Early warning method and device based on multi-source accident data, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to business scenes such as financial science and technology, and discloses an early warning method, device and equipment based on multi-source accident data and a medium. Training a similarity matching model based on the dimension reduction feature set to generate a similar index structure, analyzing the real-time sensor data stream, converting the real-time sensor data stream into a real-time feature vector, and mapping the real-time feature vector to the similar index structure to obtain a candidate accident record set; and performing similarity comparison on the candidate accident record set and the real-time feature vector to generate a similarity index, and outputting abnormal early warning information when the similarity index exceeds a preset threshold. According to the method, the historical accident data and the real-time sensor data are fused, rapid retrieval and comparison are achieved in a similar index structure, and the real-time performance and accuracy of early warning are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a pre-warning method and device based on multi-source accident data, equipment and medium. BACKGROUND

[0002] In the field of financial technology business, vehicle insurance scene, real-time vehicle pre-warning technology has been widely used in risk control and accident prevention, but the existing scheme still has many deficiencies.

[0003] In the prior art, multi-source accident data is often scattered in vehicle operation records, collision event sequences, claim and survey data, voice reports and other heterogeneous sources, lacking a unified gathering and storage mechanism, making it difficult to form a consistent and comprehensive accident database, affecting the reliability and timeliness of subsequent retrieval and comparison.

[0004] In the processing of accident characteristics, the dimensions, value ranges and formats of different source data are significantly different, lacking a unified standardization mechanism, which will cause deviation in similarity measurement. At the same time, the accident characteristic dimension is high and redundant, and the similarity calculation efficiency is low in high-dimensional space, which is difficult to meet the low delay requirement of real-time pre-warning.

[0005] In the aspect of similarity retrieval, the existing system mostly relies on sequential comparison or simple matching, lacking an efficient index structure for similarity search, with slow retrieval speed and low recall rate, resulting in insufficient and inaccurate positioning of candidate samples. In addition, real-time sensor data stream is complex in form, with inconsistent field naming and value, lacking a unified mapping mechanism for historical accident characteristics, making it difficult to ensure that real-time characteristics and historical samples are in the same representation space.

[0006] Even if the candidate accident records are obtained, the existing method is still not uniform in similarity measurement caliber and threshold determination, the similarity index lacks consistent definition, the threshold setting is unstable, and false positives or false negatives are prone to occur, making it difficult to meet the stability and availability requirements of real-time pre-warning for vehicle insurance risk control. SUMMARY

[0007] The main purpose of the present application is to provide a pre-warning method, device, equipment and storage medium based on multi-source accident data, aiming to solve the technical problem that the existing technology lacks a unified multi-source accident data processing and real-time feature mapping mechanism, making it difficult to perform efficient and accurate similarity comparison between historical accident data and real-time sensor data in the same feature space, thereby making it difficult to output stable and reliable pre-warning information in time.

[0008] To achieve the above purpose, the present application provides a pre-warning method based on multi-source accident data, comprising:

[0009] acquire multi-source accident data, generate a vehicle accident database based on the multi-source accident data;

[0010] perform feature standardization processing on accident features in the vehicle accident database to obtain an accident feature set;

[0011] perform principal component dimension reduction processing on the accident feature set to obtain a dimension-reduced feature set;

[0012] train a similarity matching model based on the dimension-reduced feature set to generate a similarity index structure;

[0013] acquire a real-time sensor data stream, parse feature fields in the real-time sensor data stream, and convert the feature fields into a real-time feature vector;

[0014] map the real-time feature vector to the similarity index structure to obtain a candidate accident record set;

[0015] perform similarity comparison on the candidate accident record set and the real-time feature vector to generate a similarity index, and output an abnormal early warning information when the similarity index exceeds a preset threshold.

[0016] Further, to achieve the above object, the application provides a warning device based on multi-source accident data, comprising:

[0017] an accident database construction module configured to acquire multi-source accident data and generate a vehicle accident database based on the multi-source accident data;

[0018] a feature standardization module configured to perform feature standardization processing on accident features in the vehicle accident database to obtain an accident feature set;

[0019] a feature dimension reduction module configured to perform principal component dimension reduction processing on the accident feature set to obtain a dimension-reduced feature set;

[0020] a similarity matching model training module configured to train a similarity matching model based on the dimension-reduced feature set to generate a similarity index structure;

[0021] a real-time feature vector generation module configured to acquire a real-time sensor data stream, parse feature fields in the real-time sensor data stream, and convert the feature fields into a real-time feature vector;

[0022] a candidate accident record retrieval module configured to map the real-time feature vector to the similarity index structure to obtain a candidate accident record set;

[0023] a similarity comparison and early warning module configured to perform similarity comparison on the candidate accident record set and the real-time feature vector to generate a similarity index, and output an abnormal early warning information when the similarity index exceeds a preset threshold.

[0024] Further, to achieve the above object, the present application also provides a computer device, comprising a memory, a processor and a multi-source accident data based early warning program stored in the memory and executable on the processor, wherein the multi-source accident data based early warning program realizes the steps of the multi-source accident data based early warning method when executed by the processor.

[0025] Further, to achieve the above object, the present application also provides a computer readable storage medium, wherein the storage medium stores a multi-source accident data based early warning program, and the multi-source accident data based early warning program realizes the steps of the multi-source accident data based early warning method when executed by a processor.

[0026] Beneficial effects: The present application relates to the field of artificial intelligence technology, and can be applied to business scenarios such as financial technology. A multi-source accident data based early warning method, device, equipment and medium are disclosed, comprising: acquiring multi-source accident data and generating a vehicle accident database, performing feature standardization processing on accident features in the vehicle accident database to form an accident feature set, performing principal component dimension reduction on the accident feature set to obtain a reduced dimension feature set, training a similarity matching model based on the reduced dimension feature set to generate a similarity index structure, receiving a real-time sensor data stream and parsing a feature field to convert it into a real-time feature vector, mapping the real-time feature vector to the similarity index structure to obtain a candidate accident record set, performing similarity comparison on the candidate accident record set and the real-time feature vector to generate a similarity index, and outputting abnormal early warning information when the similarity index exceeds a preset threshold. The present application unifies the mapping of historical accident features and real-time feature vectors to a similarity index structure by constructing a vehicle accident database and combining a real-time sensor data stream, realizes the rapid retrieval of a candidate accident record set, and generates a similarity index through similarity comparison and triggers abnormal early warning information, thereby ensuring real-time performance while improving the accuracy of comparison results and the reliability of early warning. BRIEF DESCRIPTION OF DRAWINGS

[0027] The present application will be further described below with reference to the accompanying drawings and embodiments. In the drawings:

[0028] Figure 1 An application environment diagram of the multi-source accident data based early warning method in an embodiment of the present application;

[0029] Figure 2 A flow diagram of the multi-source accident data based early warning method in an embodiment of the present application;

[0030] Figure 3 A functional module diagram of a preferred embodiment of the multi-source accident data based early warning device of the present application;

[0031] Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0032] Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0033] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0034] The early warning method based on multi-source accident data provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can acquire multi-source accident data from the client and generate a vehicle accident database. It performs feature standardization on the accident features in the database to form an accident feature set, performs principal component analysis (PCA) to obtain a dimensionality-reduced feature set, trains a similarity matching model based on the dimensionality-reduced feature set to generate a similarity index structure, receives real-time sensor data streams, parses feature fields to convert them into real-time feature vectors, maps the real-time feature vectors to the similarity index structure to obtain a candidate accident record set, compares the candidate accident record set with the real-time feature vectors to generate a similarity index, and outputs an anomaly warning when the similarity index exceeds a preset threshold. This invention constructs a vehicle accident database and combines it with real-time sensor data streams, uniformly mapping historical accident features and real-time feature vectors to a similarity index structure, enabling rapid retrieval of candidate accident record sets. Then, it generates a similarity index through similarity comparison and triggers anomaly warning information, thereby improving the accuracy of comparison results and the reliability of warnings while ensuring real-time performance. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The present invention will now be described in detail through specific embodiments.

[0035] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the early warning method based on multi-source accident data provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0036] like Figure 2 As shown, the early warning method based on multi-source accident data proposed in this invention includes the following steps:

[0037] S10, acquire multi-source accident data, and generate a vehicle accident database based on the multi-source accident data;

[0038] In this embodiment, the process of acquiring multi-source accident data covers multiple data sources related to vehicle operation. Vehicle state data is usually provided by the on-board diagnostic system, which collects basic operating parameters such as engine speed, vehicle speed, throttle pedal position, brake status, etc. through the vehicle control unit, which can accurately reflect the dynamic characteristics of the vehicle before and after the accident. Collision event sequence data is derived from the event data recorder, which records the physical quantities captured by devices such as acceleration sensors and gyroscopes at the moment of vehicle collision, which can reflect the impact direction and impact strength. Accident environment attribute data comes from accident liability determination documents or on-site investigation materials, including weather conditions, road conditions, visibility, etc. Road topology data is usually obtained by analyzing photos or monitoring videos taken on site, and spatial layout information such as the number of lanes, road curvature, intersection structure, etc. is extracted through image recognition technology. User portrait data is mainly derived from historical claims records, including driver's driving habits, violation records, accident frequency, etc., which can be used as a reference index for driving risk. Accident scene description data is obtained by semantic analysis of the report voice or text, which can supplement the semantic understanding of the accident occurrence.

[0039] In the process of generating the vehicle accident database, it is necessary to associate different sources of data through a unified accident identifier. The accident identifier is usually generated by combining the time, location, and vehicle unique identification code of the accident, which can ensure the logical correspondence of cross-source data. After data association is completed, the original data needs to be cleaned. The cleaning process includes missing value filling, such as using mean interpolation or inference based on similar accident samples; outlier processing, such as smoothing or removing abnormal speed or steering angle caused by sensor noise. The result of data cleaning is then converted into a structured format, converting multi-source heterogeneous data into a unified field set for subsequent standardization processing and modeling operations. Structured data needs to be stored in the database according to the pre-set data mode, which usually consists of field definition, data type and storage rule, to ensure that subsequent operations can be called and extended in a standardized way.

[0040] The storage implementation of the database usually relies on a distributed storage system. When dealing with massive historical accident data, a relational database can be chosen to ensure transaction consistency, or a columnar storage system can be chosen to optimize large-scale retrieval efficiency. In order to adapt to real-time analysis requirements, index structures can be established in the database, such as inverted indexes based on accident identifiers, to support fast query and retrieval. The vehicle accident database generated in this way not only saves multi-source information of historical accidents, but also provides a data foundation for subsequent feature standardization, dimensionality reduction and similarity comparison.

[0041] In different embodiments, the collection and integration of multi-source accident data can vary. For example, in the acquisition of vehicle state data, it can be collected through the OBD-II interface, or more high-frequency and detailed data can be collected directly through the CAN bus. In the extraction of accident environmental attributes, meteorological information can be collected automatically using environmental sensors, or it can be manually input in the form of text records. In the generation of road topology data, it can be extracted from two-dimensional photos through image recognition technology, or a three-dimensional road model can be reconstructed through laser radar data.

[0042] In the data cleaning operation, a statistical distribution-based anomaly detection method can be used, or a machine learning method can be introduced to automatically identify abnormal samples through training of a classification model. In the handling of missing values, interpolation methods can be used to fill in time series missing values, or nearest neighbor-based prediction methods can be used. For the structured storage of the database, traditional relational databases can be selected, or graph databases can be used to express complex road topology and accident relationships. In high-concurrency application scenarios, the database can also be combined with a memory computing engine to shorten the delay of data calling and analysis.

[0043] Example: In the vehicle insurance scenario in the field of financial technology business, insurance companies often need to quickly confirm whether the accident situation is highly similar to historical cases during the claim settlement audit process. By storing vehicle state data and accident environmental attributes and other multi-source data in the vehicle accident database, claim settlement personnel can use the database to perform feature matching on newly occurring accidents. For example, when the system receives a report case voice and generates an accident scene description, it can retrieve the corresponding road topology and driver historical portrait in the database, thereby more quickly identifying the similarity of the accident to high-frequency risk accidents in the past, improving the efficiency and accuracy of claim settlement audit and risk warning.

