A test method for intelligent assisted driving test

Abstract

The application provides a test method for intelligent auxiliary driving test, constructs a test problem description grabbing model, directly collects signals uploaded by an intelligent vehicle in real time through a vehicle-mounted Ethernet and a CAN FD direct connection interface without using other data interfaces, triggers a test problem description reporting process once abnormal signals are monitored, and ensures that the vehicle meets safety requirements of real vehicle road test.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent auxiliary driving test, in particular to a test method for intelligent auxiliary driving test. BACKGROUND

[0002] With the vigorous development of new energy vehicles, intelligent auxiliary driving emerges as the times require. In order to improve the research and development efficiency, automatic test means is used to realize the test for intelligent driving. In the real vehicle road test stage, whether bug occurs is detected by capturing various sensor signals collected by the intelligent system. During the operation of the vehicle, the signal values in the normally operating vehicle, such as radar ranging signal, brake signal, vehicle speed signal or any other signal, are a stable change process no matter the data moves in any direction. If the signal value jumps at a certain moment or under certain conditions, after excluding sensor noise, the abnormal signal is compared with the preset stability threshold to determine whether it is an abnormal signal change, that is, whether bug occurs. For example: the positioning sensor outputs the satellite positioning latitude and longitude of the vehicle in real time during high-speed driving of the vehicle; under normal circumstances, the positioning signal will be output in a uniform change form according to the driving direction and speed change of the vehicle. However, if the positioning signal jumps irregularly during high-speed driving of the vehicle in thunder and lightning weather, after excluding occasional noise, it can be judged that the positioning sensor has a positioning drift problem. At this time, a bug report needs to be submitted. The bug report includes: priority: medium, does not affect the safety of the vehicle, but affects the path planning and decision of the vehicle; type: perception bug, positioning sensor, positioning anomaly; fault description: positioning signal drift under high-speed road and thunder and lightning weather environment, and description of specific speed and signal when bug occurs, and problem description, problem occurrence event and associated log file are reported at the same time.

[0003] If it is tested in a simulation environment, the intelligent driving system of the vehicle runs on the development board of the simulation device, and the development board is directly connected with the automatic test tool in the simulation environment, and the test tool can directly capture various signals. However, once it enters the real vehicle road test stage, the vehicle needs to run in the actual road scene, such as intelligent auxiliary driving development, test and other real vehicle debugging stages, which need to be tested in real vehicle. At this time, the real driving room environment in the vehicle does not allow to connect the mainboard of the vehicle with a large number of data lines as in the development stage.

[0004] The traditional method for recording test issues involves test engineers accompanying the vehicle manually recording the time, data, and symptoms of problems or bugs. After testing, based on the manually recorded issues and timestamps, vehicle sensor data or logs from the time the issue occurred are extracted as data for problem analysis. This data is then fed into the fault management platform, where developers perform subsequent work such as problem analysis and workflow management.

[0005] The current method of discovering problems and injecting them into the fault management platform requires test engineers to follow the vehicle to record issues, which wastes human resources and increases the safety risks for test personnel. Most importantly, manually injecting test issues into the platform results in long analysis cycles, preventing a quick transition to the problem analysis phase and causing inefficiency. Furthermore, the high probability of errors in manual recording and importing of issues leads to repeated confirmations, further contributing to overall inefficiency. Summary of the Invention

[0006] To address the issues of low efficiency and error-proneness in manually recording problems discovered during real-vehicle road testing in existing technologies, this invention provides a testing method for intelligent assisted driving testing. This method can automatically capture relevant data of problems discovered during testing and generate problem records, effectively improving the efficiency of bug reporting during real-vehicle road testing and reducing the probability of errors in the reporting process.

[0007] The present invention provides a testing method for intelligent assisted driving testing, characterized by comprising the following steps:

[0008] S1: Construct the input layer to receive test signals;

[0009] The input layer includes: a signal access module, a signal quality dynamic filtering unit, and a scene knowledge graph;

[0010] The signal access module interfaces with the vehicle bus and collects test-related signal data in real time, sending it to the signal quality dynamic filtering unit. The test-related signal data includes related signals and abnormal signals, with abnormal signals including fault codes and system alarms. The related signals are time-series data, denoted as X. sig ∈R T×D Where T is the time step and D is the signal dimension;

[0011] The Scenario Knowledge Graph (SKG) describes the relationships between test tasks, test scenarios, and exception types in a test problem.

[0012] The scene knowledge graph is recorded based on a directed graph G=(V,E);

[0013] Node V={vtask v scene v abnorm}, wherein v task is a test task, v scene is a test scenario, and v abnorm is an exception type;

[0014] The test task is a test content performed by a preset vehicle under test; the test scenario describes a precondition for the test task to be carried out; and the exception type describes a type of defect occurring in the test and needing to be reported;

[0015] Each node includes attribute information, which includes a signal stability threshold θ skg ;

[0016] The edge E is a correlation weight, representing the correlation strength between two nodes connected by the edge, and the value of the edge E is in the range of 0-1;

[0017] The signal quality dynamic filtering unit detects and dynamically filters sensor noise and transient interference signals through a sliding window variance; when filtering, the signal stability threshold θ skg is used to realize filtering operation on each signal; and the signal quality filtering unit outputs the filtered correlation signal and sends it to the encoder;

[0018] S2: constructing an encoder for extracting features from an input signal;

[0019] The encoder performs lightweight timing attention and parallel computing, specifically including the following operations:

[0020] For the correlation signal with high frequency variation, local timing features are extracted by using a depth separable convolution; for the low-frequency but key scene signal, long-range dependence is captured by using local attention; during feature extraction, numerical precision is compressed and adjusted;

[0021] After the encoder extracts features from the input signal, the features are fused and output as fused features F, which are sent to the decoder;

[0022] S3: constructing a decoder to generate a test question description sentence;

[0023] The decoder is internally provided with a rule base and a text converter;

[0024] The text converter analyzes the input fused features F and outputs a structured question description sentence; the fields in the question description sentence include time t, slice log path p, priority pr, type type, and description d;

[0025] The rule base defines the relationship between the priority pr and the type type, and judges the compliance of the relationship between the parsed pr and type based on a compliance function valid(pr, type);

[0026] For the problem description sentence of the compliance of the relationship between pr and type, the decoder synchronizes the problem description sentence to the fault management platform;

[0027] For data that is not compliant with the relationship between pr and type, update the problem description sentence after backtracking correction, and use the updated problem description sentence to participate in subsequent calculations;

[0028] S4: Construct a test problem description extraction model;

[0029] The test problem description extraction model comprises an input layer, an encoder and a decoder connected in sequence;

[0030] S5: Before the test starts, connect the signal access module of the input layer in the test problem description extraction model to the vehicle bus of the vehicle under test;

[0031] S6: After the test starts, the vehicle under test drives in the actual road scene according to the test scene setting and performs the test task; the signal access module real-time extracts vehicle data signals and test problem associated abnormal signals;

[0032] Once an abnormal signal occurs, the signal access module filters the received signal based on the scene knowledge graph SKG, sends the filtered signal to the encoder to extract features for problem description, obtains fused features F, and sends F to the decoder;

[0033] The decoder outputs a structured problem description sentence, synchronizes the problem description sentence to the fault management platform, and completes the submission of the problem description related to the abnormal signal.

