Method for recognizing a vehicle collision and vehicle

Abstract

This application provides a vehicle collision recognition method and a vehicle, capable of jointly extracting features from acquired raw acoustic and vibration signals to obtain a topological embedding feature vector; performing temporal causal modeling based on the topological embedding feature vector to obtain causal information; and performing multi-agent collaborative discrimination based on the causal information and the topological embedding feature vector to obtain an anomaly score vector; performing multi-source information fusion decision-making based on the anomaly score vector, causal information, and the topological embedding feature vector to obtain fusion evaluation data; determining a dynamic alarm threshold based on the real-time vehicle status; and determining the collision recognition result based on the dynamic alarm threshold and the fusion evaluation data. By introducing a joint modeling mechanism for acoustic and vibration signals, this method addresses the current problem of insensitivity to low-amplitude undercarriage impacts on vehicles.

Description

Technical Field

[0001] This application relates to the field of vehicle safety identification technology, and in particular to a vehicle collision identification method and vehicle. Background Technology

[0002] Vehicle chassis condition recognition technology primarily relies on inertial measurement units (IMUs), suspension height sensors, or vehicle attitude estimation results to determine the presence of abnormal operating conditions. However, during actual driving, especially when traversing slopes, speed bumps, or undulating roads at low speeds, vehicles may exhibit overall stable body posture and minimal changes in suspension travel, even though chassis structural components have already made contact with the ground. Such collision events are often characterized by short duration, small impact amplitude, and complex directions, making it difficult for inertial sensors such as IMUs to generate significant responses at the attitude or acceleration levels. Summary of the Invention

[0003] In view of this, the purpose of this application is to propose a vehicle collision identification method and a vehicle. To address the problem that chassis collision events are difficult to accurately identify under stable vehicle body posture conditions, this application introduces a joint modeling mechanism of acoustic signals and vibration signals to solve the problem that a single sensing scheme is not sensitive to short-term, low-amplitude undercarriage impacts, thereby achieving reliable and generalized determination of actual contact events of the vehicle chassis.

[0004] To achieve the above objectives, this application provides a method for identifying vehicle collisions, comprising:

[0005] Joint feature extraction is performed on the acquired raw acoustic signal and raw vibration signal to obtain the topological embedding feature vector; Temporal causal modeling is performed based on the topological embedding feature vector to obtain causal information, and multi-agent collaborative discrimination is performed based on the causal information and the topological embedding feature vector to obtain an anomaly score vector. Based on the anomaly scoring vector, the causal information, and the topological embedding feature vector, a multi-source information fusion decision is made to obtain fusion evaluation data; The dynamic alarm threshold is determined based on the real-time vehicle status, and the collision recognition result is determined based on the dynamic alarm threshold and the fused evaluation data.

[0006] Optionally, the step of jointly extracting features from the acquired raw acoustic signal and raw vibration signal to obtain a topological embedding feature vector includes: The original acoustic signal and the original vibration signal are time-aligned according to the vehicle clock to obtain the initial acoustic signal and the initial vibration signal; The continuous initial acoustic signal and initial vibration signal are converted into a discrete acoustic-vibration window data sequence according to a sliding window of preset time length. The acoustic vibration window data sequence is subjected to joint feature extraction and topological feature compression to obtain the topological embedding feature vector.

[0007] Optionally, the step of performing joint feature extraction and topological feature compression on the acoustic vibration window data sequence to obtain the topological embedding feature vector includes: Extract the joint acoustic and vibration features from the acoustic and vibration window data sequence to obtain the high-dimensional acoustic and vibration feature matrix; The high-dimensional acoustic-vibration feature matrix is ​​subjected to neighborhood-preserving dimensionality reduction mapping according to the preset target dimension to obtain the topological embedding feature vector.

[0008] Optionally, the causal information includes a causal weight matrix and a time-sensitive vector; the step of performing time-series causal modeling based on the topological embedding feature vector to obtain the causal information includes: The degree of mutual influence between each dimension in the topological embedding feature vector is quantified to obtain the causal weight matrix; The causal temporal pattern between acoustic and vibration features in the topological embedding feature vector is determined to obtain the time sensitivity vector.

[0009] Optionally, the step of performing multi-agent collaborative discrimination based on the causal information and the topological embedding feature vector to obtain an anomaly scoring vector includes: The dimensional features of the topological embedding feature vector are determined as the state attributes of the nodes. The multi-agent defines the weights of the directed edges between nodes based on the causal information to obtain the event response graph. The event response graph is classified and mapped according to the preset label data to obtain the anomaly score vector.

[0010] Optionally, the step of performing multi-source information fusion decision based on the anomaly scoring vector, the causal information, and the topological embedding feature vector to obtain fusion evaluation data includes: The causal strength of the collision event is determined based on the causal information. Determine the rate of change of distance between different dimensions in the topological embedding feature vector; The fusion evaluation data is obtained by weighting the anomaly scoring vector, the causal strength, and the distance change rate according to preset weight coefficients.

[0011] Optionally, a dynamic alarm threshold is determined based on the real-time vehicle status, including: The real-time driving scenario is determined based on the real-time vehicle speed and real-time suspension height in the vehicle status. Identify the target event associated with the real-time driving scenario, and dynamically adjust the default alarm threshold of the target event according to the real-time driving scenario to obtain the dynamic alarm threshold.

[0012] Optionally, determining the collision recognition result based on the dynamic alarm threshold and the fusion evaluation data includes: Based on the fusion evaluation data, determine the fusion determination event and the confidence level corresponding to the fusion determination event; Determine the target dynamic threshold corresponding to the fusion judgment event from the dynamic alarm thresholds; In response to the confidence level being greater than or equal to the target dynamic threshold, the fusion determination event is determined as the collision recognition result.

[0013] Optionally, the step of dynamically adjusting the default alarm threshold for the target event based on the real-time driving scenario to obtain the dynamic alarm threshold includes: Determine the scenario type of the real-time driving scenario; In response to the scenario type being a smooth driving scenario, the default alarm threshold is increased according to a first adjustment value corresponding to the smooth driving scenario to obtain the dynamic alarm threshold; In response to the scenario type being a bumpy driving scenario, the default alarm threshold is reduced according to a second adjustment value corresponding to the bumpy driving scenario to obtain the dynamic alarm threshold; In response to the scenario type being a parking scenario, the preset maximum threshold is determined as the dynamic alarm threshold.

[0014] Based on the same inventive concept, this application also provides a vehicle including an electronic device, the electronic device including a memory, a processor and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the method described above.

