device, which can be a terminal device or a server, wherein the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Terminal devices include but are not limited to mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, aircraft, etc. Terminal devices and servers can be directly or indirectly connected via wired or wireless communication, and this application does not limit this.

[0058] Method Embodiment

[0059] The present application provides an embodiment of a method for training a risk identification model. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer device such as a set of computer executable instructions. In addition, although the logical order is shown in the flowchart, it is only an example order provided. In some cases, the steps shown or described can be executed in a different order from that shown here. In this embodiment, the computer device is a server as an example for explanation.

[0060] See also Figure 1 , which is a flow chart of a risk identification model training method provided in an embodiment of the present application, the method comprising:

[0061] Step S101, obtaining a training sample set for training an initial risk identification model.

[0062] In step S101, the training sample set includes multivariate time series data corresponding to multiple IoT devices; the multivariate time series data is time series data of multiple indicator dimensions within a preset time period. The preset time period is a time period set according to actual needs, so it is not specifically limited here. For example, when the IoT device is a vehicle, the multivariate time series data is time series data of multiple indicator dimensions of the vehicle between 8:00 and 23:00, or time series data of multiple indicator dimensions of the vehicle between Monday and Friday; when the IoT device is an elevator, the multivariate time series data is time series data of multiple indicator dimensions of the elevator between 7:00-21:00.

[0063] In this embodiment, the IoT device may include but is not limited to any of the following types of devices: vehicles, elevators, or wearable devices.

[0064] (1) If the IoT device is a vehicle, in the technical solution of this application, the multivariate time series data corresponding to the vehicle can be collected through sensors installed on the vehicle. For example, vehicle speed, vehicle acceleration, driving habits, driving routes, braking status, geographic location and other vehicle networking data for assessing driving risks can be collected.

[0065] (2) If the IoT device is an elevator, in the technical solution of this application, multivariate time series data corresponding to the elevator can be collected to assess the failure risk of the elevator. For example, multivariate time series data such as the elevator's operating status, sensor data (such as the elevator's acceleration, vibration status, and door switch status), environmental data (such as the temperature and humidity in the elevator), and the frequency of use of the elevator can be collected.

[0066] (3) If the IoT device is a wearable device, it may be a smart watch, fitness tracker, or other device. In the technical solution of the present application, wearable devices can be used to collect multivariate time series data such as a customer's heart rate, activity level, sleep quality, etc., which can be used to assess the customer's health status and potential health risks.

[0067] It should be noted that the relevant information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) collected by this application are all information and data authorized by the user or fully authorized by all parties. For example, an interface is set between the server of this application and the relevant user or organization. Before obtaining relevant information, it is necessary to send an acquisition request to the aforementioned user or organization through the interface, and obtain relevant information after receiving the consent information fed back by the aforementioned user or organization.

[0068] In the scenario where the elevator is used as an IoT device, the server that executes the method process in the embodiment of the present application can obtain time series data of multiple indicator dimensions such as the elevator's operating status, sensor data (such as the elevator's acceleration, vibration conditions, door switch status), environmental data (such as the temperature and humidity in the elevator), and frequency of use of the elevator from the smart elevator system.

[0069] Step S102, input the training sample set into the conditional embedding representation learning network in the initial risk identification model, and the conditional embedding representation learning network converts the multivariate time series data into a spatiotemporal structure matrix.

[0070] In step S102, the spatiotemporal structure matrix is ​​used to characterize the state change process of the IoT device in the time dimension and the space dimension. The conditional embedding representation learning network in the present application can be a convolutional neural network, which can integrate conditional information into the embedding representation so that the risk identification model can better capture the characteristics and patterns of the data under different conditions.

[0071] In order to improve the accuracy of risk identification model in identifying risks, in this embodiment, Figure 2 As shown, the server can input the multivariate time series data in the training sample set into the conditional embedding representation learning network. The conditional embedding representation learning network can transform the multivariate time series data into a spatiotemporal structure matrix through conditional embedding, thereby effectively reflecting the state change process of the IoT device in the time dimension and the space dimension through the spatiotemporal structure matrix, and can provide a data basis for the subsequent extraction of target feature data.