[0044] This embodiment can achieve full-dimensional coverage of accident information by uniformly associating and structuring multi-source data such as vehicle state, collision events, environmental conditions, road topology, driver portraits, and accident scene descriptions. The unified vehicle accident database not only ensures the integrity and consistency of historical data, but also provides a reliable data foundation for subsequent feature extraction, standardization, dimensionality reduction, and similarity modeling, effectively solving the problem of inaccurate comparison results caused by insufficient information from a single data source.

[0045] S20, performing feature standardization processing on the accident features in the vehicle accident database to obtain an accident feature set;

[0046] In this embodiment, the vehicle accident database contains a large number of accident feature fields, which are derived from multiple sources of data, including vehicle status, collision events, accident environment, road topology, user profiles, and accident scene descriptions. Data from different sources differ significantly in their collection methods, units, and representations; therefore, these accident features need to be standardized before subsequent calculations and comparisons.

[0047] The first step is to extract the original accident feature fields. Database fields often include both continuous and discrete variables. For example, vehicle speed, braking distance, and steering angle are continuous variables, while accident type, weather conditions, and road grade are discrete variables. In practical processing, it is necessary to identify the data type of the fields in order to apply different standardization methods.

[0048] For numerical features, maximum and minimum value normalization can be performed to map the original values ​​to a preset interval, such as [0,1] or [-1,1], thereby eliminating differences between different units of measurement. For example, vehicle speed may be expressed in km / h, and braking distance may be expressed in meters. Normalization allows them to be compared on a uniform scale.

[0049] For categorical features, integer encoding can be used to transform different category values ​​into ordered numerical representations. For example, in the weather conditions field, sunny, rainy, and foggy days can be mapped to 0, 1, and 2 respectively, and in the accident liability field, full liability, equal liability, and no liability can be mapped to 0, 1, and 2 respectively. In this way, discrete text or symbolic data can be transformed into a computable form.

[0050] After processing numerical and categorical features separately, it is necessary to verify the consistency of dimensions. Since the number and order of fields may vary across different data sources, field alignment and dimension checks are required to ensure the final feature vectors are structurally consistent. Dimension consistency verification can be achieved by calculating whether the vector length equals the preset number of feature dimensions. If missing or redundant fields are found, they can be corrected by imputation, removal, or recoding.

[0051] Finally, the normalized and encoded feature fields are combined to generate a standardized accident feature set. This set has a unified numerical scale and structured dimensional representation, which can be directly input into the dimensionality reduction and similarity modeling stages, ensuring the comparability of data from different sources.

[0052] In different implementations, the standardization process can be adjusted according to the application scenario. For example, when standardizing numerical features, maximum and minimum value normalization can be used, or the Z-score standardization method can be used to scale the features using the mean and standard deviation to adapt to different distribution characteristics.

[0053] When encoding categorical features, integer encoding or one-hot encoding can be used to map categories to sparse vectors, which can improve the discriminative power of the representation when there are a large number of categories. If the feature dimension is too high, a category merging strategy can be used to merge rare categories to reduce dimensionality inflation.

[0054] In the dimensional consistency verification stage, the number and order of fields can be constrained through database field definition files, or feature dimensions can be dynamically detected in the data processing program, automatically filling in default values ​​when missing values ​​are found and removing redundant fields. For scenarios requiring cross-database alignment, a feature mapping table can be introduced to unify field naming and numbering.

[0055] In high-concurrency scenarios, standardized processing can also be performed in batch mode, which can simultaneously perform data type identification, normalization and encoding operations on multiple accident records, thereby improving efficiency.

[0056] Example Explanation: In the auto insurance scenario within the fintech business, insurance companies need to perform risk modeling and similarity comparison based on historical accident data. If unprocessed raw data is used directly, the speed field might be expressed in km / h, while the braking distance field is expressed in meters, causing discrepancies in similarity calculations due to different units of measurement. Simultaneously, weather fields are recorded as "sunny" or "rainy" in text form, making direct numerical calculations difficult. Through feature standardization, fields such as speed and braking distance can be normalized to a unified range, weather fields can be encoded into a computable numerical representation, and the final generated feature vectors can be ensured to be consistent in dimension. This allows different accident data to be compared on the same scale and with the same structure, improving the accuracy and reliability of the comparison.

[0057] This embodiment eliminates dimensional differences and representational inconsistencies between different data sources by standardizing accident features in the vehicle accident database, ensuring the comparability and effectiveness of the feature set in subsequent calculations. The standardized accident feature set provides a unified input condition for dimensionality reduction and similarity modeling, improving the stability of data processing and the accuracy of comparison results.

[0058] S30, perform principal component dimensionality reduction on the accident feature set to obtain a dimensionality-reduced feature set;

[0059] In this embodiment, the accident feature set contains standardized multi-dimensional feature vectors. These vectors may have a high-dimensional structure, such as hundreds of dimensions. Directly using them for similarity comparison would result in excessive computational overhead and may introduce noise. To improve computational efficiency and retain key information, principal component dimensionality reduction processing needs to be performed on the accident feature set.

[0060] First, it is necessary to determine the covariance matrix of the accident feature set. The covariance matrix is ​​a measure of the correlation between various features and can reveal the linear dependencies between different features. In a practical implementation, a matrix-form sample set can be constructed based on the feature set, and the covariance between each dimension can be calculated after mean centering.

[0061] Then, eigenvalue decomposition is performed on the covariance matrix to obtain a set of eigenvalues ​​and corresponding eigenvectors. The eigenvalues ​​represent the variance information in the corresponding direction, while the eigenvectors represent the new coordinate axis directions after dimensionality reduction. In this process, larger eigenvalues ​​mean that more data information can be retained in that direction.

[0062] Next, the eigenvectors are sorted according to their eigenvalues ​​to form an ordered sequence of principal components. Typically, the eigenvalues ​​are arranged from largest to smallest, and the corresponding eigenvectors are arranged sequentially. This ensures that the first few principal components can explain most of the variance in the data.

[0063] In an ordered principal component sequence, a threshold is set based on the cumulative variance contribution rate, such as 95% or 99%. The top few principal components that can cover this contribution rate are selected to form a principal component subset. This process can reduce feature dimensionality while preserving as much data information as possible, and reduce the impact of redundant features and noise.

[0064] Finally, the accident feature set is projected onto a low-dimensional space defined by the principal component subsets to obtain the projected feature representation, i.e., the dimensionality-reduced feature set. This set retains the main variation information of the accident features and significantly reduces the dimensionality, thus providing efficient input data for subsequent similarity modeling and index construction.

[0065] In different implementations, dimensionality reduction can be achieved using different methods. For example, it can be based on traditional principal component analysis (PCA) and achieved through eigenvalue decomposition; or it can use singular value decomposition (SVD), which has better numerical stability on large-scale datasets. For scenarios with high real-time requirements, incremental PCA can be used to dynamically update the principal component subsets as new data arrives, without needing to recalculate all the data.

[0066] Regarding threshold settings, the cumulative variance contribution rate threshold can be adjusted according to different application scenarios. In risk analysis scenarios that require high accuracy, a higher contribution rate threshold can be set, such as 99%; in early warning scenarios that emphasize real-time response, a lower threshold, such as 90%, can be selected to further reduce the computational burden.

[0067] For high-dimensional sparse features, feature selection methods can be combined with principal component analysis. First, feature selection is used to remove low-weight or low-correlation features, and then principal component analysis is performed to reduce computational complexity and noise interference.

[0068] In a distributed environment, the efficiency of processing large-scale data can be improved by parallel computation of the covariance matrix and eigenvalue decomposition. For resource-constrained edge devices, principal component subsets can be pre-computed in the cloud, and projection operations can be performed only on the edge device, reducing local computational overhead.

[0069] This embodiment significantly reduces the feature dimensionality by performing principal component analysis (PCA) on the accident feature set, thereby decreasing the computational complexity of subsequent model training and similarity comparison, while preserving the main variable information in the accident features to the greatest extent possible. This not only improves the efficiency of similarity modeling but also enhances the model's robustness to noise and redundant features. The resulting dimensionality-reduced feature set achieves higher-precision similarity matching while maintaining computational efficiency.

[0070] S40, Train a similarity matching model based on the reduced-dimensionality feature set to generate a similarity index structure;

[0071] In this embodiment, the dimensionality-reduced feature set, even after principal component projection, is still only a low-dimensional feature vector set. A similarity matching model is still needed to support subsequent rapid retrieval and comparison. The core of training the similarity matching model lies in constructing a data structure that can effectively organize and index feature vectors, thereby enabling the rapid identification of similar historical accident records when faced with real-time input vectors.

[0072] Specifically, the first step is to set parameters for the Locality Sensitive Hash (LSH) model, including hash length and the number of hash tables. Hash length defines the number of bits in a single hash signature, determining the fineness with which feature vectors are divided into hash buckets; the number of hash tables affects the overall recall and precision of the retrieval. By adjusting these two parameters, a balance can be achieved between computational overhead and retrieval accuracy.

[0073] Next, an initial set of hash functions needs to be generated. This is typically achieved using a randomized hyperplane approach, where each hyperplane defines a binary classification discriminant function to determine the hash value of the feature vector in that dimension. The set of hash functions formed by combining multiple hyperplanes can project high-dimensional vectors onto a low-dimensional hash signature space.

[0074] Next, the initial set of hash functions is trained using the dimensionality-reduced feature set to update its projection parameters. The training process typically employs an iterative optimization method, ensuring that similar vectors are mapped to the same or nearby hash buckets in the hash space, while dissimilar vectors are distributed across different hash buckets.

[0075] After training, a hash table storage structure can be constructed based on the trained set of hash functions. Each feature vector generates a corresponding hash signature through a hash function and maps it to a specific hash bucket in the hash table, forming a correspondence between vectors and hash buckets.

[0076] To validate the model's effectiveness, an accuracy evaluation of approximate nearest neighbor search needs to be performed on the hash table storage structure. This evaluation typically involves setting aside a portion of the labeled data as a validation set and calculating the proportion of actual similar vectors that are correctly retrieved. If the accuracy falls below a preset performance threshold, the hash length or the number of hash tables needs to be readjusted, and the training process repeated.

[0077] When the model's approximate nearest neighbor search accuracy on the validation set reaches or exceeds a preset threshold, the projection parameters of the trained hash function set need to be frozen to prevent drift during subsequent use. Finally, the trained hash function set is combined with a hash table storage structure to form a persistent similarity index structure, providing a foundation for real-time vector mapping and retrieval.

[0078] The methods for training similarity matching models can differ depending on the implementation. Traditional stochastic hyperplane methods can be used to generate hash functions, while deep learning-based hash embedding models can be employed, using neural networks instead of hyperplanes to achieve nonlinear mapping. Regarding parameter selection, adjustments can be made based on the actual data volume and computational resources. For example, in large-scale data scenarios, the number of hash tables can be appropriately increased to improve recall, while in resource-constrained devices, the number of tables can be reduced to decrease memory consumption.

[0079] Efficiency can be improved through distributed training, which divides the dimensionality-reduced feature set into multiple nodes for parallel computation, and finally merges them to generate a unified set of hash functions and a hash table storage structure. In scenarios with high real-time requirements, an incremental update approach can be adopted, retraining local hash functions on new data at regular intervals instead of retraining the entire dataset.

[0080] In addition, in terms of storage structure design, a compact bit array structure or sparse hash mapping can be used to reduce memory usage, while a persistence mechanism can be combined to ensure that the index structure can still be loaded quickly after the system restarts.

[0081] This embodiment trains a similarity matching model on a dimensionality-reduced feature set and generates a similarity index structure, enabling the construction of an efficient approximate nearest neighbor retrieval mechanism while maintaining feature expressive power. This process significantly reduces retrieval time complexity, allowing millisecond-level similarity queries even in a database of millions of accident records. Simultaneously, the frozen hash function set ensures the consistency and stability of the retrieval, avoiding the risk of results becoming invalid over time, thus providing reliable data support for real-time early warning.