[0034] Further features thereof are:

[0035] The test problem description extraction model further comprises a closed-loop feedback layer, which is arranged between the decoder and the fault management platform; the closed-loop feedback layer and the fault management platform establish a bidirectional communication link; the decoder outputs a structured problem description sentence, which is sent to the fault management platform via the closed-loop feedback layer; the analysis result returned by the fault management platform is sent to the closed-loop feedback layer as feedback data, and is automatically labeled as a fine-tuning sample of the model in the closed-loop feedback layer; an incremental learning module is designed in the closed-loop feedback layer, and the decoder in the model is fine-tuned online with the feedback data to improve the classification accuracy of the decoder;

[0036] The feedback data of the fault management platform is denoted as: D fb ={(F i , pr i * ,type i * ,d i * )};

[0037] Wherein, i is the i-th moment, pr i * represents the modified priority fed back by the fault management platform, type i * represents the modified type fed back by the fault management platform, d i * represents the modified description fed back by the fault management platform.

[0038] The learnable parameters of the encoder are denoted as: Θ enc , and the learnable parameters of the decoder are denoted as: Θ dec .

[0039] The incremental learning loss function is set in the incremental learning module, the encoder parameters Θ enc are frozen, and only the decoder Θ dec is fine-tuned; the incremental learning loss function L fb is:

[0040] L fb =λ1L p +λ2L tp +λ3L ds ;

[0041] In the formula, λ1, λ2, and λ3 are the learnable parameter weights of the decoder.

[0042] The priority loss function L p is a cross-entropy function: L p =-pr i * ∑ i * log P (pr i │F i ,Θ dec );

[0043] In the formula, P() represents a probability function; pr i is the i-th moment output by the test question description grabbing model.

[0044] The type loss function L tp is a cross-entropy function: L tp =-type i* ∑ i * log P (type i │F i ,Θ dec );

[0045] In the formula, type i To test the problem description, capture the type value of the model output at time i.

[0046] Description of loss function L ds The cross-entropy function: L ds = -∑logP (d i * │F i ,Θ dec );

[0047] In the formula, d i To test the problem description, capture the d value output by the model at time i.

[0048] The AdamW optimizer is used to update the decoder parameters:

[0049]

[0050] In the formula, t is time t, η is the learning rate, and λ is the learning rate. w For weight decay, ▽_Θ dec For decoder parameters Θ dec The partial derivatives;

[0051] In step S1, the dynamic filtering performed in the signal quality dynamic filtering unit removes noise through sliding window variance detection, with the window size ω:

[0052] ;

[0053] ;

[0054] In the formula, σ t 2 X″ represents the variance at time step t. sig For the filtered correlated signal, X sig,t Let θ be the associated signal corresponding to time step t. skg X is a preset signal stability threshold in the scene knowledge graph SKG. sig,i Represents time series data X sig The i-th data point in the sliding window, Interp() is the mean of the signal within the sliding window; Interp() is the linear interpolation function.

[0055] The input layer further comprises an SKG dynamic expansion module, which performs real-time clustering on the current test scene based on the original static SKG through an online DBSCAN algorithm, and if no existing scene in the SKG is matched, a new scene node is automatically generated and temporarily added to the SKG atlas; specifically comprising the following steps:

[0056] a1: Real-time scene clustering: For the current scene feature s = [S weather , S road , S speed ], calculate the similarity with all existing scene nodes using DBSCAN:

[0057] ;

[0058] where v scene，i represents the i-th scene node in the SKG picture, and σ is the Gaussian kernel parameter;

[0059] a2: Find the maximum similarity max(sim), and the scene corresponding to max(sim) is recorded as Vmax;

[0060] a3: Compare max(sim) and the new scene decision threshold τ;

[0061] If the maximum similarity max(sim) < τ, a new node v scene,new = s is generated, and the weight of the edge E is initialized as e temp = 0.3;

[0062] Otherwise, the current scene feature s is classified as an existing scene Vmax;

[0063] where e temp is the temporary weight;

[0064] In step S2, a lightweight neural network architecture Mobile-Transformer model is used to construct an encoder to output fused features F ∈ R T×D , and the encoder specifically implements the following steps:

[0065] b1: The filtered associated signal X″ sig is first subjected to a channel-wise independent convolution, and a single channel of the input signal is independently convolved to extract a single-channel local timing feature X depth , achieving lightweight model:

[0066] ;

[0067] where ★ is a convolution operation, W d is a channel-wise convolution weight, and b dReLU is the activation function; k is the size of the convolution kernel;

[0068] b2: After point-by-point convolution dimension reduction, the output results of deep convolution are cross-channel fused, dimension adjustment is realized, multi-channel features are fused, and the number of channels is compressed from D to H / 2, H is the feature dimension:

[0069] ;

[0070] In the formula, X point is the time sequence feature after dimension reduction, W p is the point-by-point convolution weight, and b p is the point-by-point convolution bias;

[0071] b3: Using local attention, capturing scene and signal correlation, calculating the attention weight in the local window for scene feature s and signal feature X point ;

[0072] Q=X point W q , K=[ X point ;s embed ]W k , V=[X point ;s embed ] W v ;

[0073] In the formula, s embed ∈R T×H / 2 is the scene embedding, W q , W k , and W v ∈R H / 2×H / 2 are mapping matrices;

[0074] The correlation features Att(Q, K, V) of the scene and the signal are extracted:

[0075] ;

[0076] In the formula, M∈R T×T is a local mask;

[0077] b4: Output the fused feature F after splicing;