[0072] In the scenario where the elevator is used as an IoT device, the elevator's operating status, sensor data (such as the elevator's acceleration, vibration status, door switch status), environmental data (such as the temperature and humidity in the elevator), and the floor height of the building where the elevator is located, the frequency of elevator use and other time series data within a preset time period are input into the conditional embedding representation learning network. The conditional embedding representation learning network can convert the above data into a spatiotemporal structure matrix that effectively reflects the changing process of the elevator's operating status at different times and on different floors. Based on this spatiotemporal structure matrix, it can help to comprehensively capture the state changes of the elevator in the time and space dimensions, and can capture the performance differences of the elevator on different floors and under environmental conditions.

[0073] See also Figure 3 , Figure 3 A flowchart of a process of constructing a space-time structure matrix provided in an embodiment of the present application, the process includes the following steps:

[0074] Step S1021, extracting conditional feature data from multivariate time series data by a conditional embedding representation learning network.

[0075] In step S1021, the conditional characteristic data includes: multiple moments in a preset time period, multiple indicator dimensions, and the data value of each indicator dimension corresponding to each moment.

[0076] Taking the multivariate time series data of the above elevator, the multivariate time series data can be shown in Table 1.

[0077] Table 1

[0078] Time(X) Acceleration (Y) Vibration condition (Y) Door switch status (Y) … t1 a1 b1 c1 … t2 a2 b2 c2 … t3 … … … …

[0079] The conditional feature data extracted from the multivariate time series data by the conditional embedding representation learning network can be shown in Table 2.

[0080] Table 2

[0081]

[0082]

[0083] It should be noted that the multivariate time series data exists in the two-dimensional form of the time dimension as shown in Table 1 above. Through the conditional embedding representation learning network, the conditional feature data in the three-dimensional form shown in Table 2 can be extracted from the multivariate time series data, which can provide a data basis for the subsequent construction of the spatiotemporal structure matrix.

[0084] Step S1022, determining the time variation relationship corresponding to the multivariate time series data by the conditional embedding representation learning network.

[0085] In step S1022, the time change relationship is the change trend of the time series data of multiple indicator dimensions in the time dimension, for example, the change trend of the running state of the elevator at different times.

[0086] Step S1023, determining the spatial variation relationship corresponding to the multivariate time series data by the conditional embedding representation learning network.

[0087] In step S1023, the spatial variation relationship is the variation trend of the time series data of multiple indicator dimensions in the spatial dimension, for example, the variation trend of the running status of the elevator on different floors.

[0088] Step S1024, the conditional embedding representation learning network constructs a spatiotemporal structure matrix based on the conditional feature data, the temporal variation relationship and the spatial variation relationship.

[0089] In this embodiment, the server can construct a conditional embedding representation learning network based on conditional feature data, time variation relationship and space variation relationship. Figure 2 The space-time structure matrix shown.

[0090] It should be noted that by constructing a spatiotemporal structure matrix based on conditional feature data, time change relationships and space change relationships through a conditional embedding representation learning network, multiple moments in multivariate time series data, multiple indicator dimensions, and the data values ​​of each indicator dimension corresponding to each moment can be more correlated in the time dimension and space dimension, and the spatiotemporal structure matrix can effectively reflect the state change process of IoT devices in the time dimension and space dimension, which can provide a data basis for the subsequent extraction of target feature data. Therefore, the risk identification model trained based on the spatiotemporal structure matrix can effectively capture the fine-grained features between multiple moments and multiple indicator dimensions within the multivariate time series data, thereby improving the accuracy of the risk identification model in identifying risks.

[0091] Step S103, inputting the spatiotemporal structure matrix into the conditional diffusion neural network in the initial risk identification model, and extracting target feature data from the spatiotemporal structure matrix by the conditional diffusion neural network.

[0092] In step S103, the target feature data includes key features required to identify the risks of IoT devices during state changes. The conditional diffusion neural network in this embodiment is a deep neural network based on a diffusion model.

[0093] In order to improve the accuracy of risk identification model in identifying risks, in this embodiment, the server can input the spatiotemporal structure matrix into the conditional diffusion neural network, and the conditional diffusion neural network can be used to identify risks. Figure 2 The reverse diffusion process shown in , extracts the key features (i.e., target feature data, which can be obtained by Figure 2 The time and indicator dimensions shown in ) make it possible to capture effective target feature data from multivariate time series data.