[0082] S50, acquire real-time sensor data stream, parse the feature fields in the real-time sensor data stream, and convert the feature fields into real-time feature vectors;

[0083] In this embodiment, the real-time sensor data stream refers to the continuous data stream generated by multiple different types of sensors during vehicle operation. This data may originate from driver actions, vehicle operating status, and external environmental conditions. To enable this raw data to enter the subsequent comparison and retrieval stages, it needs to be parsed into feature fields and further converted into computable real-time feature vectors.

[0084] First, multi-source real-time sensor data streams are integrated into the edge processing device. Different types of sensors often have different data transmission protocols and formats. For example, a driver behavior sensor may output angle values ​​and displacement, a vehicle status sensor may output rotational speed and velocity, and an environmental information sensor may output illuminance or humidity. This data needs to be uniformly accessed through a single channel and preliminarily decoded to ensure that the data can be correctly identified.

[0085] Secondly, specialized parsing operations are performed on different types of data streams to transform the raw data into semantically clear feature fields. For example, steering wheel angle and brake pedal travel are extracted from driver behavior data, vehicle speed and engine speed are extracted from vehicle status data, and light intensity and road slippage are extracted from environmental information data. The parsing process needs to combine the raw dimensional information provided by the sensors to ensure that the feature fields maintain numerical accuracy and comparability.

[0086] After feature field extraction, dimension alignment is required. Dimension alignment refers to matching features from different sources according to the feature dimension definitions in the vehicle accident database, ensuring that the parsed fields are consistent with historical data in terms of the number, order, and format of dimensions. For example, if the database defines the accident feature field order as "steering wheel angle, brake travel, vehicle speed, engine speed, light intensity, and slipperiness," then the real-time parsed feature fields need to be arranged in the same order.

[0087] Finally, real-time format conversion needs to be performed in the memory buffer of the edge processing device to combine the dimension-aligned feature fields into a numerical vector representation. This conversion process typically involves numerical normalization (such as converting angles to radians and mapping slippage levels to numerical values) and combining all fields into a vector form so that it can be used as computational input. The use of a memory buffer can reduce disk I / O latency and meet millisecond-level real-time requirements.

[0088] Parallel parsing can be employed, distributing driver behavior, vehicle status, and environmental data streams to different threads or processing cores to improve overall processing speed. Alternatively, a unified data decoding middleware can be used to convert sensor data from different protocols into a standardized message format before entering the parsing logic.

[0089] During dimensional alignment, a dynamic mapping table can be established. When new feature dimensions are added to the accident database, the real-time parsing module can dynamically adapt according to the mapping table, avoiding manual modification of the parsing logic. Scaling and normalization strategies for different fields can also be set parametrically to flexibly adapt to different vehicle models or sensor accuracies.

[0090] In the real-time format conversion stage, sparse vector representation can be chosen to reduce memory overhead, or fixed-point numbers can be used instead of floating-point numbers to improve computation speed. If the scenario requires extremely low latency, the data can also be written to shared memory for direct use by the similarity matching model, avoiding multiple data copies.

[0091] This embodiment ensures consistency in format and dimension of data from different sensors by accessing real-time sensor data streams and parsing feature fields. After further conversion into real-time feature vectors, direct comparison with a historical accident database is possible. This approach not only improves the real-time performance and accuracy of data processing but also reduces matching errors caused by data source heterogeneity, providing a solid foundation for subsequent similarity comparisons.

[0092] S60, map the real-time feature vector to the similarity index structure to obtain a candidate accident record set;

[0093] In this embodiment, the process of mapping real-time feature vectors to a similarity index structure essentially utilizes an indexing mechanism to achieve rapid approximate matching, thereby filtering out accident records from the vehicle accident database that may be similar to the real-time scenario. This operation needs to be decomposed into three levels: computation, storage, and retrieval.

[0094] First, real-time feature vectors need to be processed using a Locality Sensitive Hash (LSH) function to generate hash signatures. LSH functions maintain the similarity between high-dimensional vectors, ensuring that similar objects still fall into the same or nearby storage units after the vectors are mapped to a low-dimensional hash space. The hash function set trained on the real-time feature vectors and the dimensionality-reduced feature set uses the same projection parameters to guarantee the comparability between real-time data and database features.

[0095] Secondly, the generated hash signature is used for retrieval in a hash table storage structure with a similar index structure. The hash table storage structure consists of multiple hash buckets, each storing a set of incident record indexes under the corresponding signature. By calculating the target hash signature of the real-time feature vector and searching for hash buckets that match it or whose Hamming distance is within a certain threshold, potentially similar incident candidates can be quickly located.

[0096] Next, an incident record index set is extracted from the matched hash buckets. Each index in the index set corresponds to an incident record in the database, and the extracted set typically contains several potentially similar incident events. To avoid too many candidate results or the inclusion of outdated and worthless data, further retrieval of the extracted index set is necessary.

[0097] Once the accident record index set is determined, the system retrieves the corresponding accident records from the vehicle accident database. This operation typically involves database index access and batch query mechanisms to ensure data retrieval is completed within milliseconds.

[0098] Finally, all retrieved incident records are aggregated to form an initial candidate incident record set. This set may still contain a large number of records after aggregation, so a further filtering mechanism, such as time window filtering, is needed to retain recent incident records, giving the candidate set stronger real-world relevance. Ultimately, the filtered set becomes the candidate incident record set, providing input for subsequent similarity comparisons.

[0099] Mapping efficiency can be improved by using parallel retrieval across multiple hash tables. This involves generating multiple hash signatures for each real-time feature vector and querying them in parallel across multiple hash tables to increase recall. Alternatively, an improved hash bucket selection strategy can be employed. This strategy not only searches for buckets that perfectly match the target signature but also calculates the Hamming distance between adjacent buckets. If the distance is within a set threshold, the bucket is also included in the search, thus avoiding missing boundary cases.

[0100] When aggregating candidate records, a weighted aggregation strategy can be used to assign different weights to candidate results from different hash tables, preventing records in certain low-quality buckets from affecting the overall candidate set. Dynamic time window settings can also be used to adjust the window length based on the risk level of the vehicle's real-time operating environment. For example, the filtering time window can be shortened in highway scenarios to highlight recent accident characteristics, while it can be appropriately extended in low-speed urban areas.

[0101] For database retrieval, an in-memory indexing mechanism can be used to preload the incident index set into the cache, thereby reducing disk query latency. For larger-scale processing, incident records can be stored in shards using a distributed database system, with each shard node responding to query requests in parallel, and the result sets ultimately merged.

[0102] This embodiment achieves fast approximate search in high-dimensional data environments by mapping real-time feature vectors to a similarity index structure, avoiding the high latency caused by linear scanning. The introduction of a candidate accident record set significantly reduces the computational scale of subsequent similarity comparisons, ensuring real-time requirements. Combined with a time window filtering mechanism, the timeliness and relevance of the candidate set can be improved, thus providing more accurate reference for anomaly warnings.

[0103] S70, perform a similarity comparison between the candidate accident record set and the real-time feature vector to generate a similarity index, and output an abnormal warning message when the similarity index exceeds a preset threshold.

[0104] In this embodiment, the purpose of comparing the candidate accident record set with the real-time feature vector is to map streaming observations to a historical experience space and form a quantifiable similarity index. Then, based on the comparison result of the similarity index with a preset threshold, an anomaly warning is triggered. The candidate accident record set refers to a subset of historical accident samples retained after index retrieval and time filtering. Each sample contains an accident feature vector and structured fields such as accident type label, timestamp, and source credibility. This set originates from the previous retrieval process and maintains the same dimensions and scale as the real-time feature vector. The real-time feature vector refers to the current scene representation after edge-side parsing and format conversion. The field order and value range are aligned with the feature definitions in the vehicle accident database to facilitate direct similarity calculation. Similarity comparison refers to performing measurement calculations and set aggregation on two types of vectors in the same feature space, generating a similarity index through a two-layer structure of single-sample measurement and set-level fusion. Similarity metrics refer to comparable scalars or scalar pairs that reflect the closeness and type consistency between real-time feature vectors and candidate accident record sets. Sources include, but are not limited to, cosine similarity, weighted Euclidean similarity, Mahalanobis similarity, and hash collision ratio. Time decay weights and source credibility weights can be introduced to form a combined score. Preset thresholds refer to the boundary values ​​or intervals obtained after offline calibration using historical sample annotation sets. These can be single values ​​or multi-level intervals, used to map similarity metrics to the determination of whether to output abnormal warning information. Abnormal warning information refers to structured messages for vehicle-mounted human-machine interfaces and remote risk control platforms. It typically includes fields such as warning type label, similarity metrics, threshold number or name, timestamp, candidate sample summary, and subsequent processing channel identifier, facilitating further processing in the downstream decision-making chain.

[0105] The similarity comparison process is organized hierarchically to balance real-time performance and stability. First, the real-time feature vectors are normalized to a unit norm or recalibrated according to the training period's dimensional table to eliminate scale shifts caused by instantaneous sensor fluctuations. Then, feature vectors are extracted from each candidate accident record, and single-sample similarity is calculated using a selected metric to obtain an individual score sequence. These individual scores, along with candidate sample timestamps, source credibility, and accident type matching, are fed into a weighted function to form a weighted score sequence. Next, individual scores and weighted scores are aggregated at the set level. Aggregation strategies can include weighted mean, weighted quantiles, maximum values, or multi-component fusion (e.g., linearly combining the principal similarity component with the type consistency component) to obtain the final similarity index. To reduce noise, individual scores can be truncated or robustned before aggregation to remove extreme values. After index calculation, it is compared with a preset threshold; the comparison relationship can be greater than, within an interval, or across a threshold, as defined in the offline calibration file. When the comparison relationship is true, an anomaly warning message is generated. The message includes the index of the closest sample and a snapshot of key fields for tracing and verification, in order to support subsequent verification and closure.

[0106] Regarding metric selection, cosine metrics can be used for scenarios with high dimensionality and sparseness or strong directionality, providing a robust assessment of directional similarity. For scenarios where the noise covariance structure is available and sensitive to scale changes, Mahalanobis metrics can be used, enhancing discriminative capabilities through offline estimation of the covariance inverse matrix. For scenarios where hash indexes are already constructed and require rapid approximation, the hash collision ratio can be directly used as the initial screening score, then fused with vector metrics to improve throughput. In terms of aggregation, road segments with rapid speed changes and low false alarm costs can be aggregated with maximum values ​​to improve sensitivity; road segments in urban areas with slow traffic and high false alarm costs can be aggregated with weighted quantiles or truncated means to stabilize output. Time decay weights can be automatically calculated based on the accident record timestamps using an exponential decay function, with records closer to the current time receiving higher weights. Source credibility weights can be generated based on the quality level of the collection channel or historical consistency rate, reducing the impact of low-quality sources.

[0107] Regarding threshold configuration, multiple versions of threshold tables can be established based on road type, weather zone, and time period zone. The edge side loads matching threshold entries based on the current environmental label. Alternatively, slow adaptive processing can be implemented while ensuring stability. This involves updating the quantile threshold using the distribution of non-accident samples over a recent period, while locking the positive sample window during the accident period for verification to avoid drift. To address the scenario where the candidate accident record set is empty, the approach can degenerate to distance determination from the local reference center or directly enter a low-level prompt channel to avoid a state with no output. To improve latency performance, accident feature vectors within the candidate set can be resident in memory and grouped by type. During real-time comparison, calculations are only performed within the matching type or adjacent type groups. Simultaneously, fixed-point arithmetic or a half-precision vector library can be used to run on the edge chip, reducing the resource consumption caused by multiplication and addition. To enhance robustness, a false negative defense strategy can be introduced for real-time feature vectors. For example, short-cycle recalculation can be triggered when a single comparison result is close to the threshold, and small-window majority voting can be performed on multiple results to reduce the probability of instantaneous noise triggering.