[0078] F=Concat(X point , Att(Q,K,V));

[0079] In step S3, the text converter specifically includes the following steps:

[0080] c1: Abnormal time step positioning:

[0081] The feature activation value act of the fusion feature F at each time step is calculated i , taking the time step with the maximum act as the abnormal core time step i that is strongly related to the abnormal signal i ; max ;

[0082] ;

[0083] In the formula, F i,k is the feature value of the kth dimension corresponding to the ith time step of F;

[0084] c2: Time stamp mapping:

[0085] The calculation formula of the time t of the abnormal signal is:

[0086] t=t sys +i max × sampling interval;

[0087] In the formula, t sys is the system time stamp, indicating the start time of the current test; the sampling interval is the time length corresponding to each time step, with the unit of seconds;

[0088] c3: Convert t to a time format required by the fault management platform through the Format() function;

[0089] c4: Generate the slice log path p;

[0090] Confirm the preset log slice time value ts in the system;

[0091] According to the time t, record the signal logs of each ts seconds before and after t to generate the slice log file associated with the abnormal signal, and generate the slice log file name according to the preset format; read the pre-stored log storage path template in the system, and together with the slice log file to form the log path p;

[0092] c5: Generate priority pr;

[0093] Based on a two-layer connection network, a priority prediction branch is constituted, the input of the priority prediction branch is the time step average feature of the fusion feature F, and the output is three probability values corresponding to pr=0, 1, 2; take the pr with the maximum probability as the predicted value;

[0094] ;

[0095] In the formula, P() represents the probability function;

[0096] c5: Generate type;

[0097] Extract the scene and abnormal association feature F scene-abnorm :

[0098] F scene-abnorm =Linear(Att(Q,K,V)), F scene-abnorm ∈R T×32 ;

[0099] where Linear is a linear layer that maps the input features to a target dimension of the base neural network operation;

[0100] a type classification branch based on a fully connected network with softmax is constructed, the input of the type classification branch is F scene-abnorm , the maximum activation feature F abnorm-max at the time step:

[0101] F abnorm-max =max T i-1 F scene-abnorm,i , F abnorm–max ∈R 32 ;

[0102] The type classification branch outputs Num probability values, Num>1, corresponding to the predicted probabilities of type=0~Num-1;

[0103] Take the type with the maximum probability as the type prediction value type;

[0104] c6: generate description d;

[0105] Map the fusion feature F to semantic labels, including:

[0106] scene label: extracted from s embed ;

[0107] signal anomaly label: extracted from X point ;

[0108] priority label: mapped from pr;

[0109] and type label: mapped from type;

[0110] Output the semantic labels in a fixed format: [scene label][signal anomaly label][priority / type label];

[0111] In step S3, the method of backtracking correction includes the following steps:

[0112] d1: logical verification based on the compliance function valid(pr, type);

[0113] ;

[0114] If valid = 1, it means that the relationship between pr and type is compliant;

[0115] If valid = 0, it means that the relationship between pr and type is not compliant, and then d2 is executed;

[0116] d2: Correct the priority by attention backtracking:

[0117] ;

[0118] In the formula, pr" is the corrected priority, α i is the attention weight of the i-th time step feature F i in the fusion feature, P(pr'|F i ) is the probability of predicting the priority as (pr'|F i ) based on the i-th feature F i , and pr' is the candidate priority;

[0119] d3: Correct the type;

[0120] If the corrected pr = 2: Force the type to be adjusted to type" = {2, 3} with high probability;

[0121] If the corrected pr = 1: Force the type to be adjusted to type" = {4};

[0122] Replace the original pr and type with the corrected pr" and type" to output;

[0123] In step S3, the problem description sentence is synchronized to the fault management platform, which specifically includes the following operations:

[0124] e1: Record the field set of the fault management platform as Φ = {φ1, φ2, …, φ5};

[0125] Wherein, φ1 = created creation time; φ2 = attachment log attachment path; φ3 = priority; φ4 = issueType problem type; φ5 = summary;

[0126] e2: Record the field output by the test problem description grabbing model as: Ψ = {ψ1, ψ2, …, ψ5};

[0127] Wherein, ψ1 = time t; ψ2 = log path p; ψ3 = priority pr; ψ4 = type type; ψ5 = description d;

[0128] e3: Construct a mapping function;

[0129] ;

[0130] In the formula, Map p is a preset mapping table of platform P, Format is a format conversion function, and Default indicates direct use;

[0131] φ j represents the jth fault management platform field, ψ k represents the kth problem description statement field.

[0132] The application provides a test method for intelligent auxiliary driving test, constructs a test problem description extraction model, does not need to use other data interfaces, directly collects signals uploaded by an intelligent vehicle in real time through a vehicle-mounted Ethernet and a CAN FD direct connection interface, triggers a test problem description reporting process once an abnormal signal is monitored, ensures that a vehicle meets safety requirements of real vehicle road test, quickly generates a structured problem description statement of a test problem included in an abnormal signal based on the test problem description extraction model, and automatically uploads the structured problem description statement to a corresponding fault management platform, so that the entire process does not need manual participation, automatically realizes extraction of related data of a problem found in test, and automatically records the problem, effectively improves execution efficiency, and reduces a probability of error in a test problem reporting process. Meanwhile, a closed-loop feedback layer is arranged in the test problem description extraction model, a "frozen encoder + fine-tuning decoder" strategy is adopted, online fine-tuning of the model is performed by using feedback data, a problem type classification accuracy is improved, and a model performance attenuation problem caused by scene drift is solved. BRIEF DESCRIPTION OF DRAWINGS

[0133] Figure 1 is a model schematic diagram of RTCL-SK-TMT in the application. DETAILED DESCRIPTION

[0134] The application is applied to an intelligent auxiliary driving test environment on the basis of a scene knowledge enhanced time sequence multi-task (Scene-Knowledge Enhanced Temporal Multi-Task Transformer, SK-TMT for short) model framework in the prior art, a closed-loop feedback layer is introduced, an RTCL-SK-TMT (Real-Time Closed-Loop Feedback - Scene Knowledge Graph - Temporal Multi-Task Model) model is constructed, and the RTCL-SK-TMT model structure is specifically as shown in Figure 1 The application applies the RTCL-SK-TMT model to the scene of intelligent auxiliary driving test to construct a test problem description extraction model.

[0135] Specifically, the application includes a test method for intelligent assisted driving test, comprising the following steps.

[0136] S1: constructing an input layer to receive test signals.