[0015] As can be seen from the above, the vehicle collision identification method and vehicle provided in this application can jointly extract features from the collected raw acoustic signals and raw vibration signals to obtain a topological embedding feature vector; perform temporal causal modeling based on the topological embedding feature vector to obtain causal information; and perform multi-agent collaborative discrimination based on the causal information and the topological embedding feature vector to obtain an anomaly score vector; perform multi-source information fusion decision-making based on the anomaly score vector, causal information, and the topological embedding feature vector to obtain fusion evaluation data; determine a dynamic alarm threshold based on the real-time vehicle status; and determine the collision identification result based on the dynamic alarm threshold and the fusion evaluation data. It directly collects the structural vibration and acoustic shock wave generated when the undercarriage scrapes, achieving direct capture of the contact event itself. The collected raw acoustic signals and raw vibration signals are extremely sensitive to physical contact and are independent of the overall vehicle body posture stability, improving the accuracy of collision event identification. It automatically learns the causal relationship and time lag between acoustic and vibration features from the topological embedding feature vector, focusing on real collision events with causal logic rather than simple signal amplitude, thus achieving the distinction between real collision events and noise. A joint sound-vibration modeling and collaborative reasoning mechanism was constructed to achieve deep fusion of multimodal information. Judgments were made based on the state of the entire coupled system, achieving a leap from data fusion to decision fusion. Based on the weighted fusion of anomaly scoring vectors, causal information, and topological embedding feature vectors, a dynamic threshold adjustment mechanism based on real-time vehicle state was introduced, forming a multi-evidence verification mechanism. This mechanism effectively suppresses false alarms under complex road conditions while maintaining high sensitivity to real collision events, resulting in accurate and reliable collision identification results. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a vehicle collision recognition method according to an embodiment of this application; Figure 2 This is a schematic diagram of a vehicle collision recognition device according to an embodiment of this application; Figure 3 This is a schematic diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0019] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] In this article, it is important to understand that any number of elements in the accompanying figures is for illustrative purposes and not for limitation, and any naming is for distinction only and has no limiting meaning.

[0021] Based on the above background description, the following situations also exist in the related technologies: In related technologies, vehicle chassis scraping detection technology mainly relies on inertial measurement units, suspension height sensors, or vehicle attitude estimation. However, the chassis structure, materials, and sound insulation design vary greatly among different vehicle models, making it difficult to generalize the identification of collision events using a single signal or model. Furthermore, single-dimensional sensor detection lacks a systematic modeling of the coupling relationship between acoustic impact signals and structural vibration signals, failing to accurately depict the true physical characteristics of collision events, thus leading to missed detections or misjudgments.

[0022] Under real-world complex driving conditions, scrape detection solutions that rely on inertial measurement units, suspension height sensors, or vehicle attitude estimation have the following drawbacks: Blind spots: When a vehicle travels at low speed over a slope, speed bump, or uneven road surface, the overall vehicle posture may remain stable and the suspension travel may not change significantly, but the chassis structural components may have already made contact with the ground. In this situation, sensors that rely on the vehicle's posture cannot respond effectively.

[0023] Weak signal characteristics: Collision events often have characteristics such as short duration, small impact amplitude, and complex direction, which makes it difficult for inertial sensors such as IMUs to form significant and stable abnormal responses at the acceleration or attitude level, making it very easy to miss detection.

[0024] Insufficient generalization ability: The chassis structure, materials and sound insulation design of different car models vary greatly, making it difficult for recognition schemes based on a single signal or a single model to be adapted to multiple car platforms and lacking universality.

[0025] Lack of multimodal coupling analysis: Related technologies lack systematic modeling of the coupling relationship between acoustic impact signals and structural vibration signals, making it impossible to accurately characterize the physical characteristics of the sound-vibration relationship in collision events, thus making it difficult to distinguish real collision events from other road surface disturbances.

[0026] The vehicle collision recognition method and vehicle provided in this application address the problem of "detection blind spots" (collision events under stable vehicle posture) by abandoning indirect inference based on vehicle posture and instead directly sensing physical contact signals. By switching signal sources, the contact event itself is directly captured. Vibration sensors (MEMS accelerometers) and arrayed capacitor microphones located at key chassis positions directly collect the structural vibrations and acoustic shock waves generated when the undercarriage scrapes. These two signals are extremely sensitive to physical contact and are independent of the overall vehicle posture stability, only affected by the vehicle's state. Joint feature extraction is performed directly from the original acoustic and vibration signals, extracting short-time energy, envelope spectrum, and other physical characteristics that directly reflect the impact event, thereby directly capturing the collision event itself.

[0027] To address the issue of "weak signal characteristics" (short-duration, small-amplitude, complex-directional chassis impacts), the signal-to-noise ratio of weak signals is enhanced through "feature compression amplification" and "temporal causal focusing." Uniform Manifold Approximation and Projection (UMAP) is used for topological compression. Leveraging the nonlinear dimensionality reduction capability of the UMAP algorithm, high-dimensional features (e.g., 128 dimensions) are compressed into a low-dimensional space (8 dimensions). This process preserves and highlights the unique topological structure of weak impact events in the feature space, enabling geometric separation from normal road surface noise, effectively "amplifying" the differences at the feature level.

[0028] Causal mining is performed using the Temporal Causal Discovery Framework (TCDF) to automatically learn the causal relationships and time lags between acoustic and vibrational features from time series data. Real, physically meaningful bottom-scraping impacts produce stable, specific causal patterns across different signal channels (e.g., sound waves leading vibrations by tens of milliseconds), while random noise does not exhibit such patterns. This allows the system to focus on "signal patterns" with causal logic, rather than simply "signal amplitudes."

[0029] To address the lack of multimodal coupling analysis, a joint acoustic-vibration modeling and collaborative reasoning mechanism is constructed to achieve deep fusion of multimodal information. First, a coupling model is established using the TCDF (Traditional Coherence Principle) model. The core output of the TCDF model (causal weight matrix) is a quantitative expression of the temporal coupling relationship between acoustic and vibration signals. Then, a decentralized multi-agent (AI agent) collaborative network (SwarmNet) is used for collaborative reasoning, combining UMAP features (node ​​states) with the TCDF causal graph (edge ​​connections) to construct a graph neural network. SwarmNet simulates the propagation and interaction of signals across multiple modalities through message passing between multiple agents (nodes), ultimately making a judgment based on the state of the entire coupled system. This achieves a leap from data fusion to decision fusion, obtaining anomaly score vectors representing "no events," "mild bottoming out," or "severe bottoming out."

[0030] To address the issue of insufficient generalization ability caused by significant differences in vehicle models, a two-level adaptive architecture of "unified framework + personalized fine-tuning" is designed to avoid this problem. First, a unified feature interface is established. The front-end uses UMAP to map the raw acoustic and vibration signals from different vehicle models into the same 8-dimensional topological space, forming a universal feature representation decoupled from specific physical structures. The back-end model fine-tuning establishes an automotive-grade acoustic and vibration modal database, incorporating data from different vehicle models and operating conditions. When the algorithm is deployed to a specific vehicle model, transfer learning and fine-tuning of the core models (SwarmNet, TCDF) are performed based on a subset of data from that model. The model can also be initialized according to the vehicle model's structural parameters, enabling the general algorithm to quickly adapt to the acoustic and vibration characteristics of specific vehicle models.