[0094] In the scenario where the elevator is used as an IoT device, the server will be able to effectively reflect the spatiotemporal structure matrix of the elevator's operating status changes at different times and on different floors, and input it into the conditional diffusion neural network. The conditional diffusion neural network will extract the key data that can be directly used for elevator failure prediction and safety assessment, such as acceleration anomalies, vibration mode anomalies, door switch failures, etc., from the spatiotemporal structure matrix through the inverse diffusion process. The risk identification model can identify the potential failure risks of the elevator based on the above key data, so that the smart elevator system can take preventive measures before the problem occurs, such as notifying maintenance personnel in advance to conduct inspections or repairs, thereby effectively reducing the failure risk of the elevator.

[0095] See also Figure 4 , Figure 4 A flowchart of a target feature data extraction process provided in an embodiment of the present application, the process includes the following steps:

[0096] Step S1031, a conditional diffusion neural network is used to perform diffusion denoising on the spatiotemporal structure matrix to generate denoised data.

[0097] In step S1031, diffusion denoising is used to remove abnormal data in the spatiotemporal structure matrix. Specifically, diffusion denoising can be used to remove high-frequency noise and abnormal fluctuations in the spatiotemporal structure matrix, such as sensor errors, short-term environmental interference, etc.

[0098] In this embodiment, the conditional diffusion neural network can remove abnormal data in the spatiotemporal structure matrix through a diffusion denoising algorithm to generate denoised data. The diffusion denoising algorithm can be any one of the algorithms used for diffusion denoising, such as isotropic diffusion and anisotropic diffusion, which are not specifically limited here.

[0099] Step S1032: Perform dimensionality reduction processing on the denoised data using a conditional diffusion neural network to obtain dimensionality reduced data.

[0100] Step S1033, extracting target feature data from the data after dimensionality reduction by a conditional diffusion neural network.

[0101] It should be noted that by performing diffusion denoising on the space-time structure matrix to generate denoised data, high-frequency noise and abnormal fluctuations in the space-time structure matrix can be removed; performing dimensionality reduction on the denoised data can further remove the noise in the data, so that the conditional diffusion neural network can extract accurate target feature data from the reduced-dimensional data, thereby enabling the risk identification model trained based on the accurate target feature data to identify risks more accurately.

[0102] Step S104, training the initial risk identification model based on the spatiotemporal structure matrix and the target feature data to obtain a risk identification model.

[0103] In step S104, the risk identification model is used to identify the risk of the IoT device to be identified; the type of the IoT device to be identified is the same as the type of the multiple IoT devices. For example, when the type of the multiple IoT devices is elevators, the risk identification model is used to identify the risk of the elevator to be identified.

[0104] In this embodiment, the server can train a risk identification model with a higher risk identification accuracy based on the spatiotemporal structure matrix transformed from multivariate time series data and the effective target feature data extracted from the spatiotemporal structure matrix, and then the risk identification model that completes the model training can accurately identify the risks of the IoT devices to be identified.

[0105] In order to further improve the accuracy of risk identification by the risk identification model, the server of the embodiment of the present application can perform feature enhancement processing on the target feature data through the attention fusion network. Figure 5 , Figure 5 A flowchart of a model training process provided in an embodiment of the present application, the process includes the following steps:

[0106] Step S1041, inputting the conditional feature data and the target feature data into the attention fusion network in the initial risk identification model, and the attention fusion network extracts multiple key features from the conditional feature data and the target feature data.

[0107] In this embodiment, if Figure 2 As shown, the server can input conditional feature data and target feature data including time, indicator dimensions and data values ​​into the attention fusion network, and the attention fusion network extracts multiple key features required to identify the risks of IoT devices during state changes from the conditional feature data and target feature data.

[0108] In step S1042, the attention fusion network performs weighted processing on each key feature based on the attention mechanism to obtain multiple weighted key features.

[0109] In step S1043, the attention fusion network performs feature fusion on the weighted multiple key features to obtain feature enhancement data.

[0110] Step S1044: training the initial risk identification model based on the feature enhancement data to obtain a risk identification model.