[0137] The input layer includes a signal access module, a signal quality dynamic filtering unit, and a scene knowledge graph.

[0138] The input layer accelerates real-time signal direct sampling and preprocessing, solves the problems of low efficiency and high risk in the test process, and discards the traditional "signal storage and batch processing" mode. Through the vehicle Ethernet and CAN FD direct connection interface, real-time test associated signal data is collected and sent to the signal quality dynamic filtering unit. The test associated signal data includes associated signals and abnormal signals. The associated signals are all signals associated with the test task among the signals uploaded by the vehicle to the vehicle enterprise cloud platform, such as vehicle speed, sensor data, etc. The abnormal signals include fault codes and system alarms. The associated signals are time series data, denoted as: X sig ∈R T×D , where T is the time step and D is the signal dimension. In this embodiment, the signal is input into the model in real time at a frequency of 100Hz, and the delay is controlled within 10ms. In this embodiment, the D=10-dimensional sensor signal is shown in Table 1.

[0139] Table 1: Sensor signal example

[0140]

[0141] The signal quality dynamic filtering unit automatically filters sensor noise (such as radar clutter) and transient interference signals (such as CAN bus glitches) through sliding window variance detection, and the threshold is the signal stability threshold θ skg in the scene knowledge graph, which is dynamically adjusted to avoid invalid data triggering false recording.

[0142] In step S1, the dynamic filtering in the signal quality dynamic filtering unit removes noise through sliding window variance detection, and the window size ω:

[0143] ;

[0144] ;

[0145] In the formula, σ t 2 represents the variance within the window at time step t, X sig is the filtered associated signal, X sig,t is the associated signal corresponding to time step t, and θ skgThe signal stability threshold preset in the scene knowledge graph SKG, the specific value is set in advance according to the actual situation of the vehicle, for example, the radar signal threshold in the "high-speed scene" is lower. sig,i The time series data X is represented as sig The i-th data point in the sliding window is The signal mean value in the sliding window is; Interp() is a linear interpolation function. The window size ω is dynamically adjusted according to the signal type, for example, ω=5 for radar signals and ω=3 for vehicle speed signals.

[0146] The scene knowledge graph SKG describes the relationship between test tasks, test scenarios and abnormal types in the test problem;

[0147] The scene knowledge graph is recorded in the form of a directed graph G=(V,E);

[0148] The nodes V={v task ,v scene ,v abnorm}, wherein v task is a test task, v scene is a test scenario, and v abnorm is an abnormal type;

[0149] The test task is the test content preset for the tested vehicle to perform; the test scenario describes the precondition for the test task to be carried out; the abnormal type describes the type of defect that needs to be reported in the test;

[0150] Each node includes attribute information, which includes: signal stability threshold θ skg ;

[0151] The edge E is the correlation weight, which represents the correlation strength between the two nodes connected by the edge, and the value of the edge E is in the range of 0~1.

[0152] The nodes included in this embodiment are as follows:

[0153] v_task (test task):

[0154] v_task1: high-speed NOA (navigation assisted driving);

[0155] v_task2: urban road ACC (adaptive cruise control);

[0156] v_task3: rural road LCC (lane centering control);

[0157] v_scene (scene node):

[0158] v_scene1: high-speed sunny day (s=[0,0,110]);

[0159] v_scene2: Urban Rainy Day (s=[1,1,30]);

[0160] v_scene3: Rural Overcast Day (s=[3,2,50]);

[0161] v_abnorm (Abnormal Type):

[0162] v_abnorm1: Positioning Drift (INS Error Too Large);

[0163] v_abnorm2: Radar Occlusion (Distance Signal Jump);

[0164] v_abnorm3: Control Delay (Steering / Brake Response Slow);

[0165] v_abnorm4: Visual Failure (Camera Lane Recognition Error);

[0166] v_abnorm5: Environmental Interference (Light Too Strong / Too Weak);

[0167] Edge Weight (E, Correlation Strength 0~1):

[0168] Because the scene knowledge graph is a directed graph, the edge from node v_task1 to node v_abnorm1 is represented by v_task1→v_abnorm1; The specific information is as follows:

[0169] v_task1 (Highway NOA)→v_abnorm1 (Positioning Drift): e=0.8 (Highway requires high positioning accuracy);

[0170] v_task1→v_abnorm2 (Radar Occlusion): e=0.7 (Highway relies on forward radar);

[0171] v_task2 (Urban ACC)→v_abnorm3 (Control Delay): e=0.6 (Urban car-following requires fast response);

[0172] v_task3→v_abnorm5 (Environmental Interference): e=0.5 (Urban light / occlusion is complex);

[0173] v_task2 (Rural LCC)→v_abnorm4 (Visual Failure): e=0.7 (Rural lane lines are not clear, relying on vision).

[0174] In this method, when determining abnormal signals, the signal value is not considered in isolation, but combined with the signal stability threshold θ skg and the scene knowledge graph SKG preset rules to filter out potential abnormal signals and confirm whether to trigger recording.

[0175] Firstly, based on dynamic noise filtering, invalid interference is removed, and suspected abnormalities are locked. The model judges whether each 10-dimensional real-time signal exceeds the scene-specific stability threshold θ through sliding window variance detection skg t 2 Compare σ skg with θ t 2 If σ skg > θ 55 , the signal is marked as a “suspected abnormal signal”; if it does not exceed the threshold, it is determined to be a “normal signal fluctuation” and does not trigger the subsequent process. For example: in the high-speed NOA scene, the forward radar distance jumps to 200m at t=55 time step, and the window variance σ 2 =428>5, triggering the “suspected abnormal” label.

[0176] However, only the “suspected abnormal” label is not enough to trigger the record, and further verification is required through the “task-scene-exception association rule” of SKG. SKG defines the association weight e of different test tasks such as high-speed NOA, urban ACC, and abnormal types; the edge e takes value 0~1, and the higher the weight, the more attention is needed. For example: high-speed NOA (v_task1) and “position drift (v_abnorm1)” have an association weight e=0.8 (high attention); urban ACC (v_task2) and “environmental interference (v_abnorm5)” have an association weight e=0.5 (medium attention). Verification logic: if the abnormal type corresponding to the “suspected abnormal signal” (such as radar jump corresponding to v_abnorm2) is associated with the current test task with an association weight e≥0.3 (0.3 is the default threshold of the temporary scene), it is confirmed that the abnormality is “recorded abnormality”, triggering the generation of subsequent t, p, pr, type, d; if the weight is <0.3, such as radar jump in the rural LCC scene, the association weight is not defined, and the subsequent problem record process is not triggered.