[0031] To enhance the robustness of collision event recognition, a multi-model fusion decision is adopted, which weights and fuses the outputs of UMAP (structural change), TCDF (causal logic), and SwarmNet (anomaly score vector). A dynamic threshold adjustment mechanism based on vehicle speed and suspension displacement is introduced to form a multi-evidence verification mechanism, which can effectively suppress false alarms under complex road conditions and maintain high sensitivity to real collision events.

[0032] In summary, the approach shifts from relying on inertial attitude to directly sensing acoustic and vibrational physical contact; from single signal / model to multimodal causal collaborative modeling; and from fixed models to an adaptive learning framework. By introducing a joint modeling mechanism for acoustic and vibrational signals, the approach addresses the problem of single sensing schemes being insensitive to short-term, low-amplitude undercarriage impacts, thereby achieving reliable and generalized determination of actual contact events on the vehicle chassis.

[0033] The vehicle collision recognition method provided by the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0034] In some embodiments, such as Figure 1As shown, the vehicle collision identification method includes steps 101-104.

[0035] Step 101: Perform joint feature extraction on the acquired raw acoustic signal and raw vibration signal to obtain the topological embedding feature vector.

[0036] In practice, switching signal sources to avoid noise interference requires prioritizing the acquisition of multi-source acoustic and vibration signals and the synchronization of a unified sampling frequency. The core objective of switching signal sources is to establish stable, unified, and high-quality original acoustic and vibration signals, providing effective basic data for subsequent multi-model algorithms. The process of determining the topological embedding feature vector includes steps S1-S3.

[0037] Step S1 involves multi-source signal synchronization and windowing processing. The inputs are the original acoustic signal A1 (48kHz), the original vibration signal A2 (10kHz), and the vehicle time reference A3. The processing logic aligns the timestamps of A1 and A2 based on GPS or a high-precision clock protocol, unifying them to the same time axis. Subsequently, the signals are bandpass filtered to remove noise, and sliced ​​using a 500-millisecond sliding window (50% overlap) to convert the continuous signal into discrete window data. The output is an acoustic-vibration window data sequence D1, where each data window contains aligned acoustic waveforms (24,000 points) and vibration data (5,000 points), serving as the basic unit for subsequent analysis.

[0038] Step S2: High-dimensional joint feature extraction.

[0039] The input is the synchronized acoustic-vibration window data sequence D1 output from step S1. The processing logic extracts joint time-domain and frequency-domain features from each acoustic-vibration window data sequence D1, such as short-time energy, instantaneous frequency, zero-crossing rate, and envelope spectrum, for both the acoustic and vibration signals. The output is a high-dimensional feature matrix D2, for example, a 128-dimensional matrix, which comprehensively characterizes the physical properties of the signal within the window.

[0040] Step S3: Topological feature compression based on UMAP.

[0041] The input is a high-dimensional feature matrix D2. The processing logic involves inputting the high-dimensional feature matrix D2 into the UMAP algorithm. UMAP is a non-linear dimensionality reduction method whose core principle is to preserve the local topological structure in high-dimensional data. It achieves data compression by constructing a nearest-neighbor graph and finding matching embeddings in a low-dimensional space. Optionally, the target dimension is set to 8 dimensions. The output is an 8-dimensional topological embedding feature vector D3. This step significantly reduces the data dimensionality while preserving the key feature structure of collision events, facilitating subsequent model processing.

[0042] During vehicle operation under typical road conditions (such as speed bumps, gravel roads, unpaved surfaces, and underground parking garage ramps), vibration signals and acoustic impact signals are collected by vibration sensors (such as MEMS accelerometers) deployed at key locations on the chassis and arrayed capacitor microphones placed on body structural components or underbody protection plates. The sampling frequency for vibration signals is 10kHz, and the sampling frequency for acoustic impact signals is 48kHz. Because the two types of sensor signals have different time bases and sampling densities, they require unified synchronization processing.

[0043] First, based on GPS timestamps and the high-precision clock synchronization protocol within the vehicle's electronic architecture, the original acoustic and vibration signals are aligned to a unified reference time axis. Second, bandpass filtering is applied to the original acoustic and vibration signals to remove environmental background noise and broadband vibration artifacts. A sliding window slicing mechanism is then used for signal pane processing, with a window size of 500ms and an overlap rate of 50%, ensuring that the signal within each time window contains complete impact characteristics.

[0044] Based on this, all signal windows are cached in a circular buffer queue, providing the downstream model with raw acoustic signals, raw vibration signals, and timestamp information simultaneously, forming a triplet input data format. The final output is an aligned sequence of acoustic and vibration signal windows, with each record containing the raw acoustic waveform, vibration channel vector, synchronization timestamp, and vehicle driving status identifier, serving as the basic data source for subsequent feature compression and modeling stages.

[0045] After acquiring synchronized raw acoustic and vibration signals, a UMAP-based acoustic-vibration topology embedding space and feature compression are performed on the raw acoustic and vibration signals to facilitate subsequent processing. By introducing the nonlinear manifold learning algorithm UMAP, topology-preserving compression mapping is performed on the multi-channel acoustic-vibration signals, aiming to retain the weak but highly characteristic signal topology relationships in collision events.

[0046] In each sliding window dataset, acoustic and vibration signals are first uniformly clipped to the same length, and then combined with time-domain and frequency-domain features such as short-time energy, instantaneous frequency, zero-crossing rate, and short-time envelope are used to construct a high-dimensional feature matrix. This matrix is ​​then input into the UMAP embedding engine to perform a neighborhood-preserving low-dimensional mapping process.

[0047] The UMAP algorithm constructs a local K-nearest neighbor graph in a high-dimensional space and then searches for point sets in a low-dimensional space that preserve the original adjacency probability distribution to retain key local structural features. This process enables significant geometric separation between bottom-scraping impact signals and other road surface disturbances, while compressing data dimensionality and reducing the computational load on subsequent models. Optionally, an 8-dimensional embedding dimension is chosen for UMAP, and experiments have demonstrated that this dimension balances expressive power and inference efficiency. The final output is an 8-dimensional topological embedding feature vector corresponding to each sliding window data compression, preserving the topological continuity between acoustic and vibration signals while serving as a unified input interface for subsequent causal modeling and anomaly detection models, thus improving the generalization ability of the recognition process.

[0048] Step 102: Perform temporal causal modeling based on topological embedding feature vectors to obtain causal information, and perform multi-agent collaborative discrimination based on causal information and topological embedding feature vectors to obtain anomaly score vectors.