[0111] It should be noted that the attention fusion network can perform feature enhancement on the target feature data by weighting and fusing multiple key features extracted from the conditional feature data and the target feature data, so that the risk identification model trained based on the feature enhanced data can identify risks more accurately.

[0112] Furthermore, in order to train a risk identification model with a higher risk identification accuracy, in this embodiment, the server can update the parameters of the conditional diffusion neural network by contrastive learning. Figure 6 , Figure 6 A flowchart of a parameter updating process of a conditional diffusion neural network provided in an embodiment of the present application, the process includes the following steps:

[0113] Step S601, obtaining a first space-time structure matrix or a second space-time structure matrix.

[0114] In step S601, the first spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at the first moment; the second spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at the second moment; the second moment is later than the first moment, and the second moment is separated from the first moment by a preset time length.

[0115] In this embodiment, the preset time length can be the time interval between two adjacent iterations of the conditional embedding representation learning network. Thus, the first space-time structure matrix can represent the space-time structure matrix output by the conditional embedding representation learning network at N-1 iterations, and the second space-time structure matrix can represent the space-time structure matrix output by the conditional embedding representation learning network at N iterations, where N is an integer greater than 1.

[0116] Step S602, obtaining first target feature data and second target feature data.

[0117] In step S602, the first target feature data is the target feature data extracted from the first spatiotemporal structure matrix by the conditional diffusion neural network, and the second target feature data is the target feature data extracted from the second spatiotemporal structure matrix by the conditional diffusion neural network.

[0118] In this embodiment, the first target feature data may represent the target feature data output by the conditional diffusion neural network at N-1 iterations, and the second target feature data may represent the target feature data output by the conditional diffusion neural network at N iterations.

[0119] Step S603, updating the parameters of the conditional diffusion neural network based on the first target feature data and the second target feature data, and at least one of the first spatiotemporal structure matrix and the second spatiotemporal structure matrix, to obtain an updated conditional diffusion neural network.

[0120] In step S603, the updated conditional diffusion neural network is used to extract target feature data from the third spatiotemporal structure matrix, where the third spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at the third moment, where the third moment is separated from the second moment by a preset time length, and the third moment is later than the second moment. The third spatiotemporal structure matrix may be the spatiotemporal structure matrix output by the conditional embedding representation learning network at the N+1th iteration.

[0121] For example, if the first moment is the moment when the conditional embedding representation learning network outputs the first spatiotemporal structure matrix after 99 iterations, and if the second moment is the moment when the conditional embedding representation learning network outputs the second spatiotemporal structure matrix after 100 iterations, then the third moment is the moment when the conditional embedding representation learning network outputs the third spatiotemporal structure matrix after 101 iterations; the updated conditional diffusion neural network obtained by updating the parameters of the conditional diffusion neural network based on the first spatiotemporal structure matrix output by the conditional embedding representation learning network after 99 iterations and the second spatiotemporal structure matrix output by the conditional embedding representation learning network after 100 iterations can extract more accurate target feature data from the third spatiotemporal structure matrix output by the conditional embedding representation learning network after 101 iterations.

[0122] Specifically, in this embodiment, the server can construct a positive sample pair based on the first target feature data and the second target feature data; then the server can construct a negative sample pair based on the first target feature data and the first spatiotemporal structure matrix, or construct a negative sample pair based on the second target feature data and the second spatiotemporal structure matrix; finally, the parameters of the conditional diffusion neural network are updated based on the positive sample pair and the negative sample pair to obtain an updated conditional diffusion neural network.

[0123] In this embodiment, since the first target feature data can represent the target feature data output by the conditional diffusion neural network at N-1 iterations, and the second target feature data can represent the target feature data output by the conditional diffusion neural network at N iterations, the difference between the first target feature data and the second target feature data is small. Therefore, in order to enable the conditional diffusion neural network to learn similar features between feature data, the server can construct a positive sample pair based on the first target feature data and the second target feature data.