[0108] Regarding message generation, anomaly warning information adopts a fixed-field serialization format. Fields include warning type label, similarity index, threshold identifier, timestamp, index of the closest sample, summary table of the closest samples, visual alert level, and suggested action identifier. After generation at the edge, it is immediately written to a local reliable queue and pushed in parallel to the vehicle display and remote interface. If the network is unavailable, the most recent messages are retained and replayed with timestamps to ensure traceability. To facilitate offline review, a key feature comparison table for explanation can be included in the message, such as side-by-side values ​​for vehicle speed, steering angle, braking distance, and light intensity, to aid in manual verification.

[0109] This embodiment compresses the candidate accident record set into a single similarity index through similarity comparison and aggregation. This index is then compared with a preset threshold to determine whether to output an abnormal warning. With the candidate size much smaller than the entire database, the computational complexity is approximately linearly related to the number of candidates, significantly reducing latency and resource consumption. The hierarchical design of metrics and aggregation ensures stable performance under different road environments and data quality conditions. Multiple versions of the threshold table and time decay weights further enhance timeliness and controllability. Thus, while ensuring real-time performance, the false alarm rate and false negative rate can be kept within acceptable ranges through offline calibration and online fine-tuning.

[0110] This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as fintech. It discloses a method, device, equipment, and medium for early warning based on multi-source accident data, comprising: acquiring multi-source accident data and generating a vehicle accident database; performing feature standardization processing on accident features in the vehicle accident database to form an accident feature set; performing principal component dimensionality reduction on the accident feature set to obtain a dimensionality-reduced feature set; training a similarity matching model based on the dimensionality-reduced feature set to generate a similarity index structure; receiving real-time sensor data streams and parsing feature fields to convert them into real-time feature vectors; mapping the real-time feature vectors to the similarity index structure to obtain a candidate accident record set; performing similarity comparison between the candidate accident record set and the real-time feature vectors to generate a similarity index; and outputting an anomaly warning information when the similarity index exceeds a preset threshold. This invention constructs a vehicle accident database and combines it with real-time sensor data streams, uniformly mapping historical accident features and real-time feature vectors to a similarity index structure to achieve rapid retrieval of candidate accident record sets. Then, it generates a similarity index through similarity comparison and triggers anomaly warning information, thereby improving the accuracy of comparison results and the reliability of warnings while ensuring real-time performance.

[0111] In one embodiment, step S10 above includes:

[0112] S101: Obtain vehicle status data from the vehicle diagnostic system, obtain collision event sequence data from the event data recorder, obtain accident environmental attribute data from the accident liability determination letter, extract road topology data from on-site inspection photos, obtain user profile data from claims records, and parse accident scene description data from the report voice text.

[0113] S102, by associating the vehicle status data, collision event sequence data, accident environment attribute data, road topology data, user profile data, and accident scene description data with the accident identifier, an associated accident record is generated;

[0114] S103, Perform data cleaning operation on the associated accident records to process missing values ​​and outliers, and generate cleaned accident records;

[0115] S104, convert the cleaned accident record into a structured data format;

[0116] S105, Store the cleaned accident records in the structured data format according to the preset data mode to generate a vehicle accident database.

[0117] In this embodiment, the goal of acquiring multi-source accident data is to form a comprehensive observation set covering vehicle status, collision process, environmental conditions, road structure, insurance and claims profiles, and verbal descriptions, providing stable input for subsequent unified modeling. Vehicle status data comes from the vehicle diagnostic system, and fields typically include vehicle speed, engine speed, throttle opening, brake switch, steering angle, longitudinal and lateral acceleration, and stability system intervention markers. During acquisition, timestamps are recorded using the vehicle controller time base, and the raw frames are organized into a time sequence according to the vehicle identification code and acquisition channel label. Collision event sequence data comes from the event data recorder, containing high-frequency samples in short time windows before and after the collision, and fields cover airbag ignition, collision pulse acceleration, collision direction, and restraint system status. To ensure comparability, the start and end times of the event window need to be aligned with the time axis of the vehicle status data. Accident environmental attribute data comes from the liability determination report. Structured fields focus on accident location, road grade, weather conditions, visibility, ground conditions, traffic control methods, number of involved parties, and basic information. Due to the mixture of text and tables, a combination of layout parsing and natural language parsing is used to extract key phrases from headers, footers, form cells, and clause paragraphs, and place names are mapped to latitude and longitude. Road topology data is extracted from on-site survey photos, including the number of lanes, lane line types, intersection shapes, entrance and exit locations, sign and marking elements, obstructions, and guardrails. Perspective relationships are restored through geometric correction and distortion correction, and road elements are identified through object detection and semantic segmentation. A planar approximate topology is generated using pixel coordinates and camera pose estimation. User profile data comes from claims records, focusing on vehicle usage frequency, travel time distribution, historical claim tags, vehicle age, and policy information, emphasizing behavioral and asset characteristics related to accident risk. This part needs to be aligned with privacy and compliance policies, retaining only fields necessary for modeling. The accident scene description data comes from the voice text of the report. First, the voice is transcribed and colloquialized. Then, event extraction and spatiotemporal element extraction are used to convert the subject, action, time, location, road elements and relative motion relationships in the narrative into structured fragments, and cross-references are established with objective observations from other sources.

[0118] The purpose of linking accident records is to aggregate observations from multiple channels into a complete view of the same accident. An accident identifier, serving as a globally unique key, is generated by combining the nearest neighbor intervals of occurrence time, spatially adjacent areas, vehicle identification codes, and data collection terminal markers. Rules include time overlap verification, spatial distance threshold verification, vehicle identifier consistency verification, and trigger event consistency verification. When multiple candidates meet the conditions, a conflict resolution strategy selects the key-value pair with the highest confidence, and a source list and confidence weight are written into the record. Subsequently, vehicle status data, collision event sequence data, accident environmental attribute data, road topology data, user profile data, and accident scene description data are linked using the accident identifier to generate a linked accident record. This record, at the accident level, retains index pointers to the original time-series segments and snapshots of summarized statistical features, ensuring both low-latency queries and the ability to trace back to original evidence.

[0119] Data cleaning operations revolve around four categories of issues: missing data, anomalies, duplication, and conflicts. For missing data, short-duration intermittent sensor interruptions are filled using neighborhood interpolation or similar sample conditions. Essential tags are set for core fields affecting critical judgments; missing fields are marked and added to the quality assessment list. For anomalies, unreasonable ranges or jumps are judged using a combination of physical constraints and equipment status. Samples that do not meet mechanical and road conditions are removed, and confidence reduction weights are added to samples with critical values. For duplication, fields repeatedly reported from multiple channels are recorded using a strategy combining time priority and source priority, retaining one authoritative record, while other records are retained as redundant backups. For conflicts, when text and sensor sources contradict each other on key elements (e.g., collision direction, weather), a judgment score is constructed based on source credibility, evidence quantity, and evidence consistency. The higher score is selected, and a conflict summary and selection criteria are recorded for later review. After cleaning, a post-cleaning incident record is generated, containing quality tags, an anomaly list, and processing traces, ensuring traceability throughout the entire process.

[0120] The structured data format clearly defines field names, data types, value ranges, time and spatial coordinate systems, units and dimensions, encoding dictionaries, and reference relationships. Time uses a unified time base and time zone standard, space uses a unified coordinate system and completes map engine matching, and units are stored in a unified dimension and the original unit mapping is recorded. Text fields provide word segmentation and phrase standardization results and original text mapping pointers, while image-derived fields provide conversion parameters from pixel measurements to physical measurements. To reduce coupling and improve scalability, a versioned field list and compatibility strategy are introduced; new fields are filled with blanks in older versions without disrupting existing queries. The transformation process maps the cleaned accident records to concrete instances of the structured data format, outputting both fact tables and dimension table references, such as entity relationships like accident facts, vehicle dimensions, driving behavior dimensions, road dimensions, environmental dimensions, and text event dimensions.

[0121] The default data model defines constraints and indexing strategies for long-term operation of the storage organization. The storage organization balances row-based and column-based approaches based on query and analysis needs; accident facts for high-concurrency writes and conditional retrieval are arranged in a row-based layout and bound to primary keys and time indexes; field sets for statistical aggregation and model training are arranged in a column-based layout and pruned by partitions, splitting according to partition keys such as accident occurrence time, regional grid, and road grade to reduce the scanning range. Secondary indexes required for similarity retrieval and spatiotemporal queries include time range indexes, spatial grid indexes, accident type inverted indexes, and vehicle identification code indexes; references to image and audio derivatives are linked to foreign keys and object storage keys to avoid directly placing large amounts of binary data into transaction tables. The write path uses an idempotent interface; transaction logs verify the uniqueness of accident identifiers and the necessity of fields before writing to the database, and failed records are placed in a waiting queue and replayed periodically; after writing to the database, change events are generated and delivered to downstream training and index building channels to maintain data flow consistency. At this point, the vehicle accident database is completed and has the dual capability of supporting both online retrieval and offline training.

[0122] This embodiment establishes a stable and traceable data foundation for the vehicle accident database by uniformly collecting multi-source accident data, cross-source association based on accident identifiers, cleaning for missing and anomalies, and structural transformation and hierarchical storage for analysis and retrieval. Inconsistencies in time, space, and dimensions caused by differences in sources are eliminated at the entry and transformation stages; cross-source conflicts are recorded and adjudicated at the association and cleaning stages; and performance targets for querying and training are achieved through preset data patterns and multi-level indexes. Therefore, subsequent standardization, dimensionality reduction, similarity retrieval, and early warning stages do not require repeated processing of entry-level noise and structural differences, resulting in reduced overall latency and resource consumption, improved sample coverage and field completeness, and enhanced usability and interpretability of historical experience.

[0123] In one embodiment, step S20 above includes:

[0124] S201, Extract the original accident feature fields from the vehicle accident database;

[0125] S202, identify the characteristic data type of the original accident characteristic field to distinguish between numerical characteristic fields and categorical characteristic fields;

[0126] S203, Perform maximum and minimum value normalization processing on the numerical feature field to map the feature value to a preset normalization interval to obtain the normalized numerical feature field;

[0127] S204, Perform integer encoding processing on the categorical feature field to convert discrete feature values ​​into ordered numerical representations, and obtain the encoded categorical feature field;

[0128] S205, verify the dimensional consistency between the normalized numerical feature field and the encoded categorical feature field;

[0129] S206. Combine the normalized numerical feature fields and the encoded categorical feature fields after successful verification to generate an accident feature set.

[0130] In this embodiment, when extracting original accident feature fields from the vehicle accident database, a field mapping list is used as a constraint. The list records the field name, source table, time granularity, unit, value range, and default strategy, and also provides a fixed order of field to location index. The extraction process prioritizes records with timestamps and vehicle identification information, aggregates fields with the same name from different sources into a single column according to the mapping list, and retains the source pointer for easy backtracking; for measurements with units, unit unification is completed first before proceeding to subsequent processing; for text-type enumerations, token standardization and abnormal character cleanup are used to remove non-semantic differences, ensuring that synonymous writing is mapped to the same tag; for time-type fields, if they are used as direct input, they have been split into derived items such as intraday time period, weekday indication, and holiday indication before entering this stage. During the extraction stage, only these derived items are loaded according to the mapping list to avoid repeated derivation in this stage.

[0131] When identifying the characteristic data type of the original accident feature fields, two types of information are used for collaborative decision-making: metadata annotation and data distribution characteristics. Metadata provides the field declaration type and allowed value set, serving as the primary basis. When metadata is missing or ambiguous, data distribution tests are used for judgment. Numerical feature fields are characterized by a predominance of samples that can be parsed into numerical values, an interquartile range greater than zero, and a limited proportion of discrete values. Categorical feature fields are characterized by a finite and stable set of unique values, and a one-to-one mapping between the character set and the dictionary table. Boolean and switch-type observations are processed as categorical feature fields; encoded numerical values ​​such as levels and gears are also classified as categorical feature fields if they do not constitute continuous measurements. After type identification is completed, type labels consistent with the field mapping list are output, and the identification basis and version are recorded so that the same judgment criteria can be used for training and inference.