[0049] In practice, after obtaining the topological embedding feature vector, causal mining and multidimensional anomaly detection are required to initially determine whether collision events exist. The initial determination result is represented as an anomaly score vector, which consists of the probability / score of three events: "no event," "mild scraping," and "severe scraping." The process of determining the anomaly score vector includes steps S4 and S5.

[0050] Step S4: Temporal causal relationship mining based on TCDF.

[0051] The input is a continuous sequence of acoustic and vibration feature vectors (D3) consisting of multiple topological embedding feature vectors (i.e., 8-dimensional feature vectors of multiple time windows). The processing logic involves inputting the continuous sequence of D3 vectors into the TCDF model. TCDF aims to automatically discover temporal causal relationships between variables in multi-channel time series (i.e., which channel's change leads and influences another channel). It learns the causal direction, time lag, and influence strength between feature dimensions through a convolutional attention mechanism. The output contains an 8x8 causal weight matrix and an 8-dimensional time-sensitivity vector containing causal information D4. Causal information D4 quantifies the interaction patterns of each acoustic and vibration feature dimension before and after the collision event.

[0052] Step S5: Multidimensional anomaly detection based on SwarmNet.

[0053] The input consists of a topological embedding feature vector D3 (representing node states) and causal information D4 (defining graph connectivity). The processing logic uses the SwarmNet model to integrate D3 and D4 into a single graph structure, where each dimension of D3 represents a node in the graph, and the causal weights of D4 define the connections between nodes. The model mimics swarm intelligence for collaborative reasoning, iteratively evaluating the input features over multiple rounds. The output is a 3-dimensional anomaly score vector D5, where the three dimensions correspond to the probability or score of three event categories: "no event," "mild bottoming out," and "severe bottoming out."

[0054] The process of extracting temporal causal relationships and bottom-scraping impact patterns based on TCDF is as follows: The TCDF model automatically mines the temporal causal relationships between different channels from the UMAP-reduced acoustic-vibration feature sequence, which is used to identify the temporal evolution structure of acoustic-vibration coupling modes in chassis collision events. The TCDF model uses a convolutional attention mechanism to construct a temporal causal graph. By performing dynamic graphing and edge weight training on multi-channel time series, it identifies the channel response sequence and influence intensity before and after the impact event. In this process, the UMAP features within each sliding window are treated as a channel node. The model expands on the sliding window sequence to construct a convolutional receptive field. A trainable masking mechanism determines the temporal causal direction, thereby outputting the causal path matrix between acoustic and vibration channels and the saliency score of each path. In engineering implementation, this module can identify causal patterns such as acoustic response leading vibration response by tens of milliseconds, and a long duration of response in a certain frequency band, effectively separating normal structural rebound sound from substantial structural impact events. The final model will output a multi-channel causal graph and a time-sensitive intensity vector. These outputs will be used as structural information in the graph embedding part of the downstream SwarmNet model to improve the robustness of bottom-scraping anomaly pattern recognition.

[0055] This section describes the multidimensional anomaly detection process using the SwarmNet model. The SwarmNet model is used to detect collision events from inputs that integrate UMAP features and TCDF causal graph information. SwarmNet is a neural network architecture that incorporates the concept of swarm intelligence, possessing high-dimensional dynamic information collaborative computing capabilities, and is suitable for processing multi-channel, heterogeneous, and nonlinearly coupled input features. In determining the anomaly scoring vector, each sliding window input contains the topological embedding feature vector D3 from the UMAP output and the causal information D4 inferred from the TCDF. The SwarmNet input channel module integrates these into node states and graph connectivity, while simultaneously introducing a multi-agent path search mechanism to simulate signal propagation paths, thus forming an event response graph. The core SwarmNet network includes a feature proxy pool, a policy sampler, and a structural feedback unit, guiding the model to converge to a stable detection strategy in complex inputs through multiple iterations. During the training phase, supervised training is performed using labeled, real-world scraping data. The labels originate from contact switches triggered by scraping the bottom of the test vehicle and human auditory judgment results. The model outputs event type scores / probabilities at the sliding window level, including scores / probabilities for three events: "no event," "minor underrun," and "severe underrun," represented by anomaly score vectors. This provides the primary data support for subsequent fusion mechanisms. The SwarmNet model exhibits strong cross-scene transfer capabilities and input modality robustness, adapting to the acoustic and vibration characteristics of different vehicle models and road conditions, thus improving generalization ability.

[0056] Step 103: Make a multi-source information fusion decision based on the anomaly scoring vector, causal information and topological embedding feature vector to obtain fusion evaluation data.

[0057] Step 104: Determine the dynamic alarm threshold based on the real-time vehicle status, and determine the collision recognition result based on the dynamic alarm threshold and the fusion evaluation data.

[0058] In practice, the score / probability of each event can be directly determined based on the anomaly scoring vector, yielding a preliminary judgment result. For example, an anomaly scoring vector of (2, 4, 9) indicates a score of 2 for no event, 4 for a minor collision, and 9 for a severe collision, preliminarily confirming a severe collision. Similarly, an anomaly scoring vector of (5%, 10%, 85%) indicates a probability of 5% for no event, 10% for a minor collision, and 85% for a severe collision, also preliminarily confirming a severe collision. To further ensure the accuracy of the recognition results, causal information and topological embedding feature vectors are used to verify the preliminary judgment result of the anomaly scoring vector, and dynamic evaluation is performed based on the real-time vehicle status to determine the final collision recognition result. The fusion verification process is shown in step S6.

[0059] Step S6: Multi-model fusion decision and dynamic threshold adjustment. The inputs are the anomaly score vector D5 output by SwarmNet, the causal information D4 output by TCDF, the topological embedding feature vector D3 output by UMAP, and additional vehicle state information A4 (such as vehicle speed and suspension displacement). The processing logic involves constructing a fusion decision framework. The anomaly score vector D5 output by SwarmNet is used as the primary criterion, while the topological embedding feature vector D3 output by UMAP and the causal information D4 output by TCDF are referenced for weighted decision-making. Furthermore, an adaptive threshold adjustment mechanism is introduced to dynamically adjust the threshold for triggering collision event determination based on the real-time vehicle state (A4) to reduce the false alarm rate. The output is the final collision identification result O1 for each sliding window, i.e., no event / minor undercarriage damage / severe undercarriage damage and its confidence level.