[0124] Furthermore, since the first spatiotemporal structure matrix can represent the spatiotemporal structure matrix output by the conditional embedding representation learning network at N-1 iterations, the second spatiotemporal structure matrix can represent the spatiotemporal structure matrix output by the conditional embedding representation learning network at N iterations; the first spatiotemporal structure matrix and the first target feature data are data of the same iteration, and the similarity between the first spatiotemporal structure matrix and the first target feature data is low, and the second spatiotemporal structure matrix and the second target feature data are data of the same iteration, and the similarity between the second spatiotemporal structure matrix and the second target feature data is low; therefore, the difference between the first spatiotemporal structure matrix and the first target feature data is large, and the difference between the second spatiotemporal structure matrix and the second target feature data is large, so that in order to enable the conditional diffusion neural network to learn the differences between feature data and thereby form clear category boundaries in the feature space, the server constructs a negative sample pair based on the first target feature data and the first spatiotemporal structure matrix, or constructs a negative sample pair based on the second target feature data and the second spatiotemporal structure matrix.

[0125] Therefore, by updating the parameters of the conditional diffusion neural network based on the above positive sample pairs and negative sample pairs, the conditional diffusion neural network can learn the similar features between feature data and the differences between feature data, so that the conditional diffusion neural network can extract more accurate target feature data, and then train a risk identification model with higher risk identification accuracy. In addition, the training efficiency of the risk identification model can also be improved through the comparative learning method.

[0126] In an optional embodiment, when the type of multiple IoT devices is vehicles, the server can input time series data such as vehicle speed, vehicle acceleration, driving habits, driving routes, braking status, geographic location, etc. collected from sensors on multiple vehicles within a preset time period into the conditional embedding representation learning network in the initial risk identification model, and the conditional embedding representation learning network converts the above time series data (i.e., the multivariate time series data of the vehicle) into a spatiotemporal structure matrix for characterizing the driving behavior change process of the vehicle at different times and locations; then the server can input the spatiotemporal structure matrix into the conditional diffusion neural network, and the conditional diffusion neural network extracts the vehicle's emergency braking frequency, cornering speed and other key features that effectively reflect the driver's behavioral risks from the spatiotemporal structure matrix through an inverse diffusion process, and then trains the initial risk identification model based on the above key features and the spatiotemporal structure matrix to obtain a risk identification model that can accurately identify the risks existing in the vehicle driving process. Among them, in the process of extracting key features from the spatiotemporal structure matrix by the conditional diffusion neural network through the inverse diffusion process, the conditional diffusion neural network can remove abnormal data caused by environmental noise (such as slippery roads, strong winds, etc.) or sensor errors in the spatiotemporal structure matrix through the diffusion denoising algorithm, thereby reducing the interference of noise on risk identification and further improving the accuracy of risk identification model in identifying risks.

[0127] Through the training method of the risk identification model provided in the embodiment of the present application, multivariate time series data can be converted into a space-time structure matrix, so that the state change process of the Internet of Things device in the time dimension and the space dimension can be effectively reflected through the space-time structure matrix; the key features (i.e., target feature data) required to effectively capture the risks existing in the state change process of the Internet of Things device are extracted from the space-time structure matrix through the conditional diffusion neural network, thereby achieving the capture of effective target feature data from the multivariate time series data; therefore, based on the space-time structure matrix converted from the multivariate time series data and the effective target feature data extracted from the space-time structure matrix, a risk identification model with a higher risk identification accuracy can be trained, and then the risk identification model that completes the model training can accurately identify the risks existing in the Internet of Things devices to be identified.

[0128] In addition, by performing diffusion denoising on the spatiotemporal structure matrix, the conditional diffusion neural network can extract accurate target feature data based on the denoised data, so that the risk identification model trained based on the accurate target feature data can identify risks more accurately. By updating the parameters of the conditional diffusion neural network through comparative learning, the conditional diffusion neural network can extract more accurate target feature data, and then train a risk identification model with higher risk identification accuracy.

[0129] Device Embodiment

[0130] The present application embodiment provides a training device for a risk identification model, wherein: Figure 7 A schematic diagram of a risk identification model training device provided in an embodiment of the present application is shown in FIG. Figure 7 As shown, the device includes: an acquisition module 11, a data conversion module 12, a feature extraction module 13, and a training module 14. Figure 7 You can see the connection relationship between several modules.