[0177] The signal quality dynamic filtering unit detects and dynamically filters sensor noise and transient interference signals through sliding window variance; the filtering is based on the signal stability threshold θ defined in the scene knowledge graph SKG skg to realize filtering operation on each signal; the signal quality filtering unit outputs the filtered associated signal and sends it to the encoder.

[0178] When the model is constructed, the technical personnel presets the test scene in the SKG according to the test task, but the vehicle will encounter situations beyond the preset scene in the actual test process, and once a bug is encountered in a scene beyond the preset, it also needs to be reported. In order to make the model more practical, the input layer also includes: a SKG dynamic expansion module, which performs real-time clustering on the current test scene based on the original static SKG through an online DBSCAN algorithm, and if no matching scene in the SKG is found, a new scene node is automatically generated and temporarily added to the SKG atlas; solve the record missing problem caused by "new scene missing identification". Specifically, the following steps are included.

[0179] a1: Real-time scene clustering: For the current scene feature s = [s weather , s road , s speed ], calculate the similarity with all existing scene nodes using DBSCAN. In this embodiment, the similarity is calculated based on Gaussian similarity:

[0180] ;

[0181] where v scene，i represents the i-th scene node in the SKG picture, and sigma is the Gaussian kernel parameter.

[0182] a2: Find the maximum similarity max(sim), and the scene corresponding to max(sim) is recorded as Vmax.

[0183] a3: Compare max(sim) and the new scene judgment threshold tau; The specific threshold tau is set according to the vehicle test accuracy, and in this embodiment, tau = 0.6;

[0184] If the maximum similarity max(sim) < tau, a new node v scene,new = s is generated, and the weight of the edge E is initialized to e temp = 0.3.

[0185] Otherwise, the current scene feature s is classified as an existing scene Vmax. e temp is a temporary weight, and the initial value of the weight is 0.3.

[0186] When the maximum similarity max(sim) < tau, a new scene node is generated and temporarily added to the SKG atlas. The reason for temporary addition is that in the later stage, the technical personnel needs to manually confirm whether this new scene is typical or not. If it is a typical scene with a high probability, rather than a scene with a very small probability, it will be permanently added to the SKG and considered for execution in other test tasks. If it is a scene with a very small probability, it will not be added to the SKG permanently.

[0187] Suppose the current scene feature s = [sweather , s road , s speed Encoding rules are as follows:

[0188] s weather (Weather): 0 = sunny, 1 = rainy, 2 = snowy, 3 = cloudy;

[0189] s road (Road type): 0 = highway, 1 = urban trunk road, 2 = rural road, 3 = construction section;

[0190] s speed (Average speed): Directly take the average of the last 10 seconds (km / h, such as 110, 30, etc.);

[0191] New scene judgment threshold τ = 0.6;

[0192] Calculate the Gaussian similarity (σ = 10) of the current scene s and all v scene in SKG, take the maximum value max(sim): if max(sim) ≥ 0.6: belong to the existing scene (such as s = [0, 0, 105] and v_scene1 = [0, 0, 110] sim ≈ 0.88 ≥ 0.6, judged as sunny highway). If max(sim) < 0.6: generate a new node (such as s = [2, 3, 40] (snowy construction road, 40 km / h), and the sim of all existing scenes is less than 0.6, add v_scene4 = [2, 3, 40], temporary edge weight e temp = 0.3).

[0193] S2: Build an encoder to extract features from the input signal.

[0194] Lightweight temporal attention and parallel computing are performed in the encoder to solve real-time inference problems on edge devices, and a hybrid structure of depth separable convolution combined with local attention is adopted. Specifically, the following operations are included:

[0195] For high-frequency changing signals such as steering wheel angle, use depth separable convolution to extract local temporal features, which can effectively reduce the computational complexity;

[0196] For low-frequency but key scene signals such as test task switching or weather mutation, use local attention to capture long-range dependencies, such as window size = 5 time steps, which balances efficiency and relevance;

[0197] In the feature extraction process, the numerical precision is compressed and adjusted to reduce the subsequent calculation amount. In this embodiment, the weight is quantized from 32-bit floating point to 16-bit integer, and the inference speed is increased by 2 times, which meets the real-time requirements of the vehicle-mounted edge (NVIDIA Jetson AGX), such as single sample inference ≤50ms.

[0198] The encoder fuses the features extracted from the input signal and outputs the fused features F.

[0199] In step S2, a lightweight neural network architecture Mobile-Transformer model is used to construct the encoder, and the fused features F ∈ R T×D with a dimension of 128 are output.

[0200] b1: Use deep separable convolution to extract local time sequence features;

[0201] First, the filtered associated signal X" sig First, the single channel local time sequence feature X depth is extracted by independently convolving each channel of the input signal, which realizes the lightweight of the model:

[0202] ;

[0203] In the formula, ★ represents convolution operation; W d is the channel-wise convolution weight, which does not need to be manually set and is automatically updated iteratively; b d is the channel-wise convolution bias, and ReLU is the activation function; k is the convolution kernel size, and in this embodiment, k = 3.

[0204] b2: Then, based on 1x1 convolution, the output result of the deep convolution is reduced in dimension by point-wise convolution, which realizes cross-channel fusion and dimension adjustment, and fuses multi-channel features, such as combining the features of vehicle speed and radar distance; and the number of channels is compressed from D to H / 2, which reduces the dimension for the subsequent attention module, realizes multi-channel feature fusion and dimension compression, and H is the feature dimension, which is 128 in this embodiment:

[0205] ;

[0206] In the formula, X point is the reduced time sequence feature; W p is the point-wise convolution weight, which does not need to be manually set and is automatically updated iteratively; b p is the point-wise convolution bias.

[0207] b3: Use local attention to capture scene and signal correlation, and capture scene features s and signal features X point, compute the attention weight within the local window, in this embodiment, the window size k att = 5;

[0208] Q = X point W q , K = [X point ; s embed ]W k , V = [X point ; s embed ]W v ;

[0209] where s embed ∈ R T×H / 2 is the scene embedding, W q , W k , W v ∈ R H / 2×H / 2 are the mapping matrices.

[0210] The correlation feature Att(Q, K, V) of the scene and the signal is extracted:

[0211] ;

[0212] where M ∈ R T×T is a local mask.