[0060] The collision identification result is primarily determined by a three-model fusion decision and an adaptive threshold adjustment mechanism. The outputs of UMAP, TCDF, and SwarmNet are fused to construct a unified collision event identification framework, enabling adaptive output adjustment for different road conditions and vehicle platforms. The fusion mechanism is based on a decision-level integration strategy. First, the output features of the three sub-models are collected: the anomaly score vector D5 from SwarmNet, the causal information D4 from TCDF, and the topological embedding feature vector D3 from UMAP. Then, the fusion strategy logic is constructed at the control end, where UMAP output provides a reference for the rate of signal structure change, TCDF provides temporal causal criteria support, and SwarmNet, as the main discrimination module, provides classification scores. In the fusion module, a learnable threshold adjuster is introduced. This component dynamically adjusts the trigger threshold based on the vehicle's operating status (vehicle speed, road roughness estimation, suspension displacement trend), thereby suppressing the false alarm rate and enhancing the detection rate of very short-term collision events. The final collision recognition result is an event recognition label (e.g., no event, minor scrape, or severe scrape) and confidence level (or score) for each sliding window, which can be retrieved by the vehicle control system or archived by the data logging system.

[0061] After obtaining the collision recognition results, in order to ensure the accuracy of the model on different vehicles, it is necessary to fine-tune the model according to the vehicle structure parameter A5. The fine-tuning process is shown in step S7.

[0062] Step S7: Database Construction and Vehicle-Specific Adaptive Fine-Tuning. Inputs include collision recognition result O1, raw acoustic and vibration signals, and vehicle structural parameters A5. The fine-tuning process is an offline learning and optimization process. First, an automotive-grade acoustic and vibration modal database is constructed using collected data from multiple vehicle models and operating conditions. Then, based on this database, transfer learning and fine-tuning are performed on a subset of data from a specific vehicle model, enabling the model to adapt to the characteristics of different chassis structures and materials. The output is the optimized model parameters O2 for the specific vehicle model, thereby improving the algorithm's generalization ability and recognition accuracy across different vehicle platforms.

[0063] The mechanism for constructing an automotive-grade acoustic and vibration modal database and fine-tuning a personalized model revolves around the issues of model generalization ability and adaptation to different vehicle models. An automotive-grade acoustic and vibration modal database is established, and personalized model fine-tuning is performed based on this database. In the database construction phase, raw acoustic and vibration data collected from different test vehicles under various typical operating conditions are organized and indexed by vehicle model-VIN-test condition, forming a modal sample library containing multi-dimensional fields such as raw waveforms, event labels, structural parameters, and road condition categories. Based on this, feature labels are added to different chassis structures (such as high ground clearance SUVs, sports sedans, and differences in front and rear skid plate materials), forming a structured set of difference parameters. In the model fine-tuning phase, the SwarmNet model is transferred and trained. After loading pre-trained weights, fine-tuning is performed on a subset of data corresponding to a specific vehicle model, ensuring that the model maintains global capabilities while also considering the specific vehicle model's undercarriage scraping characteristics. Simultaneously, the causal structure in the TCDF is reinitialized with the adjacency bias matrix based on the vehicle model's material parameters, improving the matching degree of the causal graph under different structures. Through targeted fine-tuning, the entire algorithm system possesses advantages in reconfigurability and low-cost deployment when facing the promotion of multiple vehicle platforms by automakers, enhancing the commercial applicability and implementation potential of the patented solution.

[0064] In summary, through a complete and closed-loop processing flow of raw signal acquisition → feature extraction and compression → temporal causal modeling → intelligent agent collaborative discrimination → multi-source information fusion decision-making → continuous learning and optimization, accurate and robust identification of vehicle chassis collision events can be achieved.

[0065] In some embodiments, joint feature extraction is performed on the acquired raw acoustic signal and raw vibration signal to obtain a topological embedding feature vector, including: The original acoustic signal and the original vibration signal are time-aligned according to the vehicle clock to obtain the initial acoustic signal and the initial vibration signal; The continuous initial acoustic signal and initial vibration signal are converted into a discrete acoustic-vibration window data sequence according to a sliding window of preset time length. Joint feature extraction and topological feature compression are performed on the acoustic vibration window data sequence to obtain the topological embedding feature vector.

[0066] The process involves joint feature extraction and topological feature compression of the acoustic vibration window data sequence to obtain a topological embedding feature vector, including: The joint acoustic and vibration features are extracted from the acoustic and vibration window data sequence to obtain the high-dimensional acoustic and vibration feature matrix; Based on the preset target dimension, the high-dimensional acoustic and vibration feature matrix is ​​subjected to neighborhood-preserving dimensionality reduction mapping to obtain the topological embedding feature vector.

[0067] In practice, the original acoustic and vibration signals are first synchronized and panned. First, a filtering preprocessing step is performed, bandpass filtering the original signals to remove environmental noise (such as wind noise and engine vibration) and irrelevant frequency interference. The time alignment process involves synchronizing the original acoustic signals (48kHz) and vibration signals (10kHz) from different sensors to a unified time reference using the vehicle clock. Based on timestamp interpolation or resampling algorithms, the heterogeneous sensor data are mapped to the same time axis, ensuring that the physical time relationship between the acoustic and vibration signals is preserved.

[0068] The sliding window slicing process converts continuous initial acoustic and vibration signals into discrete acoustic-vibration window data sequences based on a sliding window of preset time length. For example, a sliding window with a fixed duration of 500 milliseconds can be used to segment continuous initial acoustic and vibration signals, with a window overlap rate of 50%, ensuring that transient events (such as short-term bottoming) can be completely captured in one or more windows. Sliding windows are a common method in time-series signal processing. By using overlapping windows, information loss is reduced, and infinitely long signals are transformed into finite-length sequences, making them suitable for real-time or batch analysis.

[0069] The time alignment and sliding window slicing processes enable data standardization and event localization. They address the issues of varying sampling rates and inconsistent time bases across multiple sensors, providing time-aligned standardized data for subsequent analysis. By converting continuous signals into discrete time windows, they facilitate independent feature extraction and event detection for each window, adapting to the short-lived and sudden nature of collision events.

[0070] Collision events manifest as specific transient patterns in acoustic and vibration signals (such as high-frequency impacts and resonance attenuation). Time-frequency domain features can effectively capture these patterns, making them more conducive to model learning than the original signal. Therefore, for the initially processed data, high-dimensional joint feature extraction is performed first. For each synchronized acoustic and vibration data window, multiple types of features are calculated. Time-domain features include short-time energy (signal strength), zero-crossing rate (signal frequency roughness), and envelope (amplitude variation trend). Frequency-domain features include instantaneous frequency and spectral centroid extracted through short-time Fourier transform.

[0071] The features of acoustic and vibration signals are concatenated to form a high-dimensional acoustic-vibration feature matrix (e.g., 128-dimensional). The original waveform is transformed into a feature set that better reflects the physical characteristics of the collision event, such as the energy concentration and frequency component changes of the impact event. This provides rich information input for subsequent dimensionality reduction steps, avoiding the computational burden of directly processing high-dimensional raw data.