[0131] Among them, the acquisition module 11 is used to obtain a training sample set for training the initial risk identification model; the training sample set includes multivariate time series data corresponding to multiple IoT devices; the multivariate time series data is time series data of multiple indicator dimensions within a preset time period;

[0132] The data conversion module 12 is used to input the training sample set into the conditional embedding representation learning network in the initial risk identification model, and the conditional embedding representation learning network converts the multivariate time series data into a spatiotemporal structure matrix; the spatiotemporal structure matrix is ​​used to characterize the state change process of the IoT device in the time dimension and the space dimension;

[0133] The feature extraction module 13 is used to input the spatiotemporal structure matrix into the conditional diffusion neural network in the initial risk identification model, and the conditional diffusion neural network extracts target feature data from the spatiotemporal structure matrix; the target feature data includes key features required to identify the risks existing in the IoT device during the state change process;

[0134] The training module 14 is used to train the initial risk identification model based on the spatiotemporal structure matrix and the target feature data to obtain a risk identification model; the risk identification model is used to identify the risks of the IoT devices to be identified; the type of the IoT devices to be identified is the same as the type of the multiple IoT devices.

[0135] Optionally, the feature extraction module 13 includes: a diffusion denoising processing unit, a dimensionality reduction processing unit and a feature extraction unit.

[0136] The diffusion denoising processing unit is used to perform diffusion denoising on the spatiotemporal structure matrix by a conditional diffusion neural network to generate denoised data; the diffusion denoising process is used to remove abnormal data in the spatiotemporal structure matrix;

[0137] A dimension reduction processing unit is used to perform dimension reduction processing on the denoised data by a conditional diffusion neural network to obtain dimension-reduced data;

[0138] The feature extraction unit is used to extract target feature data from the dimensionally reduced data by a conditional diffusion neural network.

[0139] Optionally, the data conversion module includes: a conditional feature extraction unit, a first determination unit, a second determination unit and a matrix construction unit.

[0140] The conditional feature extraction unit is used to extract conditional feature data from the multivariate time series data by the conditional embedding representation learning network; the conditional feature data includes: multiple moments in a preset time period, multiple indicator dimensions, and the data value of each indicator dimension corresponding to each moment;

[0141] A first determining unit is used to determine the time change relationship corresponding to the multivariate time series data by a conditional embedding representation learning network; the time change relationship is a change trend of the time series data of multiple indicator dimensions in the time dimension;

[0142] The second determining unit is used to determine the spatial variation relationship corresponding to the multivariate time series data by the conditional embedding representation learning network; the spatial variation relationship is the variation trend of the time series data of multiple indicator dimensions in the spatial dimension;

[0143] The matrix construction unit is used to construct a spatiotemporal structure matrix based on conditional feature data, time variation relationship and space variation relationship by the conditional embedding representation learning network.

[0144] Optionally, the training module includes: a key feature extraction unit, a weighted processing unit, a feature fusion unit and a model training unit.

[0145] The key feature extraction unit is used to input the condition feature data and the target feature data into the attention fusion network in the initial risk identification model, and the attention fusion network extracts multiple key features from the condition feature data and the target feature data;

[0146] A weighted processing unit, used for performing weighted processing on each key feature based on the attention mechanism by the attention fusion network to obtain a plurality of weighted key features;

[0147] A feature fusion unit is used to fuse multiple weighted key features through an attention fusion network to obtain feature enhancement data;

[0148] The model training unit is used to perform model training on the initial risk identification model based on the feature enhancement data to obtain the risk identification model.

[0149] Optionally, the training module also includes: a first acquisition unit, a second acquisition unit and a parameter updating unit.

[0150] Among them, the first acquisition unit is used to acquire the first spatiotemporal structure matrix or the second spatiotemporal structure matrix; the first spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at the first moment; the second spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at the second moment; the second moment is later than the first moment, and the second moment is separated from the first moment by a preset time length;

[0151] The second acquisition unit is used to acquire first target feature data and second target feature data; the first target feature data is target feature data extracted by the conditional diffusion neural network from the first spatiotemporal structure matrix, and the second target feature data is target feature data extracted by the conditional diffusion neural network from the second spatiotemporal structure matrix;

[0152] A parameter updating unit is used to update the parameters of the conditional diffusion neural network based on the first target feature data and the second target feature data, and at least one of the first spatiotemporal structure matrix and the second spatiotemporal structure matrix to obtain an updated conditional diffusion neural network; the updated conditional diffusion neural network is used to extract the target feature data from the third spatiotemporal structure matrix, the third spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at a third moment, the third moment and the second moment are separated by a preset time length, and the third moment is later than the second moment.