[0132] Before performing maximum / minimum value normalization on numerical feature fields, outlier screening and zero-span protection are performed. Outlier screening prunes values ​​based on reasonable upper and lower bounds, derived from business-allowed ranges and historical quantile intervals. Pruning records are retained in the quality markers for later weight adjustments. Zero-span protection addresses cases where the maximum and minimum values ​​of a sample in a certain field are equal, employing constant mapping to the interval center or regression to historical full parameters to avoid division by zero risk. The normalization interval uses fixed boundaries and is associated with a parameter table. The parameter table stores estimates of minimum and maximum values ​​and their timestamps at the field level. The training and online sides reference the same parameter table snapshot to ensure consistency. During mapping, calculations are performed column-by-column in field order, retaining the parameter version used for each mapping, allowing for restoration of the transformation during data auditing and troubleshooting.

[0133] When performing integer encoding on categorical feature fields, an encoding dictionary and unknown value strategy are first constructed. The encoding dictionary is maintained on a field-by-field basis, including category labels and corresponding integer numbers, number deactivation labels and effective times, ensuring stable and traceable numbers. When a new category appears, a new number is assigned according to incremental rules, and historical numbers are not reordered. To control the tail of sparse discrete values, extremely low-frequency categories can be merged into general numbers during the dictionary construction stage. Unknown values ​​and missing samples are uniformly mapped to reserved numbers, and an unknown source label is added to the quality label, which can be assigned independent processing weights later during the modeling stage. During the encoding process, the column order of the field mapping list is strictly followed, and the number space is not shared across fields to avoid cross-column confusion. The encoded output is an integer sequence, but it only serves as a numerical representation of discrete identities and does not introduce size order meaning in subsequent similarity and dimensionality reduction processing. If necessary, the encoded values ​​can be rescaled by combining weights or frequencies to make the projection of discrete numbers in the numerical space more stable.

[0134] When verifying the dimensional consistency between the normalized numerical feature fields and the encoded categorical feature fields, a dual verification method is employed. The first verification is structural, checking whether the length of the vector after concatenation of the current sample matches the total length given in the field mapping list, and whether the data types of each column match the expected types. The second verification is content, checking whether the proportion of null values, out-of-bounds values, and unknown codes in key columns are within the threshold range, and generating alarm entries for any detected anomalies and adding them to the monitoring queue. Dimensional consistency relies on a unified field order index. After successful verification, an immutable column order hash is written for rapid comparison in subsequent stages, avoiding systematic bias caused by column misalignment during the similarity calculation stage.

[0135] After successful combination validation, the normalized numerical feature fields and the encoded categorical feature fields are concatenated as vectors to generate an incident feature set. The concatenation order is fixed and consistent with the field mapping list. The vector structure simultaneously carries parameter versions, dictionary versions, column order hashes, and quality markers, forming a processing trace bound to the sample. The incident feature set is output to two types of storage areas: one is a high-throughput key-value store for online processing, using incident identifiers or sample identifiers as keys and vectors as values, facilitating low-latency retrieval; the other is a columnar store for offline training and evaluation, organized by time and spatial partitions, supporting batch scanning and statistical aggregation. The generation process provides an idempotent interface and duplicate detection to avoid duplicate samples caused by replay.

[0136] This embodiment identifies the original accident feature fields by type, normalizes the maximum and minimum values ​​at the field granularity, and encodes them into integers. After double verification of dimensional consistency and content thresholds, the data is then concatenated into an accident feature set in a fixed order, achieving uniformity in the three dimensions of scale, enumeration, and structure. The normalization parameters and encoding dictionary are managed using versioning, and the same parameter snapshot is used in both training and online stages to reduce similarity mismatch caused by offsets on both sides. Column order hashing and quality tags are embedded with the vector, allowing subsequent dimensionality reduction and similarity retrieval to directly consume the standardized output without repeated cleaning and alignment. This reduces computational overhead, stabilizes latency, and improves the comparability between samples, providing a stable and traceable input for subsequent principal component dimensionality reduction and similarity index construction.

[0137] In one embodiment, step S30 above includes:

[0138] S301, Determine the covariance matrix of the accident feature set;

[0139] S302, decompose the eigenvalues ​​and eigenvectors of the covariance matrix;

[0140] S303, Arrange the eigenvectors in descending order according to the magnitude of the eigenvalues ​​to generate an ordered principal component sequence;

[0141] S304, Based on the ordered principal component sequence, select continuous principal components whose cumulative variance contribution rate exceeds a preset contribution rate threshold, and generate a principal component subset;

[0142] S305, the accident feature set is projected onto the low-dimensional feature space defined by the principal component subset to obtain the projected feature representation as the dimensionality-reduced feature set.

[0143] In this embodiment, before the accident feature set enters the dimensionality reduction stage, mean alignment and scaling are performed to ensure that each column enters the statistical calculation with a uniform metric. Mean alignment is completed by subtracting the historical mean of each column within the sample batch; the historical mean is read from the parameter snapshot. Scaling checks for any mixed-in non-numerical entries or missing bits; if any are found, the process immediately reverts to the quality marker in the standardization stage for removal or supplementation, ensuring the feasibility and stability of covariance calculation. When the sample size is large, a block accumulation strategy is used to maintain the overall mean and second moment, avoiding memory overflow caused by loading all at once and providing a sufficient data foundation for the stable solution of the subsequent covariance matrix.

[0144] A numerically stable second-order statistical accumulation process is used to determine the covariance matrix of the accident feature set. Vector batches are read in blocks, and the global mean, global sum of squares, and cumulative cross terms are updated batch by batch, ultimately calculating a symmetric second-order covariance matrix. To suppress the influence of extreme samples, an anomaly percentage threshold is introduced for each batch; batches exceeding the threshold only participate in mean calculation and not in cumulative cross terms. A minimal diagonal regularization term is applied to the diagonal elements to increase the condition number of subsequent decompositions. For scenarios where the matrix dimension is much larger than the number of samples, an equivalent matrix form is automatically switched to reduce computational complexity, and intermediate products are cached for auditing and reproducibility.

[0145] When decomposing the eigenvalues ​​and eigenvectors of the covariance matrix, a solution strategy is selected based on the data size and hardware capabilities. For small to medium dimensions, exact decomposition is used to ensure vector orthogonality and eigenvalue non-negativity; for large dimensions, random approximation or truncated singular value decomposition is used to shorten computation time while preserving variance information. The convergence index and upper bound of the residuals are recorded during the decomposition process. If the preset accuracy is not reached, the approximate rank is automatically increased or the step size is reduced for repeated iterations. The obtained eigenvectors are uniformly normalized and a one-to-one mapping is established with the corresponding eigenvalues. This mapping relationship is written into the parameter snapshot to ensure that the training and online ends share the same projection basis.

[0146] When sorting eigenvectors in descending order based on their eigenvalues, the eigenvalues ​​are used as the sorting keys, and the eigenvectors are used as the sorting objects, strictly maintaining a one-to-one correspondence. In high-dimensional scenarios, the sorting engine employs parallel key-value pair sorting to avoid memory fragmentation. The sorted output is an ordered principal component sequence, containing column indices, vector content, and corresponding eigenvalues. The sequence is bound to a version number and timestamp during generation and archived in an immutable form. Subsequent filtering uses this sequence as the sole input source to prevent drift caused by repeated calculations.

[0147] When selecting continuous principal components whose cumulative variance contribution rate exceeds a preset contribution rate threshold based on an ordered principal component sequence, the prefix sum of the normalized eigenvalues ​​is first calculated and accumulated along the sequence from front to back until the threshold is reached. A set of vectors consecutively starting from the first term is then selected to form a principal component subset. The threshold is determined from a configuration center and can be adjusted within a controlled range according to the application's resources and real-time requirements. When the smallest index reaching the threshold experiences fluctuation, a time smoothing strategy is used to delay the switch within a small range to ensure the stability of the online vector dimension. The principal component subset output includes the subset size, vector stack, and cumulative contribution rate value, along with a consistent hash for online verification.

[0148] When projecting the accident feature set onto the low-dimensional feature space defined by the principal component subsets, matrix multiplication is used to achieve linear mapping. The computation graph adaptively partitions the input batch size at runtime, prioritizing single-precision or mixed-precision methods to improve throughput, and enabling streaming multiplication in memory-constrained edge environments to reduce peak usage. Before projection, the column order hash and parameter snapshot version are checked again; if inconsistent, projection is rejected and an alarm is reported. After projection, a dimensionality-reduced feature set is generated, along with accompanying metadata, including the version of the principal component subset used, cumulative contribution rate, input batch identifier, and discarded sample count, ensuring that subsequent similarity index training and real-time mapping can directly consume this result without repeated verification.

[0149] This embodiment ensures numerical stability through block accumulation and diagonal regularization of the covariance matrix. Feature decomposition adaptively selects precise or approximate solution strategies based on data scale. The sorting process outputs an ordered principal component sequence based on eigenvalues. Principal component selection employs continuous truncation of cumulative variance contribution rates combined with smooth switching. Linear projection utilizes streaming and mixed-precision computation in resource-constrained environments, ultimately forming a traceable, version-consistent, and dimensionally stable dimensionality-reduced feature set. This reduces high-dimensional redundancy, suppresses noise amplification, lowers the storage and retrieval burden of subsequent similarity indexes, and maintains a consistent projection basis at both the training and online ends, reducing the drift risk and latency fluctuations in similarity calculations.

[0150] In one embodiment, step S40 above includes:

[0151] S401, set the hash length parameter and hash table number parameter of the locality-sensitive hash model;

[0152] S402, Generate a set of random hyperplanes as an initial set of hash functions based on the hash length parameter;

[0153] S403, use the dimensionality reduction feature set to train the projection parameters of the initial hash function set, and generate the trained hash function set;

[0154] S404, construct a hash table storage structure based on the set of trained hash functions, and map each feature vector in the dimensionality reduction feature set to the corresponding hash bucket in the hash table storage structure;

[0155] S405, Use the hash table storage structure to verify the accuracy of the approximate nearest neighbor search;

[0156] S406, When the accuracy of the approximate nearest neighbor search is lower than the preset model performance threshold, adjust the hash length parameter or hash table quantity parameter and retrain;

[0157] S407, when the accuracy of the approximate nearest neighbor search is higher than or equal to the preset model performance threshold, freeze the projection parameters of the trained hash function set, and output the trained hash function set and hash table storage structure as a similarity index structure.

[0158] In this embodiment, the training input is a dimensionality-reduced feature set, with the vector dimension consistent with the output of the previous step. First, the hash length parameter and hash table quantity parameter of the Locality Sensitive Hash (LSH) model are determined, as they jointly determine the signature bit length and parallel table size. When determining the parameters, the vector dimension, total sample size, target retrieval latency, and target recall metric are read to establish a constraint relationship between the target signature collision probability and the candidate set size. Initial values ​​are given through offline evaluation curves or empirical mapping tables, and feasibility is verified within the limits of memory and processor resources. After the parameters are determined, they are fixed to a configuration snapshot, recording the version number and timestamp, providing a basis for subsequent reproduction and rollback.

[0159] The generation of the set of random hyperplanes as the initial set of hash functions follows a one-to-one quantitative relationship with the hash length parameter. Each hyperplane is represented by a normal vector with the same dimension as the dimensionality reduction feature set. The normal vector components are sampled from a symmetric distribution and uniformly normalized to avoid directional bias. To ensure the independence of signatures between different tables, a separate set of normal vectors is sampled for each table, and an orthogonality approximation test is performed after sampling. If the angle between adjacent normal vectors is too small, they are resampled and replaced to reduce redundant direction discrimination. The generation process incorporates metadata about the random seed and sampling strategy for online diagnostics.

[0160] The projection parameter training takes the dimensionality-reduced feature set as input, calculates the symbol for each vector and the initial hash function set to form a signature, and measures the signature collision distribution and the separation of nearest neighbor pairs. If local cluster samples are found to be highly overlapping in the signature space but separable in the vector space, a slight rotation or offset correction is made to the corresponding hyperplane normal vector; if some hyperplanes have almost no separating effect on the entire sample, they are replaced with new random directions until the separation and balance reach the set thresholds. The training output forms the post-trained hash function set, and is bound to a snapshot version and statistical summary, including indicators such as signature entropy, bit balance, and collision moment distribution, for subsequent consistency verification.