[0072] Then, UMAP-based topological feature compression is used to reduce high-dimensional features (e.g., 128-dimensional) to low-dimensional (8-dimensional), significantly reducing the computational complexity of subsequent models. It preserves the topological separability of collision events and normal road surface noise in the feature space; similar events cluster in the low-dimensional space, while different events are far apart, improving the robustness of subsequent classification. The output high-dimensional acoustic and vibration feature matrix is ​​input into the UMAP algorithm. The algorithm first calculates the nearest neighbors of each data point in the high-dimensional space, constructing a weighted graph to represent the local topological structure. Random points are initialized in the target low-dimensional space (set to 8-dimensional), and the loss function is optimized to make the distance relationships between points in the low-dimensional space match the adjacency probability distribution in the high-dimensional space as closely as possible. UMAP assumes that the high-dimensional data is distributed on a low-dimensional manifold. It constructs high-dimensional adjacency relationships through fuzzy topological representation and optimizes them in the low dimension using cross-entropy loss, thereby preserving the local and global structure of the data.

[0073] In some embodiments, causal information includes a causal weight matrix and a time-sensitive vector; causal information is obtained by performing time-series causal modeling based on topological embedding feature vectors, including: The degree of mutual influence between each dimension in the topological embedding feature vector is quantified to obtain the causal weight matrix; By determining the causal temporal pattern between acoustic and vibrational features in the topological embedding feature vector, a time-sensitivity vector is obtained.

[0074] In practical implementation, temporal causal relationship mining based on TCDF requires inputting topological embedding feature vectors (an 8-dimensional vector for each time window) as multi-channel time series data into the TCDF model. TCDF uses a one-dimensional convolutional neural network combined with an attention mechanism to learn a trainable mask for each channel. This mask is used to infer the causal direction and time lag between channels (i.e., which channel's change precedes another channel). In collision events, acoustic signals (airborne propagation) are typically captured by sensors faster than vibration signals (structural propagation). TCDF quantifies this temporal difference by quantifying the degree of mutual influence between each dimension of the topological embedding feature vector, obtaining a causal weight matrix. By determining the causal temporal pattern between acoustic and vibration features in the topological embedding feature vector, a time sensitivity vector is obtained, providing a physical basis for event determination.

[0075] The TCDF model outputs an 8x8 causal weight matrix representing the causal strength between dimensions, and a time-sensitivity vector indicating causal delay. This enables automatic identification of causal timing patterns between acoustic and vibration signals in collision events, such as "sound pulse preceding structural vibration response," a physical characteristic of collision events. This helps distinguish between genuine ground-level collisions (with causal order) and random noise (without stable causality), reducing false positives.

[0076] In some embodiments, multi-agent collaborative discrimination is performed based on causal information and topological embedding feature vectors to obtain an anomaly scoring vector, including: The dimensional features of the topological embedding feature vector are determined as the state attributes of the nodes, and the multi-agent defines the weights of the directed edges between nodes based on causal information to obtain the event response graph. The event response graph is classified and mapped based on the preset label data to obtain the anomaly score vector.

[0077] In practice, the process of determining the event response diagram includes: Graph initialization: The eight dimensions of the topological embedding feature vector are treated as eight "agent" nodes. The initial state of each node is its corresponding feature value. The causal weight matrix output by the TCDF is used to initialize the connections between nodes. A non-zero element W_{ij} in the weight matrix indicates the existence of an edge from node j (cause) to node i (effect), and the weight of the edge is the causal strength W_{ij}. This forms a directed weighted graph that reflects the potential causal relationships between features. The time-sensitivity vector and the causal weight matrix together define the dynamic temporal properties of the connections and interactions between nodes (features) in the SwarmNet graph neural network. This allows the network to not only know whether there is a causal relationship between features (weight matrix) but also the time scale of this influence (sensitivity vector), thus enabling more accurate temporal pattern inference.

[0078] Multi-agent cooperative reasoning (graph information propagation and update): This process simulates how agents (nodes) reach global consensus through communication (passing messages along edges). Message Passing: In each iteration, each node aggregates information from its neighbors (which may be incoming or outgoing neighbors depending on the directed edges). The passed information typically includes the state of the neighboring nodes and the weights of their edges with the current node. For example, a node representing "high-frequency acoustic energy" might receive a message from a node representing "low-frequency vibration amplitude," with the message strength adjusted by the causal weights between them.

[0079] State Update: After receiving messages from all its neighbors, each node updates its own state by combining its current state with the aggregated messages using a learnable update function (usually a small neural network). This function is the embodiment of the "policy sampler" or "agent policy".

[0080] Multiple iterations: The above process is repeated multiple times. After several iterations, the state of each node no longer only contains the initial UMAP features, but also incorporates information from other relevant nodes in the graph, modulated by causal relationships. Ultimately, the states of all nodes together constitute a distributed representation of the entire system (acoustic-vibration coupling mode) within the current time window, i.e., the dynamically evolved event response graph. The event response graph captures specific signal propagation and interaction patterns in collision events.

[0081] The process of generating anomaly score vectors includes: Graph-Level Readout: After multiple iterations and the node states have stabilized, a global, classification-oriented representation needs to be extracted from the entire event response graph. SwarmNet uses a readout function (also a learnable network) to aggregate the final states of all nodes. A common approach is to use global pooling (such as summation, averaging, or attention-weighted summation) to compress the state vectors of the eight nodes into a single, fixed-length graph embedding vector. This vector integrates the topological information of the entire graph and the node features.

[0082] Classification and scoring: The graph embedding vector obtained in the previous step is input into a classification layer (usually a fully connected neural network). This classification layer, trained on labeled data (the documentation mentions "supervised training using real-world scraping data"), is able to map the graph embeddings to three event categories.

[0083] Output anomaly score vector: The output of the classification layer is a 3-dimensional anomaly score vector D5. This vector is usually normalized to a probability distribution using the Softmax function, and its three components represent the probability / score of the model judging the current time window as "no event", "mild bottoming out", and "severe bottoming out", respectively.

[0084] In some embodiments, a multi-source information fusion decision is made based on anomaly scoring vectors, causal information, and topological embedding feature vectors to obtain fusion evaluation data, including: Determine the causal strength of the collision event based on causal information; Determine the rate of change of distance between different dimensions in the topological embedding feature vector; The abnormal scoring vector, causal strength, and distance change rate are weighted according to the preset weight coefficients to obtain the fusion evaluation data.

[0085] In practical implementation, the core objective of multi-source information fusion decision-making is to effectively integrate the outputs of three independent models (UMAP, TCDF, and SwarmNet) to form a unified, stable final decision system that can adapt to different driving environments. It aims to solve the problems of misjudgment and missed detection that may exist with a single model, and to improve the overall robustness of the system through environmental perception.

[0086] The fusion process is a decision-level integration, not a feature-level fusion. It first extracts intermediate features or results with clear physical or discriminative significance from each sub-model: The master criterion from SwarmNet (anomaly score vector D5) is a 3-dimensional anomaly score vector that directly represents the model's probability / score for the three categories: "no event," "mild bottoming out," and "severe bottoming out." This is the core basis for the fusion decision and can be used directly.