[0153] Optionally, the parameter updating unit includes:

[0154] Wherein, the first construction subunit is used to construct a positive sample pair based on the first target feature data and the second target feature data;

[0155] A second construction subunit is used to construct a negative sample pair based on the first target feature data and the first spatiotemporal structure matrix, or to construct a negative sample pair based on the second target feature data and the second spatiotemporal structure matrix;

[0156] The parameter updating subunit is used to update the parameters of the conditional diffusion neural network based on the positive sample pairs and the negative sample pairs to obtain an updated conditional diffusion neural network.

[0157] Storage Medium Embodiments

[0158] The embodiment of the present application provides a computer-readable storage medium, on which a program is stored, wherein when the program is executed by a processor, some or all of the steps in the risk identification model training method described in the aforementioned method embodiment of the present application are implemented. The storage medium can be a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, and other media that can store program codes.

[0159] Processor Embodiments

[0160] An embodiment of the present application provides a processor for running a program, wherein part or all of the steps in the risk identification model training method introduced in the aforementioned method embodiment are executed when the program is running.

[0161] It should be noted that each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is merely schematic, in which the unit described as a separate component may or may not be physically separated, and the component prompted as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0162] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for training a risk identification model, characterized in that: include: Obtaining a training sample set for training an initial risk identification model; The training sample set includes multivariate time series data corresponding to a plurality of IoT devices; The multivariate time series data is time series data of multiple indicator dimensions within a preset time period; Inputting the training sample set into the conditional embedding representation learning network in the initial risk identification model, the conditional embedding representation learning network converts the multivariate time series data into a spatiotemporal structure matrix; the spatiotemporal structure matrix is ​​used to characterize the state change process of the IoT device in the time dimension and the space dimension; Inputting the spatiotemporal structure matrix into the conditional diffusion neural network in the initial risk identification model, and extracting target feature data from the spatiotemporal structure matrix by the conditional diffusion neural network; The target feature data includes key features required to identify the risks of the IoT device during the state change process; The initial risk identification model is trained based on the spatiotemporal structure matrix and the target feature data to obtain a risk identification model; the risk identification model is used to perform risk identification on the IoT device to be identified; The type of the Internet of Things device to be identified is the same as the type of the multiple Internet of Things devices.

2. The method according to claim 1, characterized in that The step of extracting target feature data from the spatiotemporal structure matrix by the conditional diffusion neural network includes: The conditional diffusion neural network performs diffusion denoising on the space-time structure matrix to generate denoised data; the diffusion denoising is used to remove abnormal data in the space-time structure matrix; The conditional diffusion neural network performs dimensionality reduction processing on the denoised data to obtain dimensionality reduced data; The conditional diffusion neural network extracts the target feature data from the dimensionally reduced data.

3. The method according to claim 1, characterized in that The step of inputting the training sample set into the conditional embedding representation learning network in the initial risk identification model, and converting the multivariate time series data into a spatiotemporal structure matrix by the conditional embedding representation learning network, comprises: The conditional embedding representation learning network extracts conditional feature data from the multivariate time series data; the conditional feature data includes: a plurality of moments within the preset time period, a plurality of the indicator dimensions, and a data value of each of the indicator dimensions corresponding to each of the moments; The conditional embedding representation learning network determines the time variation relationship corresponding to the multivariate time series data; the time variation relationship is the variation trend of the time series data of the multiple indicator dimensions in the time dimension; The conditional embedding representation learning network determines the spatial variation relationship corresponding to the multivariate time series data; the spatial variation relationship is the variation trend of the time series data of the multiple indicator dimensions in the spatial dimension; The conditional embedding representation learning network constructs the spatiotemporal structure matrix based on the conditional feature data, the time variation relationship and the space variation relationship.