[0213] b4: output the fused feature F after splicing;

[0214] F = Concat(X point , Att(Q, K, V)).

[0215] X point has a dimension of T × (H / 2) (T = 100 time steps, H = 128, H / 2 = 64), and is a 10-dimensional original sensor signal feature after cross-channel fusion, such as the correlation feature of “radar distance and brake pressure”.

[0216] s = [s weather , s road , s speed ], such as a high-speed sunny scene s = [0, 0, 110], where 0 = sunny, 0 = expressway, and 110 = average vehicle speed km / h.

[0217] s embed encodes the low-dimensional scene feature s into a high-dimensional vector aligned with the dimension of the signal feature X point ; the scene is “static background information” (such as high-speed and urban), and the signal is “dynamic time series data” (such as radar jump), S embed injects scene information into each time step through “copy + encoding”, so that the model can combine the scene to judge the importance of the signal at each moment.

[0218] For example, in the embodiment of high-speed NOA scene: the original scene s = [0, 0, 110] is first encoded into a 64-dimensional vector by the fully connected layer, like s embed = [0.2, 0.1,..., 0.3], a total of 64 values; because X point has 100 time steps, the 64-dimensional scene vector is copied 100 times to obtain S embed of 100x64, ensuring that the signal features at each time step can interact with the scene features.

[0219] The mapping matrix W q , W k , and W v achieve the attention adaptation conversion of the features; all three matrices are learnable weight matrices, and in this embodiment, the dimensions are unified as W q , W k , and Wv∈R 64×64 , which is to convert the original features (X point , s embed ) into query Q, key K, and value V vectors suitable for attention calculation.

[0220] For example, in the embodiment of high-speed NOA scene:

[0221] The radar feature vector at a time step in X point is: [0.5, 0.3,..., 0.4] (100x64); after conversion by W q (64x64 matrix), the query vector Q i at that time step is obtained, like [0.32, 0.18,..., 0.25], which is used for subsequent matching of scenes and signals.

[0222] The meanings of the query Q, key K, and value V of attention in the application scenario of the present application are shown in Table 2 below.

[0223] Table 2: Corresponding relationship of QKV in the application scenario of the present application

[0224]

[0225] Regarding the dimensions, [X point ; s embed ] is a column concatenation operation, and after X point (100x64) and s embed (100x64) are concatenated, the dimension is 100x128;

[0226] After conversion by W k (64x64), the final dimensions of K and V are both 100x64, which are aligned with Q (100x64).

[0227] For example, in the embodiment of high-speed NOA scenario:

[0228] Generation of Q: X at t=55 time step point (Radar feature) x W q Corresponding to Q 55 (representing "scene information that the current radar signal needs to match");

[0229] Generation of K: [X point ; s embed ] at t=55 time step (radar feature + high-speed scene feature) x W k Corresponding to K 55 (representing "label of the current radar in the high-speed scene");

[0230] Generation of V: [X point ; s embed ] at t=55 time step (radar feature + high-speed scene feature) x W v Corresponding to V 55 (representing "core risk information of the current radar in the high-speed scene").

[0231] Local mask M as the range limiter of attention, which is a binary matrix of M ∈ R T×T , only in the position M ij= 1 within the local window, and the rest of the position M ij = 0, and the window size k att = 5 in the formula, that is, only 2 time steps before and after each time step are concerned, a total of 5 time points. By using the local mask M, the model is prevented from paying attention to irrelevant time steps at a long distance (such as the radar signal at t=55 not needing to pay attention to the vehicle speed signal at t=10), thereby reducing the amount of calculation and focusing on short-term time sequence correlation, so that the calculation result meets the local timeliness of the vehicle-mounted signal, such as the radar anomaly usually lasting for 5-10 ms.

[0232] In the embodiment of high-speed NOA scenario: T takes the value of 100, t=55 time step; window size k att = 5, so the attention range of t=55 is t=53, 54, 55, 56, 57 (a total of 5 time steps); in the mask matrix M, only M[55, 53]=M[55, 54]=M[55, 55]=M[55, 56]=M[55, 57]=1, and the rest of the positions (such as M[55, 10], M[55, 90]) are all 0; in subsequent calculation, only the 5 values within the window are retained in the matching result of Q and K, and the rest are set to zero by the mask.

[0233] The application calculates the attention weight through softmax normalization. The role is to convert the matching score (original similarity) of Q and K into a normalized weight between 0 and 1, ensuring that the sum of the weights is 1. Softmax normalizes the scores within the window, such as: the scores [2.1, 1.8, 3.5, 2.0, 1.5] within the t=55 window are normalized to obtain the weights [0.2, 0.15, 0.4, 0.18, 0.07].

[0234] The extracted scene and signal correlation feature Att(Q, K, V) is obtained. In the calculation formula of Att(Q, K, V), the normalized attention weight is multiplied by V (core information) to obtain a "feature that integrates scene and signal correlation", the dimension is consistent with V (100x64); Time steps with high weights V will be highlighted, and time steps with low weights will be weakened, and finally the Att feature focuses on the key information of the signal that is strongly related to the scene, such as radar abnormal signals in high-speed scenes.

[0235] In the embodiment of the high-speed NOA scene:

[0236] The core information (risk level) of V within the window: t=53 corresponds to 0.3 (low risk), t=54 corresponds to 0.4 (medium risk), t=55 corresponds to 0.9 (high risk), t=56 corresponds to 0.5 (medium risk), and t=57 corresponds to 0.2 (low risk);

[0237] Att 55 = 0.2x0.3 + 0.15x0.4 + 0.4x0.9 + 0.18x0.5 + 0.07x0.2 = 0.06 + 0.06 + 0.36 + 0.09 + 0.014 = 0.584;

[0238] The final Att 55 =0.584 (high-risk feature), accurately capturing the risk correlation of the t=55 radar anomaly in the high-speed scene.

[0239] The fused feature F = Concat (X point , Att) obtained in the embodiment of the high-speed NOA scene is: 100x(64+64)=100x128, which retains the "original signal feature" and "scene and signal correlation feature" at the same time, providing a basis for subsequent decoder anomaly classification.

[0240] In this application, the local attention module uses the "query - match - extract" logic of Q, K, and V, combined with s embedThe scene injection and the range limitation of M achieve two key goals: association capture: let the model focus on different signals in different scenes (such as radar in high speed and camera in city), avoid "scene-independent misjudgment" (such as the risk of radar jump in city is lower than in high speed); light weight: the window size k_att=5 only focuses on short-term time steps, and the calculation amount is reduced by about 20 times (from 100 to 5) compared with full attention, which meets the real-time inference requirements of vehicle edge (NVIDIA Jetson AGX) (single sample ≤50ms).