[0087] For causal structure support (causal information D4) from TCDF, principal component analysis or matrix norm calculation of the 8x8 causal weight matrix is ​​performed through feature extraction to obtain one or more scalars, which are used to quantify the causal strength of the causal relationship within the current time window. High causal strength indicates a strong temporal coupling pattern between signals, supporting the determination that "an event has occurred." It provides temporal logical evidence of the event's occurrence. A genuine physical bottoming should produce a stable and significant causal pattern.

[0088] For topological change support from UMAP (topological embedding feature vector D3), the rate of change of distance (such as Euclidean or cosine distance) between consecutive time windows of the 8-dimensional topological embedding feature vector is calculated through feature extraction. A sudden change in the rate of change indicates a rapid evolution of the signal structure in the compressed feature space, potentially corresponding to a shock event. This provides evidence of abrupt changes in the signal structure during the event.

[0089] The weighted fusion decision process is the process of calculating the final score of a certain type of event. The final score = w1 × anomaly score (determined based on the anomaly score vector) + w2 × causal strength + w3 × distance change rate.

[0090] Among them, w1, w2, and w3 are adjustable weight coefficients, with w1 usually being the largest to reflect the dominant position of the abnormal rating vector.

[0091] Optionally, rules can be set, for example, to only classify a collision event as a collision if the SwarmNet score exceeds a base threshold, the TCDF causality strength is greater than threshold α, and the distance change rate is greater than threshold β. This effectively filters out isolated noise false alarms that may be generated by SwarmNet's individual judgment. Based on the fused "final score," it is mapped to three levels: "no event," "mild undershoot," and "severe undershoot," and a preliminary confidence level is generated to obtain the fused evaluation data. In some embodiments, determining a dynamic alarm threshold based on real-time vehicle status includes: The real-time driving scenario is determined based on the vehicle's real-time speed and real-time suspension height. Identify the target events associated with the real-time driving scenario, and dynamically adjust the default alarm thresholds for the target events based on the real-time driving scenario to obtain dynamic alarm thresholds.

[0092] Among them, the default alarm threshold for target events is dynamically adjusted based on the real-time driving scenario to obtain the dynamic alarm threshold, which includes: Determine the scenario type for the real-time driving scenario; In response to the scenario type being a smooth driving scenario, the default alarm threshold is increased according to the first adjustment value corresponding to the smooth driving scenario to obtain the dynamic alarm threshold; In response to the scenario type being bumpy driving scenario, the default alarm threshold is reduced according to the second adjustment value corresponding to the bumpy driving scenario to obtain the dynamic alarm threshold; In response to the scenario type being parking, the preset maximum threshold is set as the dynamic alarm threshold.

[0093] In practice, the threshold adjustment is based on the real-time vehicle status A4.

[0094] Vehicle speed: When driving smoothly at high speeds, the probability of unexpected chassis collisions is extremely low; therefore, the detection threshold should be significantly increased to suppress false alarms caused by road noise. When traversing rough roads at low speeds, the risk of bottoming out increases; therefore, the threshold should be appropriately lowered to improve sensitivity.

[0095] Suspension displacement / height: When the suspension is severely compressed (such as when going over a speed bump), the vehicle body posture changes greatly, and there is a lot of impact noise that is not scraping the bottom. At this time, the threshold for judging "severe scraping the bottom" can be increased, but it needs to be adjusted carefully to avoid missing real contact.

[0096] The threshold adjuster can be a simple lookup table, a lightweight neural network, or a decision tree model. Taking the vehicle state as input, it determines the real-time driving scenario based on the real-time vehicle speed and suspension height, and identifies the target event associated with the real-time driving scenario (e.g., a severe collision event). It then dynamically adjusts the default alarm threshold for the target event based on the real-time driving scenario to obtain a dynamic alarm threshold. The output dynamic alarm thresholds include a first dynamic threshold corresponding to no event, a second dynamic threshold corresponding to minor undercarriage damage, and a third dynamic threshold corresponding to severe undercarriage damage.

[0097] When determining the dynamic alarm threshold, the scenario type of the real-time driving scenario must first be determined. If the scenario type is a smooth driving scenario, the probability of chassis accidental collision is extremely low, allowing for a significant increase in the alarm threshold for all target events to suppress false alarms due to road noise. In this case, the default alarm threshold is increased based on the first adjustment value corresponding to the smooth driving scenario to obtain the dynamic alarm threshold. For example, taking a first adjustment value of 0.2, when the target event is no event, the default alarm threshold of 0.3 is increased to the dynamic alarm threshold of 0.5; when the target event is minor undercarriage scraping, the default alarm threshold of 0.6 is increased to the dynamic alarm threshold of 0.8; and when the target event is severe undercarriage scraping, the default alarm threshold of 0.75 is increased to the dynamic alarm threshold of 0.95. Furthermore, different first adjustment values ​​can be configured for different target events.

[0098] If the scenario type is a bumpy driving scenario, the risk of bottoming out increases. To improve sensitivity to real collisions, the alarm thresholds for all target events are appropriately lowered. The default alarm threshold is reduced based on a second adjustment value corresponding to the bumpy driving scenario, resulting in a dynamic alarm threshold. For example, with a second adjustment value of 0.1, if the target event is no event, the default alarm threshold of 0.3 is reduced to a dynamic alarm threshold of 0.2; if the target event is minor bottoming out, the default alarm threshold of 0.6 is reduced to a dynamic alarm threshold of 0.5; and if the target event is severe bottoming out, the default alarm threshold of 0.75 is reduced to a dynamic alarm threshold of 0.65. Different values ​​for the second adjustment value can also be configured for different target events.

[0099] If the scenario type is a parking scenario and it is determined that there is almost no risk of collision, the alarm threshold for all target events is significantly increased, that is, the preset maximum threshold (e.g., 0.99) is set as the dynamic alarm threshold.

[0100] In some embodiments, determining the collision identification result based on a dynamic alarm threshold and fusion evaluation data includes: Based on the fusion assessment data, determine the fusion decision events and the corresponding confidence levels for each fusion decision event; Determine the target dynamic threshold corresponding to the fusion judgment event from the dynamic alarm threshold; When the confidence level is greater than or equal to the target dynamic threshold, the fusion decision event is determined as the collision recognition result.

[0101] In practice, if the initial judgment result output by the fusion module is that the fusion judgment event is severe bottom scraping, the third dynamic threshold corresponding to the severe collision event in the dynamic alarm threshold will be used as the target dynamic threshold; only when the fusion score (confidence) of "severe bottom scraping" is greater than the dynamically calculated current target dynamic threshold will the collision recognition result of "severe bottom scraping" be finally output.

[0102] It should be noted that the method in this embodiment can be executed by a single device, such as a computer or server. The method can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the method described.