4. The method according to claim 3, characterized in that The performing model training on the initial risk identification model based on the spatiotemporal structure matrix and the target feature data to obtain a risk identification model includes: Inputting the conditional feature data and the target feature data into the attention fusion network in the initial risk identification model, and extracting a plurality of the key features from the conditional feature data and the target feature data by the attention fusion network; The attention fusion network performs weighted processing on each of the key features based on the attention mechanism to obtain a plurality of weighted key features; The attention fusion network performs feature fusion on the weighted multiple key features to obtain feature enhancement data; The initial risk identification model is trained based on the feature enhancement data to obtain the risk identification model.

5. The method according to claim 1, characterized in that In the process of training the initial risk identification model, the method further includes: Obtain a first spatiotemporal structure matrix or a second spatiotemporal structure matrix; the first spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at the first moment; the second spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at the second moment; the second moment is later than the first moment, and the second moment is separated from the first moment by a preset time length; Acquire first target feature data and second target feature data; the first target feature data is the target feature data extracted by the conditional diffusion neural network from the first spatiotemporal structure matrix, and the second target feature data is the target feature data extracted by the conditional diffusion neural network from the second spatiotemporal structure matrix; Based on the first target feature data and the second target feature data, and at least one of the first spatiotemporal structure matrix and the second spatiotemporal structure matrix, the parameters of the conditional diffusion neural network are updated to obtain an updated conditional diffusion neural network; the updated conditional diffusion neural network is used to extract the target feature data from a third spatiotemporal structure matrix, and the third spatiotemporal structure matrix is ​​the spatiotemporal structure matrix output by the conditional embedding representation learning network at a third moment, the third moment and the second moment are separated by a preset time length, and the third moment is later than the second moment.

6. The method according to claim 5, characterized in that The updating of the parameters of the conditional diffusion neural network based on the first target feature data and the second target feature data and at least one of the first spatiotemporal structure matrix and the second spatiotemporal structure matrix to obtain an updated conditional diffusion neural network includes: Constructing a positive sample pair based on the first target feature data and the second target feature data; Constructing a negative sample pair based on the first target feature data and the first spatiotemporal structure matrix, or constructing a negative sample pair based on the second target feature data and the second spatiotemporal structure matrix; The parameters of the conditional diffusion neural network are updated based on the positive sample pairs and the negative sample pairs to obtain the updated conditional diffusion neural network.

7. A training device for a risk identification model, characterized in that: include: An acquisition module, used to acquire a training sample set for training an initial risk identification model; The training sample set includes multivariate time series data corresponding to a plurality of IoT devices; The multivariate time series data is time series data of multiple indicator dimensions within a preset time period; A data conversion module, used for inputting the training sample set into the conditional embedding representation learning network in the initial risk identification model, and the conditional embedding representation learning network converts the multivariate time series data into a spatiotemporal structure matrix; the spatiotemporal structure matrix is ​​used to characterize the state change process of the IoT device in the time dimension and the space dimension; A feature extraction module, used for inputting the spatiotemporal structure matrix into the conditional diffusion neural network in the initial risk identification model, and extracting target feature data from the spatiotemporal structure matrix by the conditional diffusion neural network; The target feature data includes key features required to identify the risks of the IoT device during the state change process; A training module, used for performing model training on the initial risk identification model based on the spatiotemporal structure matrix and the target feature data to obtain the risk identification model; The risk identification model is used to identify risks of IoT devices to be identified; The type of the Internet of Things device to be identified is the same as the type of the multiple Internet of Things devices.

8. The device according to claim 7, characterized in that The feature extraction module comprises: A diffusion denoising processing unit, used for performing diffusion denoising processing on the spatiotemporal structure matrix by the conditional diffusion neural network to generate denoised data; the diffusion denoising processing is used to remove abnormal data in the spatiotemporal structure matrix; A dimensionality reduction processing unit, used for performing dimensionality reduction processing on the denoised data by the conditional diffusion neural network to obtain dimensionality reduced data; A feature extraction unit is used to extract the target feature data from the dimensionally reduced data by the conditional diffusion neural network.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the risk identification model training method as described in claims 1-6 is implemented.

10. A processor, characterized in that: Used to run a computer program, which, when running, executes the risk identification model training method as described in claims 1-6.

Images