[0241] S3: Construct a decoder to generate a test question description sentence.

[0242] The rule base and the text converter are set inside the decoder;

[0243] The text converter analyzes the fused feature F of the input and outputs a structured question description sentence; the fields in the question description sentence include: time t, slice log path p, priority pr, type type and description d;

[0244] The rule base defines the relationship between the priority pr and the type type, and judges the compliance of the relationship between pr and type based on the compliance function valid(pr, type). For example, "high priority problems must be associated with safety-related types, such as AEB failure corresponding to control class and high priority", and "environmental problems, such as heavy rain blocking the camera, have a default medium priority".

[0245] For the question description sentence whose relationship between pr and type is compliant, the decoder synchronizes the question description sentence to the fault management platform; for the data whose relationship between pr and type is not compliant, the question description sentence is updated after backtracking and correction, and the updated question description sentence is used for subsequent calculation to ensure the logical consistency of the output. For example: in the converted question description sentence, pr is "high priority" and type is "log redundancy", the priority is automatically reduced to "low" and the description is adjusted.

[0246] In the decoder, the interface protocol template of the mainstream fault management platform, such as JIRA, Azure DevOps, etc., is stored, which contains field mapping relationship. In this embodiment, JIRA platform is used, and the field mapping relationship is as follows: JIRA "Summary" corresponds to the "problem description" output by the model. After inputting the platform type, the structured text (time, log and priority, etc.) is automatically converted into the JSON / XML format required by the platform, supporting one-key upload API call, without manual format conversion.

[0247] In step S3, the text converter specifically includes the following steps:

[0248] c1: Abnormal time step positioning:

[0249] The time steps (such as radar jump, brake pressure surge) that are strongly related to the "abnormal signal" in the fusion feature F will have high feature values due to the attention weight α i Amplify the impact of key signals. The decoder calculates the feature activation value act of the fusion feature F at each time step i ;

[0250] ,

[0251] In the formula, F i,k is the feature value of the kth dimension corresponding to the ith time step of F. act i is the average activation value of the ith time step, and the higher the value represents the more significant the signal anomaly. Take the time step with the maximum act i as the abnormal core time step i max that is strongly related to the abnormal signal.

[0252] c2: Time stamp mapping:

[0253] The calculation formula of the time t of the abnormal signal is:

[0254] t=t sys +i max × sampling interval;

[0255] In the formula, t sys is the system timestamp, indicating the start time of this round of test; the sampling interval is the time length corresponding to each time step, with the unit of second.

[0256] In this embodiment, the signal real-time access module of the input layer collects signals at a frequency of 100Hz, and each time step corresponds to a sampling interval of 10ms=0.01s, and the system timestamp t sys =2025-11-10 10:00:00.000 is recorded synchronously during collection.

[0257] c3: Convert t to a time format required by the fault management platform through the Format() function. In this embodiment, the decoder converts t to the ISO8601 standard format (such as "2025-11-04T10:00:00.550Z") through the Format() function, which meets the timestamp requirements of the fault management platform JIRA.

[0258] c4: Generate the slice log path p;

[0259] Confirm the preset log slice time value ts in the system;

[0260] According to the time t, record the signal log of each ts second before and after t to generate the slice log file associated with the abnormal signal, and generate the slice log file name according to the preset format; read the pre-stored pre-log storage path template in the system, and constitute the log path p together with the slice log file. The sensor raw data slice for positioning the abnormal period is used to facilitate subsequent backtracking analysis. In the embodiment, ts=5 seconds, and the format of the pre-stored log storage path template is: p= / fixed log storage location / abnormal core time_10s.slice, and the fixed log storage location is the fixed folder location for storing logs of each vehicle.

[0261] c5: prediction generation priority pr; the decoder generates the prediction through "priority prediction" and "logical correction", and the core is to judge the emergency degree based on the "risk association feature" in F.

[0262] The priority prediction branch is composed of a two-layer connection network, the input of the priority prediction branch is the time step average feature of the fusion feature F, and the output is three probability values corresponding to pr=0, 1, 2 (0=low, 1=medium, 2=high) probability; the maximum pr is taken as the prediction value;

[0263] ;

[0264] In the formula, P() represents a probability function. For example, if P(pr=2|F)=0.8 is the maximum, pr=2 is taken as the prediction value.

[0265] c5: generation type type;

[0266] The Att(Q, K, V) part in the fusion feature F contains scene embedding s embed For example: high-speed scene and urban scene, the decoder needs to extract scene and abnormal association feature F scene-abnorm :

[0267] F scene-abnorm = Linear(Att(Q, K, V)), F scene-abnorm ∈R T×32 ;

[0268] In the formula, Linear is a linear layer, which is a basic neural network operation that maps input features to target dimensions through weighted summation and bias, and the core function is to realize feature dimension conversion, focus on key semantic information, and adapt the feature format for subsequent attention calculation or classification tasks. It is the core bridge connecting the original feature and the task required feature. The linear layer focuses on the association dimension of the scene and the anomaly, such as radar anomaly in high-speed scene belonging to perception type and brake anomaly belonging to control type.

[0269] Predict initial type: build a type classification branch based on a softmax-based fully connected network, the input of the type classification branch is F scene-abnorm , the maximum activation feature of the time step abnorm–max , capturing the most significant abnormal class features:

[0270] F abnorm -max = max T i-1 F scene-abnorm,i , F abnorm–max ∈R 32 ;

[0271] The type classification branch outputs Num probability values, Num>1, corresponding to the prediction probability of type=0~Num-1; in this embodiment, Num=5;

[0272] Take the type with the maximum probability as the type prediction value type. For example: if (P(type=2|F)=0.75) is the maximum, then type=2.

[0273] c6: generate description d;

[0274] The description d is a structured text, and the generation process of the description d is a process of converting the feature semantics of F into natural language understandable by humans.

[0275] Mapping the fusion feature F to semantic labels includes:

[0276] Scene label: extracted from s embed , such as high-speed sunny scene (s=[0,0,110]);

[0277] Signal anomaly label: extracted from X point , such as "forward radar distance jumps to 200m" "brake pressure maintains 0bar";

[0278] Priority label: mapped from pr, such as high priority;

[0279] And type label: mapped from type, such as control class anomaly;

[0280] The semantic labels are output in a fixed format: [scene label][signal anomaly label][priority / type label].