[0103] It should be noted that the above description describes some embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0104] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a vehicle collision recognition device.

[0105] refer to Figure 2 The vehicle collision identification device includes: The joint feature extraction module 10 is configured to perform joint feature extraction on the acquired raw acoustic signal and raw vibration signal to obtain a topological embedding feature vector; The multidimensional anomaly discrimination module 20 is configured to: perform temporal causal modeling based on topological embedding feature vectors to obtain causal information, and perform multi-agent collaborative discrimination based on causal information and topological embedding feature vectors to obtain anomaly score vectors; The information fusion decision module 30 is configured to: make multi-source information fusion decisions based on anomaly scoring vectors, causal information and topological embedding feature vectors to obtain fusion evaluation data; The dynamic threshold determination module 40 is configured to: determine the dynamic alarm threshold based on the real-time vehicle status, and determine the collision recognition result based on the dynamic alarm threshold and the fusion evaluation data.

[0106] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0107] The apparatus of the above embodiments is used to implement the corresponding vehicle collision recognition method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0108] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the vehicle collision recognition method described in any of the above embodiments.

[0109] Figure 3 This embodiment illustrates a more specific hardware structure of an electronic device. The device may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0110] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0111] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0112] The input / output interface 1030 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0113] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0114] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0115] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0116] The electronic devices described above are used to implement the corresponding vehicle collision recognition methods in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0117] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the vehicle collision recognition method as described in any of the above embodiments.

[0118] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0119] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the vehicle collision recognition method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0120] Based on the same inventive concept, corresponding to any of the above embodiments, this application also provides a vehicle, including the electronic device or vehicle collision identification device of the above embodiments, and executes the vehicle collision identification method as described in any of the above embodiments through the electronic device or vehicle collision identification device of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0121] It is understood that before using the technical solutions of the various embodiments in this application, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0122] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations described in this application.

[0123] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0124] It is understood that the above notification and user authorization process is merely illustrative and does not limit the implementation of this application. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this application.

[0125] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0126] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0127] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0128] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the claims of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

Claims

1. A method for identifying vehicle collisions, characterized in that, include: Joint feature extraction is performed on the acquired raw acoustic signal and raw vibration signal to obtain the topological embedding feature vector; Temporal causal modeling is performed based on the topological embedding feature vector to obtain causal information, and multi-agent collaborative discrimination is performed based on the causal information and the topological embedding feature vector to obtain an anomaly score vector. Based on the anomaly scoring vector, the causal information, and the topological embedding feature vector, a multi-source information fusion decision is made to obtain fusion evaluation data; The dynamic alarm threshold is determined based on the real-time vehicle status, and the collision recognition result is determined based on the dynamic alarm threshold and the fused evaluation data.

2. The vehicle collision identification method according to claim 1, characterized in that, The joint feature extraction of the acquired raw acoustic signal and raw vibration signal to obtain the topological embedding feature vector includes: The original acoustic signal and the original vibration signal are time-aligned according to the vehicle clock to obtain the initial acoustic signal and the initial vibration signal; The continuous initial acoustic signal and initial vibration signal are converted into a discrete acoustic-vibration window data sequence according to a sliding window of preset time length. The acoustic vibration window data sequence is subjected to joint feature extraction and topological feature compression to obtain the topological embedding feature vector.

3. The vehicle collision identification method according to claim 2, characterized in that, The process of performing joint feature extraction and topological feature compression on the acoustic vibration window data sequence to obtain the topological embedding feature vector includes: Extract the joint acoustic and vibration features from the acoustic and vibration window data sequence to obtain the high-dimensional acoustic and vibration feature matrix; The high-dimensional acoustic-vibration feature matrix is ​​subjected to neighborhood-preserving dimensionality reduction mapping according to the preset target dimension to obtain the topological embedding feature vector.

4. The vehicle collision identification method according to claim 1, characterized in that, The causal information includes a causal weight matrix and a time-sensitive vector; the step of performing time-series causal modeling based on the topological embedding feature vector to obtain causal information includes: The degree of mutual influence between each dimension in the topological embedding feature vector is quantified to obtain the causal weight matrix; The causal temporal pattern between acoustic and vibration features in the topological embedding feature vector is determined to obtain the time sensitivity vector.

5. The vehicle collision identification method according to claim 1 or 4, characterized in that, The step of performing multi-agent collaborative discrimination based on the causal information and the topological embedding feature vector to obtain an anomaly scoring vector includes: The dimensional features of the topological embedding feature vector are determined as the state attributes of the nodes. The multi-agent defines the weights of the directed edges between nodes based on the causal information to obtain the event response graph. The event response graph is classified and mapped according to the preset label data to obtain the anomaly score vector.

6. The vehicle collision identification method according to claim 1, characterized in that, The step of performing multi-source information fusion decision based on the anomaly scoring vector, the causal information, and the topological embedding feature vector to obtain fusion evaluation data includes: The causal strength of the collision event is determined based on the causal information. Determine the rate of change of distance between different dimensions in the topological embedding feature vector; The fusion evaluation data is obtained by weighting the anomaly scoring vector, the causal strength, and the distance change rate according to preset weight coefficients.

7. The vehicle collision identification method according to claim 1, characterized in that, The step of determining the dynamic alarm threshold based on real-time vehicle status includes: The real-time driving scenario is determined based on the real-time vehicle speed and real-time suspension height in the vehicle status. Identify the target event associated with the real-time driving scenario, and dynamically adjust the default alarm threshold of the target event according to the real-time driving scenario to obtain the dynamic alarm threshold.

8. The vehicle collision identification method according to claim 7, characterized in that, The step of determining the collision recognition result based on the dynamic alarm threshold and the fusion evaluation data includes: Based on the fusion evaluation data, determine the fusion determination event and the confidence level corresponding to the fusion determination event; Determine the target dynamic threshold corresponding to the fusion judgment event from the dynamic alarm thresholds; In response to the confidence level being greater than or equal to the target dynamic threshold, the fusion determination event is determined as the collision recognition result.

9. The vehicle collision identification method according to claim 7, characterized in that, The step of dynamically adjusting the default alarm threshold for the target event based on the real-time driving scenario to obtain the dynamic alarm threshold includes: Determine the scenario type of the real-time driving scenario; In response to the scenario type being a smooth driving scenario, the default alarm threshold is increased according to a first adjustment value corresponding to the smooth driving scenario to obtain the dynamic alarm threshold; In response to the scenario type being a bumpy driving scenario, the default alarm threshold is reduced according to a second adjustment value corresponding to the bumpy driving scenario to obtain the dynamic alarm threshold; In response to the scenario type being a parking scenario, the preset maximum threshold is determined as the dynamic alarm threshold.

10. A vehicle, characterized in that, The device includes an electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the program, implements the method as claimed in any one of claims 1 to 9.

Images