[0161] The hash table storage structure is constructed based on the trained hash function set. For each table, a bucket address space and bucket metadata area are first established, with a one-to-one correspondence between bucket identifiers and signature values. Entries within a bucket are stored as compact arrays containing incident record indexes and vector digest fingerprints. Subsequently, an in-table signature is generated for each feature vector in the reduced-dimensional feature set, mapped to the corresponding hash bucket, and written to the incident record index and vector digest fingerprint. When the bucket capacity is about to reach a threshold, a bucket expansion strategy is triggered, reallocating entries within the bucket to the expanded address area to maintain stable latency for insertion and retrieval. After all vector insertions are completed, a full consistency scan is performed to verify signatures and bucket identifiers, deduplicated indexes, and digest fingerprint checksums, ensuring the storage structure can be directly consumed by the query end.

[0162] Nearest neighbor search accuracy verification uses an independent validation set or reserved samples. For each validation vector, signature matching and bucket-level retrieval are first performed within the hash table storage structure to obtain a candidate index set. Then, the precise similarity with the validation vector is calculated in the original vector space, and the proportion of true nearest neighbors covered by the candidate set is statistically analyzed to obtain recall and localization accuracy metrics. To determine stability, verification covers multiple load intervals and different collision density intervals, and accuracy curves and variance intervals are output. If the nearest neighbor search accuracy is lower than the preset model performance threshold, the hash length parameter or hash table number parameter is adjusted in order of priority: increasing the hash length reduces the bucket size and decreases false recalls, while increasing the hash table number improves the hit rate under multi-view projection. After parameter changes, hyperplane generation, projection parameter training, index construction, and verification are re-executed until the metrics reach the threshold or the resource limit triggers the termination strategy.

[0163] When the accuracy of the near nearest neighbor search is higher than or equal to a preset model performance threshold, the freezing process begins. The freezing action marks the normal vectors and offsets of the trained hash function set as read-only, outputting them along with the hash table storage structure as a similarity index structure. Simultaneously, a complete version list is generated, including parameter values, function set summary hashes, bucket space layout, validation statistical summaries, and data effective time windows. The similarity index structure provides read-only query interfaces and controlled append update interfaces to ensure stability and consistency during online retrieval. To adapt to edge environments, the function set and bucket metadata can be serialized into a compact format and loaded into a high-performance memory area. Bucket entries are placed in local persistent media in the form of paginated blocks. During queries, the in-memory index is hit first, followed by a lookup of the source entry block, maintaining retrieval latency within a set range.

[0164] This embodiment establishes a closed-loop iteration between parameter determination, hyperplane generation, projection correction, index construction, and accuracy verification. The output similar index structure achieves a higher near-nearest neighbor hit rate and a lower false recall probability under the same resource constraints. After training, the hash function set is frozen and versioned, ensuring consistency between online and offline query behavior and reducing retrieval drift. Bucket space balancing and expansion strategies reduce hotspot conflicts and long-tail query latency. Serialized layout and memory-first loading methods reduce edge device usage and cold start time. Without increasing the perceptual computing burden, the candidate set size is compressed, retrieval latency is stabilized, and similar record coverage is improved, providing a traceable, maintainable, and scalable index foundation for downstream similarity comparison and early warning determination.

[0165] In one embodiment, step S50 above includes:

[0166] S501 allows parallel access to driver behavior sensor data streams, vehicle status sensor data streams, and environmental information sensor data streams at the edge processing device.

[0167] S502, Perform behavioral feature parsing on the driver behavior sensor data stream to extract the steering wheel angle feature field and the brake pedal travel feature field;

[0168] S503, Perform state feature parsing on the vehicle state sensor data stream to extract vehicle speed feature field and engine speed feature field;

[0169] S504, Perform environmental feature parsing on the environmental information sensor data stream to extract the light intensity feature field and the road slipperiness feature field;

[0170] S505, based on the feature dimension definition of the vehicle accident database, the steering wheel angle feature field, brake pedal travel feature field, vehicle speed feature field, engine speed feature field, light intensity feature field and road slipperiness feature field are dimensionally aligned;

[0171] S506 performs real-time format conversion on the dimension-aligned feature fields in the edge processing device's memory buffer to generate real-time feature vectors.

[0172] In this embodiment, multiple input channels for real-time sensor data streams are established on the edge processing device. Acquisition interfaces, caching strategies, and time synchronization strategies are configured for the driver behavior sensor data stream, vehicle status sensor data stream, and environmental information sensor data stream, respectively. The multiple input channels are aligned with a unified time base, and a combination of arrival timestamps and device clock correction is used to eliminate cross-channel jitter. To avoid input congestion, each input channel has its own receive buffer and packet loss compensation strategy, and channel identifiers and frame identifiers are recorded at the head of the queue to ensure accurate location of the data according to its source and time window during subsequent parsing.

[0173] Behavioral feature parsing performs a field extraction process based on the driver behavior sensor data stream. The steering wheel angle feature field is obtained by mapping the original sequence from the angle encoder or angle estimation model. During parsing, de-jittering and short-window smoothing are performed first, then the dimensions, positive and negative directions, and zero-point reference are determined according to the feature dimensions defined in the vehicle accident database. The output is a standardized representation of the steering wheel angle feature field. The brake pedal travel feature field is obtained from the pedal travel displacement or pedal force proxy. The parsing process is consistent with that of the steering wheel angle feature field, and the acquisition time range and quality label are appended at the end of the parsing to ensure that subsequent dimension alignment can be cross-validated over time. Behavioral feature parsing is executed as a separate unit on the edge processing device, using fixed input buffers and read-only output buffers to avoid cross-thread contention.

[0174] State feature parsing performs unified extraction on the vehicle state sensor data stream. Vehicle speed feature fields are selected from consistent sources based on wheel speed or overall vehicle speed estimates, and peak removal and frame loss interpolation are performed. Engine speed feature fields are extracted from powertrain speed sources and averaged using a short window to eliminate transient fluctuations caused by cogging effects. During the parsing phase, both types of fields are mapped to a unified unit and scale based on the feature dimensions defined in the vehicle accident database, and source identifiers and acquisition channel numbers are appended to support tracking and backtracking. State feature parsing and behavioral feature parsing share a time synchronization service, aligning vehicle speed feature fields and engine speed feature fields to the same sampling cross-section through a unified clock snapshot.

[0175] Environmental feature parsing performs field extraction and confidence estimation on environmental information sensor data streams. The illumination intensity feature field is obtained from ambient light sensor data or image brightness estimation sequences. It undergoes black level correction and drift cancellation, followed by unit and range mapping based on the feature dimension definitions of the vehicle accident database. The road slippage feature field is provided by ground adhesion estimation, tire slippage detection, or environmental perception fusion results. During parsing, consistency checks and noise thresholding are performed, marking incomplete or conflicting segments as requiring completion. Environmental feature parsing outputs confidence scores and missing data markers, facilitating selective completion or skipping during the dimension alignment stage.

[0176] Dimension alignment uses the feature dimensions of the vehicle accident database as the target architecture, and completes the structural splicing of the three parsed outputs of behavior, state, and environment according to field name, field order, and field unit. The alignment process first checks the timestamp coverage of the steering wheel angle feature field, brake pedal travel feature field, vehicle speed feature field, engine speed feature field, light intensity feature field, and road slippage feature field, constructing the same sampling section using a minimum containment window. For missing entries, nearest neighbor interpolation or model interpolation is used, and a missing mask is retained in the alignment matrix to ensure that subsequent real-time format conversion can identify and process them. After dimension alignment, an intermediate structure with complete fields, consistent units, and a fixed order is formed, while retaining the field source, credibility, and processing traces for easy auditing and playback.

[0177] Real-time format conversion is performed within the edge processing device's memory buffer, transforming the aligned field structure into a real-time feature vector. The conversion process constructs a fixed-length vector using a contiguous memory layout, filling in the steering wheel angle, brake pedal travel, vehicle speed, engine speed, light intensity, and road slippage feature fields according to the feature dimensions defined in the vehicle accident database. Unfilled slots are filled using a missing mask strategy. To meet low latency requirements, the memory buffer employs a circular queue and zero-copy dequeueing, directly exposing the real-time feature vector to downstream processes and avoiding secondary serialization. After real-time format conversion, a timestamp snapshot and vector checksum are recorded to ensure consistency with previous parsed records and reproducibility during fault diagnosis.

[0178] In real-time vehicle accident warning systems, edge devices are not necessarily single hardware units, but rather processing systems composed of multiple functional modules. First, the onboard computing unit undertakes the main local computing tasks. Examples include embedded industrial computers, single-board computers, or ECU expansion boards within the vehicle. These units can operate stably in the vehicle environment and complete sensor data parsing and preprocessing in milliseconds. Second, the sensor interface gateway acts as a bridge between the edge device and different data sources. It collects vehicle status information via the CAN bus or OBD-II interface, integrates environmental data from cameras, radar, light and humidity sensors, and accesses driver operation sensors, thereby ensuring the parallel acquisition of behavioral, status, and environmental data.

[0179] In terms of communication, edge devices typically include 4G, 5G, or V2X communication modules for real-time data exchange between the vehicle and the external environment or remote systems. Simultaneously, data interaction between different sub-modules is ensured within the vehicle via Ethernet or Wi-Fi. To support rapid data flow in real-time processing, edge devices are often equipped with high-performance memory buffers and local solid-state storage to store copies of partial accident databases or perform temporary caching, thereby avoiding latency caused by frequent external transmissions. Furthermore, to ensure data security and device stability, edge devices integrate hardware encryption units or trusted execution environments for encrypted processing of sensitive data. These, along with device self-monitoring and fault diagnosis modules, ensure continuous operation even in environments with extreme temperatures, vibrations, or network instability.

[0180] Overall, edge devices encompass computing, interfaces, communication, storage, and security, with each component working closely together to enable the local collection, parsing, fusion, and real-time feature vector generation of driver behavior data streams, vehicle status data streams, and environmental information data streams. This design reduces reliance on remote computing resources and significantly shortens accident warning response time, laying an efficient and reliable foundation for subsequent similarity comparisons and anomaly warning outputs.

[0181] This embodiment achieves stable output of real-time feature vectors with complete fields, consistent units, and fixed order by using time synchronization and congestion control of multiple input channels, separate parsing of driver behavior sensor data streams, vehicle status sensor data streams, and environmental information sensor data streams, dimension alignment based on the feature dimensions of the vehicle accident database, and real-time format conversion within the memory buffer. Steering wheel angle feature fields, brake pedal travel feature fields, vehicle speed feature fields, engine speed feature fields, light intensity feature fields, and road slippage feature fields are aligned within a unified time segment, reducing error propagation caused by cross-source offsets. Zero-copy dequeueing and contiguous memory layout shorten the generation path and reduce resource consumption, providing a low-latency, traceable, and easily reproducible data entry point for subsequent mapping and retrieval processes.

[0182] In one embodiment, step S60 above includes:

[0183] S601, Perform locality-sensitive hashing on the real-time feature vector to generate the target hash signature;

[0184] S602, in the hash table storage structure with the similar index structure, find the hash bucket that matches the target hash signature;

[0185] S603, Extract the set of incident record indexes from the matching hash buckets;

[0186] S604, retrieve the corresponding accident record from the vehicle accident database according to the accident record index set;

[0187] S605, aggregate all retrieved accident records to generate an initial candidate accident record set;

[0188] S606, perform time window filtering on the initial candidate accident record set and retain accident records within a preset time range;

[0189] S607 outputs the filtered incident records as the final set of candidate incident records.

[0190] In this embodiment, a candidate accident record set is generated by using real-time feature vectors as input and a similarity index structure as the retrieval carrier through locality-sensitive hashing (LSH) and index retrieval. The real-time feature vectors are derived from the parallel parsing and format conversion output of the edge processing device, and their dimensions and scale are consistent with the feature dimensions in the vehicle accident database. The similarity index structure is derived from the output of the similarity matching model during the training phase, and is composed of a set of trained hash functions and a hash table storage structure. The hash table storage structure internally organizes the accident record index entries using hash buckets.

[0191] First, locality-sensitive hashing (LSH) is performed to generate the target hash signature. To maintain consistency with the training phase, the real-time feature vectors undergo the same centering and dimension alignment as in the training phase before entering the signature stage, with missing components padded according to the default strategy used in the training phase. The target hash signature is obtained by applying each of the trained hash functions sequentially. Each function performs a projection and sign determination on the vector, and multiple determination results are concatenated in a predetermined order to form the signature. The signature bit width and the number of hash tables are consistent with those in the training phase to avoid cross-structure offsets during retrieval. To improve signature stability, a narrowband jitter suppression strategy can be introduced near the threshold to reduce the impact of boundary effects on bucket routing.

[0192] Next, a hash bucket matching the target hash signature is searched in a hash table storage structure with a similar index structure. Each hash table is located to a hash bucket using the target hash signature as the key. If the located bucket is empty, a signature proximity expansion search is initiated, expanding layer by layer in order of increasing signature bit difference. Adjacent buckets with a difference of one bit from the target hash signature are visited first, and then the search is expanded layer by layer to avoid the accumulation of latency caused by large-scale traversal. The expansion threshold and expansion order reuse the performance parameters from the training phase to ensure that the search path is consistent with the offline evaluation.

[0193] The system extracts a set of incident record indexes from the matching hash buckets. Each entry in a hash bucket uses the incident identifier as the primary key, along with the database partition marker and record offset, to avoid full table scans during secondary searches. Index sets returned from multiple hash tables are merged and deduplicated; identical incident identifiers are retained only once. The system also records the hit count and the table number of the hit, used for subsequent aggregation and sorting. To control the size of a single search, when the index set size exceeds a threshold, it is truncated after being weighted by hit count and table number, retaining only the most representative entries for the next stage.

[0194] Retrieves corresponding accident records from the vehicle accident database based on the accident record index set. The retrieval process employs a partitioned parallel approach, tagging the aggregated index by partition before triggering batch reads to reduce random I / O. Read fields are limited to accident feature vectors, accident timestamps, accident type tags, and accident identifiers to avoid loading irrelevant information that could consume bandwidth. The read order remains consistent with the index set's sorting, facilitating the direct inheritance of weight information during the aggregation phase.

[0195] All retrieved incident records are aggregated to generate an initial candidate incident record set. Records from different hash tables are aggregated based on hit count, table position, and signature consistency to construct an aggregation score. A single aggregation entry is formed for the same incident identifier, accompanied by an incident feature vector, timestamp, and incident type label. After the aggregation process is complete, two statistics are provided: deduplication ratio and aggregation depth, used for operational monitoring and subsequent threshold maintenance.

[0196] The initial candidate incident record set is filtered using a time window to retain incident records within a preset time range. Timestamp fields are uniformly converted to a unified time zone before filtering according to the preset time range. Records with missing timestamps or abnormal timestamp formats are placed in an exception set and not included in this round of output. The time window boundaries are consistent with the business strategy and can be executed based on natural day boundaries or a sliding window. When the time window overlaps with the database partition boundaries, alignment with the partition boundaries is prioritized to reduce the cost of cross-partition access.

[0197] The filtered incident records are output as the final candidate incident record set. The output structure includes an incident identifier, incident feature vector, incident timestamp, incident type label, and aggregate score. The field order is consistent with the input protocol of the downstream similarity comparison unit. To facilitate subsequent traceability, each batch of output includes a retrieval batch number, target hash signature, and extended level information used, forming an auditable retrieval chain. To suppress set expansion in extreme cases, an upper limit can be set for the set size in the output stage. When the upper limit is exceeded, the set is truncated in descending order of aggregate score, and the truncation event is recorded.

[0198] Example Explanation: In the fintech field, especially in auto insurance risk management and claims early warning scenarios, a complete accident similarity comparison and early warning process manifests as a continuous chain spanning historical data construction, feature engineering, model training, and real-time early warning output. First, multi-source accident data needs to be collected, including status data generated by the vehicle diagnostic system, collision sequences collected by the event recorder, environmental attributes in the accident liability determination report, road topology in on-site inspection photos, user profiles in claims records, and accident scene descriptions in the reported audio and text. These multi-source data are linked through accident identifiers to form a unified accident record. Data cleaning is performed on these records to eliminate missing and outlier values, and the data is converted into a unified data format before being stored in a database, constructing a vehicle accident database as the foundation for subsequent processing.

[0199] After extracting the original accident feature fields from the database, standardization processing is performed on different types of data. Numerical feature fields are normalized using maximum and minimum values, mapping all data to a unified range; categorical feature fields are converted into ordered numerical representations using integer encoding. Subsequently, the consistency of feature dimensions is checked and maintained, and the processed features are merged to generate an accident feature set. This set contains multi-dimensional features of vehicle operating status, driving behavior, environmental factors, and accident semantic information, possessing a unified structure that facilitates subsequent modeling.

[0200] To reduce the redundancy and computational overhead associated with high-dimensional features, principal component analysis (PCA) is performed on the accident feature set to generate a dimensionality-reduced feature set. Specifically, the feature covariance matrix is ​​calculated, and its eigenvalues ​​and eigenvectors are decomposed. Principal component subsets are selected based on their variance contribution rates, and the original features are projected into a low-dimensional space. This step significantly reduces feature dimensionality while preserving information content, improving the efficiency of model training and real-time computation.

[0201] A similarity matching model is trained based on a dimensionality-reduced feature set, and a locality-sensitive hashing (LSH) method is used to generate a similarity index structure. During training, the hash length and the number of hash tables are set, a random hyperplane is generated as the initial hash function, and the projection parameters are optimized on the dimensionality-reduced feature set to form the trained hash function set. These functions are used to construct hash tables, mapping feature vectors to hash buckets. The model performance is evaluated by retrieval accuracy; if the target is not met, the parameters are dynamically adjusted until a threshold is reached, ultimately forming a stable similarity index structure. This structure supports approximate nearest neighbor search and can efficiently retrieve potentially similar accident records.

[0202] In real-time operation, the edge processing device receives data streams from multiple sensors, including driver behavior sensors, vehicle status sensors, and environmental information sensors. Feature parsing is performed on each of these data streams to extract feature fields such as steering wheel angle, brake pedal travel, vehicle speed, engine speed, light intensity, and road slippage. Alignment is then performed according to the feature dimensions defined in the database, and real-time format conversion is completed within the edge memory buffer to generate real-time feature vectors. This ensures that the real-time collected data has a consistent structure with the features in the historical database, enabling smooth entry into the similarity retrieval process.

[0203] When mapping real-time feature vectors to a similar index structure, locality-sensitive hashing (LSH) is first performed to generate the target hash signature. Then, matching hash buckets are searched in the hash table, and the incident record index set is extracted. Next, corresponding incident records are retrieved from the database based on these indexes, and aggregated to obtain an initial candidate incident record set. After filtering this set through a time window, incident data within a preset time range is retained, and the output is the final candidate incident record set. This step ensures both the comprehensiveness of the recall and the timeliness of the results.

[0204] Finally, a similarity comparison is performed between the candidate accident record set and the real-time feature vector. Similarity indices are calculated for each accident record and the real-time vector using methods such as cosine similarity. Records exceeding a threshold are filtered out, and the frequency distribution of their type labels is further analyzed to determine the most frequent accident types. Based on this result, the system generates anomaly warning information and outputs it in real-time through the in-vehicle display terminal. In fintech-enabled auto insurance scenarios, this process can quickly identify potential high-risk situations during vehicle operation, issue early warnings, and reduce accident rates and claims costs.

[0205] This embodiment reduces the error bucket rate by reusing the processing consistency of the training phase before the real-time feature vector enters the similarity index structure, ensuring that the target hash signature and the offline index remain isomorphic. It avoids retrieval failures caused by empty buckets and suppresses the amplification of duplicate records by employing signature proximity expansion and multi-table merging for deduplication in the hash table storage structure. It shortens the database access path and bandwidth usage by using partitioned parallel batch retrieval driven by the incident record index set and limiting the fields read. Time window filtering removes records that are historically distant and contribute little to the current context, resulting in an output set that balances scale and timeliness. Approximate nearest neighbor recall is achieved without introducing high-overhead model inference, forming a stable and traceable candidate incident record set. This provides a low-latency, low-resource-consumption, and easily maintainable input set for subsequent similarity comparison and threshold judgment.

[0206] In one embodiment, a warning device based on multi-source accident data is provided, which corresponds one-to-one with the warning method based on multi-source accident data in the above embodiments. (Refer to...)Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the early warning device based on multi-source accident data of the present invention. The modules include an accident database construction module 10, a feature standardization module 20, a feature dimensionality reduction module 30, a similarity matching model training module 40, a real-time feature vector generation module 50, a candidate accident record retrieval module 60, and a similarity comparison and early warning module 70. Detailed descriptions of each functional module are as follows:

[0207] Accident database construction module 10 is used to acquire multi-source accident data and generate a vehicle accident database based on the multi-source accident data;

[0208] Feature standardization module 20 is used to perform feature standardization processing on the accident features in the vehicle accident database to obtain an accident feature set;

[0209] Feature dimensionality reduction module 30 is used to perform principal component dimensionality reduction processing on the accident feature set to obtain a dimensionality-reduced feature set;

[0210] The similarity matching model training module 40 is used to train a similarity matching model based on the dimensionality reduction feature set and generate a similarity index structure.

[0211] The real-time feature vector generation module 50 is used to acquire real-time sensor data stream, parse the feature fields in the real-time sensor data stream, and convert the feature fields into real-time feature vectors.

[0212] The candidate accident record retrieval module 60 is used to map the real-time feature vector to the similarity index structure to obtain a candidate accident record set.

[0213] The similarity comparison and early warning module 70 is used to perform similarity comparison between the candidate accident record set and the real-time feature vector, generate a similarity index, and output abnormal early warning information when the similarity index exceeds a preset threshold.

[0214] In one embodiment, the accident database construction module 10 is specifically used for:

[0215] Vehicle status data is obtained from the vehicle diagnostic system, collision event sequence data is obtained from the event data recorder, accident environmental attribute data is obtained from the accident liability determination letter, road topology data is extracted from on-site inspection photos, user profile data is obtained from claims records, and accident scene description data is parsed from the voice text of the accident report.

[0216] By associating the vehicle status data, collision event sequence data, accident environment attribute data, road topology data, user profile data, and accident scene description data with the accident identifier, associated accident records are generated.

[0217] Perform data cleaning operations on the associated incident records to handle missing and outlier values ​​and generate cleaned incident records;

[0218] The cleaned accident records are converted into a structured data format;

[0219] The cleaned accident records in the structured data format are stored according to a preset data mode to generate a vehicle accident database.

[0220] In one embodiment, the feature standardization module 20 is specifically used for:

[0221] Extract the original accident feature fields from the vehicle accident database;

[0222] Identify the data type of the original accident feature fields to distinguish between numerical feature fields and categorical feature fields;

[0223] Perform maximum and minimum value normalization on the numerical feature fields to map the feature values ​​to a preset normalization interval, and obtain the normalized numerical feature fields.

[0224] Integer encoding is performed on the categorical feature fields to convert discrete feature values ​​into ordered numerical representations, resulting in encoded categorical feature fields.

[0225] Verify the dimensional consistency between the normalized numerical feature fields and the encoded categorical feature fields;

[0226] By combining the normalized numerical feature fields that have passed the combined verification with the coded categorical feature fields, an accident feature set is generated.

[0227] In one embodiment, the feature dimensionality reduction module 30 is specifically used for:

[0228] Determine the covariance matrix of the accident feature set;

[0229] Decompose the eigenvalues ​​and eigenvectors of the covariance matrix;

[0230] The eigenvectors are sorted in descending order according to their eigenvalues ​​to generate an ordered principal component sequence.

[0231] Based on the ordered principal component sequence, continuous principal components whose cumulative variance contribution rate exceeds a preset contribution rate threshold are selected to generate a principal component subset.

[0232] The accident feature set is projected onto the low-dimensional feature space defined by the principal component subsets to obtain the projected feature representation as the dimensionality-reduced feature set.

[0233] In one embodiment, the similarity matching model training module 40 is specifically used for:

[0234] Set the hash length parameter and hash table number parameter for the locality-sensitive hash model;

[0235] A set of random hyperplanes is generated based on the hash length parameter as the initial set of hash functions;

[0236] The projection parameters of the initial hash function set are trained using the dimensionality reduction feature set to generate the trained hash function set;

[0237] A hash table storage structure is constructed based on the set of trained hash functions, and each feature vector in the dimensionality reduction feature set is mapped to the corresponding hash bucket in the hash table storage structure;

[0238] The accuracy of the approximate nearest neighbor search was verified using the hash table storage structure.

[0239] When the accuracy of the approximate nearest neighbor search is lower than the preset model performance threshold, adjust the hash length parameter or the hash table number parameter and retrain.

[0240] When the accuracy of the approximate nearest neighbor search is higher than or equal to the preset model performance threshold, the projection parameters of the trained hash function set are frozen, and the trained hash function set and hash table storage structure are output as a similarity index structure.

[0241] In one embodiment, the real-time feature vector generation module 50 is specifically used for:

[0242] The driver behavior sensor data stream, vehicle status sensor data stream, and environmental information sensor data stream are accessed in parallel at the edge processing device;

[0243] Perform behavioral feature parsing on the driver behavior sensor data stream to extract steering wheel angle feature fields and brake pedal travel feature fields;

[0244] Perform state feature parsing on the vehicle state sensor data stream to extract vehicle speed feature fields and engine speed feature fields;

[0245] Environmental feature parsing is performed on the environmental information sensor data stream to extract the light intensity feature field and the road slipperiness feature field;

[0246] Based on the feature dimension definition of the vehicle accident database, the steering wheel angle feature field, brake pedal travel feature field, vehicle speed feature field, engine speed feature field, light intensity feature field, and road slipperiness feature field are dimensionally aligned;

[0247] Real-time format conversion is performed on the dimension-aligned feature fields in the memory buffer of the edge processing device to generate real-time feature vectors.

[0248] In one embodiment, the candidate accident record retrieval module 60 is specifically used for:

[0249] Locality-sensitive hashing is performed on the real-time feature vector to generate the target hash signature;

[0250] In the hash table storage structure with the similar index structure, search for the hash bucket that matches the target hash signature;

[0251] Extract the incident record index set from the matching hash bucket;

[0252] Based on the accident record index set, the corresponding accident record is retrieved from the vehicle accident database;

[0253] Aggregate all retrieved accident records to generate an initial candidate accident record set;

[0254] The initial candidate accident record set is filtered by a time window to retain accident records within a preset time range;

[0255] The filtered incident records are output as the final set of candidate incident records.

[0256] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side early warning method based on multi-source incident data.

[0257] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of a warning method based on multi-source accident data.

[0258] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0259] Acquire multi-source accident data and generate a vehicle accident database based on the multi-source accident data;

[0260] Perform feature standardization processing on the accident features in the vehicle accident database to obtain an accident feature set;

[0261] Principal component dimensionality reduction is performed on the accident feature set to obtain a dimensionality-reduced feature set;

[0262] A similarity matching model is trained based on the reduced-dimensional feature set to generate a similarity index structure;

[0263] Acquire real-time sensor data stream, parse the feature fields in the real-time sensor data stream, and convert the feature fields into real-time feature vectors;

[0264] The real-time feature vectors are mapped to the similarity index structure to obtain a set of candidate accident records;

[0265] The candidate accident record set is compared with the real-time feature vector to generate a similarity index. When the similarity index exceeds a preset threshold, an abnormal warning message is output.

[0266] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0267] Acquire multi-source accident data and generate a vehicle accident database based on the multi-source accident data;

[0268] Perform feature standardization processing on the accident features in the vehicle accident database to obtain an accident feature set;

[0269] Principal component dimensionality reduction is performed on the accident feature set to obtain a dimensionality-reduced feature set;

[0270] A similarity matching model is trained based on the reduced-dimensional feature set to generate a similarity index structure;

[0271] Acquire real-time sensor data stream, parse the feature fields in the real-time sensor data stream, and convert the feature fields into real-time feature vectors;

[0272] The real-time feature vectors are mapped to the similarity index structure to obtain a set of candidate accident records;

[0273] The candidate accident record set is compared with the real-time feature vector to generate a similarity index. When the similarity index exceeds a preset threshold, an abnormal warning message is output.

[0274] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0275] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0276] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0277] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A pre-warning method based on multi-source accident data, characterized in that, The method comprises the following steps: Obtaining multi-source accident data, generating a vehicle accident database based on the multi-source accident data; Performing feature standardization processing on the accident features in the vehicle accident database to obtain an accident feature set; Performing principal component dimension reduction processing on the accident feature set to obtain a reduced dimension feature set; Training a similarity matching model based on the reduced dimension feature set to generate a similarity index structure; Obtaining a real-time sensor data stream, parsing the feature fields in the real-time sensor data stream, and converting the feature fields into a real-time feature vector; Mapping the real-time feature vector to the similarity index structure to obtain a candidate accident record set; Performing similarity comparison on the candidate accident record set and the real-time feature vector to generate a similarity index, and outputting an abnormal warning information when the similarity index exceeds a preset threshold.

2. The early warning method based on multi-source accident data according to claim 1, wherein, Obtaining multi-source accident data, generating a vehicle accident database based on the multi-source accident data, comprising: Obtaining vehicle state data from a vehicle diagnostic system, obtaining collision event sequence data from an event data recorder, obtaining accident environment attribute data from an accident liability determination book, extracting road topology data from scene investigation photos, obtaining user portrait data from claim settlement records, and parsing accident scene description data from complaint voice text; Generating associated accident records by associating the vehicle state data, the collision event sequence data, the accident environment attribute data, the road topology data, the user portrait data and the accident scene description data through an accident identifier; Performing data cleaning operation on the associated accident records to process missing values and abnormal values, and generating cleaned accident records; Converting the cleaned accident records into a structured data format; Storing the cleaned accident records in the structured data format according to a preset data mode to generate a vehicle accident database.

3. The early warning method based on multi-source accident data of claim 1, wherein, Performing feature standardization processing on the accident features in the vehicle accident database to obtain an accident feature set, comprising: Extracting original accident feature fields from the vehicle accident database; Identifying the feature data types of the original accident feature fields to distinguish numerical value type feature fields and category type feature fields; Performing maximum and minimum value normalization processing on the numerical value type feature fields to map the feature values to a preset normalization interval to obtain normalized numerical value type feature fields; Performing integer encoding processing on the category type feature fields to convert discrete feature values into ordered numerical values to obtain encoded category type feature fields; Verifying the dimension consistency of the normalized numerical value type feature fields and the encoded category type feature fields; Combining the normalized numerical value type feature fields and the encoded category type feature fields that pass the verification to generate an accident feature set.

4. The early warning method based on multi-source accident data of claim 1, wherein, Performing principal component dimension reduction processing on the accident feature set to obtain a reduced dimension feature set, comprising: Determining the covariance matrix of the accident feature set; Decomposing the eigenvalues and eigenvectors of the covariance matrix; Arranging the eigenvectors in descending order according to the eigenvalue size to generate an ordered principal component sequence; Based on the ordered principal component sequence, selecting continuous principal components whose cumulative variance contribution rate exceeds a preset contribution rate threshold to generate a principal component subset; Projecting the accident feature set to a low-dimensional feature space defined by the principal component subset to obtain a projected feature representation as a reduced dimension feature set.

5. The early warning method based on multi-source accident data of claim 1, wherein, Training a similarity matching model based on the reduced dimension feature set to generate a similarity index structure, including: Setting a hash length parameter and a hash table number parameter of a local sensitive hash model; Generating a set of random hyperplanes as an initial hash function set according to the hash length parameter; Training projection parameters of the initial hash function set using the reduced dimension feature set to generate a trained hash function set; Based on the trained hash function set, constructing a hash table storage structure, and mapping each feature vector in the reduced dimension feature set to a corresponding hash bucket in the hash table storage structure; Verifying the approximate nearest neighbor search accuracy rate using the hash table storage structure; When the approximate nearest neighbor search accuracy rate is lower than a preset model performance threshold, adjusting the hash length parameter or the hash table number parameter and retraining; When the approximate nearest neighbor search accuracy rate is higher than or equal to the preset model performance threshold, freezing the projection parameters of the trained hash function set, and outputting the trained hash function set and the hash table storage structure as the similarity index structure.

6. The early warning method based on multi-source accident data of claim 1, wherein, Obtaining a real-time sensor data stream, parsing a feature field in the real-time sensor data stream, and converting the feature field into a real-time feature vector, including: Parallelly accessing a driver behavior sensor data stream, a vehicle state sensor data stream, and an environmental information sensor data stream on an edge processing device; Performing behavior feature parsing on the driver behavior sensor data stream to extract a steering wheel angle feature field and a brake pedal stroke feature field; Performing state feature parsing on the vehicle state sensor data stream to extract a vehicle speed feature field and an engine speed feature field; Performing environmental feature parsing on the environmental information sensor data stream to extract an illumination intensity feature field and a road wetness feature field; Aligning the dimensions of the steering wheel angle feature field, the brake pedal stroke feature field, the vehicle speed feature field, the engine speed feature field, the illumination intensity feature field, and the road wetness feature field according to a feature dimension definition of a vehicle accident database; Performing real-time format conversion on the dimension-aligned feature fields in an edge processing device memory buffer to generate a real-time feature vector.

7. The early warning method based on multi-source accident data of claim 1, wherein, Mapping the real-time feature vector to the similarity index structure to obtain a candidate accident record set, including: Performing local sensitive hash processing on the real-time feature vector to generate a target hash signature; In the hash table storage structure of the similarity index structure, finding a hash bucket matching the target hash signature; Extracting an accident record index set from the matched hash bucket; According to the accident record index set, retrieving corresponding accident records from the vehicle accident database; Aggregating all retrieved accident records to generate an initial candidate accident record set; Performing time window filtering on the initial candidate accident record set to retain accident records within a preset time range; Outputting the filtered accident records as the final candidate accident record set.

8. A warning device based on multi-source accident data, characterized in that, The early warning device based on multi-source accident data includes: An accident database construction module is configured to acquire multi-source accident data, and generate a vehicle accident database based on the multi-source accident data; A feature standardization module is configured to perform feature standardization processing on accident features in the vehicle accident database, to obtain an accident feature set; A feature dimension reduction module is configured to perform principal component dimension reduction processing on the accident feature set, to obtain a dimension-reduced feature set; A similarity matching model training module is configured to train a similarity matching model based on the dimension-reduced feature set, to generate a similarity index structure; A real-time feature vector generation module is configured to acquire a real-time sensor data stream, parse feature fields in the real-time sensor data stream, and convert the feature fields into a real-time feature vector; A candidate accident record retrieval module is configured to map the real-time feature vector to the similarity index structure, to obtain a candidate accident record set; A similarity comparison and early warning module is configured to perform similarity comparison on the candidate accident record set and the real-time feature vector, to generate a similarity index, and output an abnormal early warning information when the similarity index exceeds a preset threshold.

9. A computer device, comprising: The computer device includes a memory, a processor, and a multi-source accident data-based early warning program stored on the memory and executable on the processor. When the multi-source accident data-based early warning program is executed by the processor, the steps of the multi-source accident data-based early warning method according to any one of claims 1-7 are implemented.

10. A computer-readable storage medium, characterized in that, The storage medium stores a multi-source accident data-based early warning program. When the multi-source accident data-based early warning program is executed by the processor, the steps of the multi-source accident data-based early warning method according to any one of claims 1-7 are implemented.

Images