[0281] The output of the decoder includes: time t, slice log path p, priority pr, type type, description d (fault description), and the core formula focuses on priority and type logical constraints and structured generation.

[0282] In this embodiment, the priority pr∈{0,1,2}, where 0=low, indicating no impact on safety, such as illumination intensity fluctuation; 1=medium, indicating attention needed, such as radar occasionally jumping; 2=high, indicating emergency handling, such as abnormal brake pressure;

[0283] type∈{0,1,2,3,4}, where 0=sensing, such as radar or camera signal abnormality; 1=decision, such as path planning error; 2=control, such as steering wheel or brake response abnormality; 3=hardware, such as sensor hardware failure; 4=environment, such as heavy rain causing signal interference.

[0284] Description d: needs to include scene, signal abnormality details, such as: "high-speed sunny scene (s=[0,0,110]), brake pressure (dimension 4) suddenly increased to 15 bar at t=100s, window variance σ²=12>θ skg =6 (high priority control class abnormality)".

[0285] The decoder analyzes the obtained priority and needs to meet the rules defined in the scene knowledge graph SKG with the abnormality type type, so in order to improve the accuracy of the results, the application also sets a method of logical verification and backtracking correction. Specifically, in step S3, the backtracking correction method includes the following steps.

[0286] d1: logical verification based on compliance function valid(pr,type);

[0287] ;

[0288] If valid=1, it means that the relationship between pr and type is compliant;

[0289] If valid=0, it means that the relationship between pr and type is not compliant, then d2 is executed.

[0290] d2: correct the priority through attention backtracking:

[0291] ;

[0292] In the formula, pr" is the corrected priority. α i is the attention weight of the i-th time step feature F i in the fusion feature, α i reflects the importance of F i to the current fault, the higher the weight, the greater the influence of F i , according to different scenes, different signals are focused on; that is, the weight obtained in the Softmax normalization.

[0293] P(pr'|F i ) is the probability of the i-th feature F iPredicted priority is (pr' | F i ) probability, such as F i Show steering wheel signal jump, P (pr' | F i ) probability is high.

[0294] pr' is the candidate priority (0, 1, 2).

[0295] Assume that pr=2, type=0, which corresponds to the combination of high priority and perception type, which does not meet the definition in SKG and belongs to violation.

[0296] d3: correct type;

[0297] If pr=2 after correction: force type to adjust to type"={2 control type, 3 hardware type} the category with high probability;

[0298] If pr=1 after correction: force type to adjust to type"={4 environment type};

[0299] Replace the original pr and type with the corrected pr" and type" to output.

[0300] In step S3, the problem description statement is synchronized to the fault management platform, which includes the following operations:

[0301] e1: Let the fault management platform field set be Φ={φ1,φ2,…,φ5};

[0302] Among them, φ1=created creation time; φ2=attachment log attachment path; φ3=priority priority; φ4=issueType problem type; φ5=summary problem summary.

[0303] e2: The field output by the test problem description grabbing model is recorded as: Ψ={ψ1,ψ2,…,ψ5};

[0304] Among them, ψ1= time t; ψ2= log path p; ψ3= priority pr; ψ4= type type; ψ5= description d;

[0305] In this embodiment, for example: ψ1=2025-11-04T10:00:00Z, ψ2 is " / logs / segment_100.csv", ψ3=2, ψ4=2, and ψ5 is: "High-speed sunny scene (s=[0,0,110]), brake pressure (dimension 4) increases to 15bar at t=100s, window variance σ²=12>θ skg =6 (high priority control type exception)".

[0306] e3: Build a mapping function;

[0307] ;

[0308] In the formula, Map p is the preset mapping table of the platform P, Format is the format conversion function, such as the timestamp format function for converting the timestamp into the ISO8601 format; Default indicates direct use.

[0309] The preset mapping table in this embodiment is as follows:

[0310] ψ1 corresponds to φ1, which converts the timestamp into the ISO8601 time; ψ2 corresponds to φ2, which multiplexes the path by using Default(); ψ3 corresponds to φ3, which converts pr=2 into High in the platform; ψ4 corresponds to φ4, which converts type=2 into "Control Abnormality" in the platform; and ψ5 corresponds to φ5, which simplifies d into an abstract.

[0311] The final fused feature F is the core output of the encoder, which is spliced from two parts: the first part is the cross-channel time sequence feature X of the 10-dimensional sensor signal output by the depth separable convolution point , such as the correlation trend of radar distance and brake pressure; and the second part is the correlation feature Att(Q, K, V) of the fusion scene embedding output by the local attention, such as the risk weight of the radar abnormality in the high-speed scene.

[0312] In this embodiment, the structured text output by the model is constructed in a JSON file corresponding to the JIRA platform. The specific JSON file is as follows:

[0313] {"timestamp": "2024-11-03T16:42:33.789Z", / / Time (ISO8601 format, adapted to the timestamp requirements of the platform)

[0314] "log_slice_path": " / vehicle_logs / 20241103 / LCC_164233_10s.slice", / / Slice log path

[0315] "priority": "High", / / Priority (consistent with the High / Medium / Low specification of JIRA)

[0316] "issue_type": "Control", / / Issue type (consistent with the platform classification standard)

[0317] "description": "Urban LCC test: Vehicle deviated from lane by 1.2mwhen passing a large truck, with steering angle response delayed by0.8s(verified by both camera and radar)", / / Problem description (contains multi-sensor verification information)

[0318] "platform_meta": {"project_key": "ADAS-TEST", "assignee": "control_eng_team"} / / Platform metadata (automatically assigned to the responsible person).

[0319] If the test problem description grabbing model submits a problem description with errors or mistakes, the description of the error in the fault management platform will be modified during the analysis of the bug by the R&D personnel; and in order to improve the accuracy of the model, the test problem description grabbing model in the present application is also provided with a closed-loop feedback layer, which is set between the decoder and the fault management platform; the closed-loop feedback layer and the fault management platform establish a bidirectional communication link. The forward link is that the structured problem description sentence output by the decoder is sent to the fault management platform through the closed-loop feedback layer; the reverse link is that the analysis result returned by the fault management platform is sent to the closed-loop feedback layer as feedback data, which is automatically labeled as a fine-tuning sample of the model in the closed-loop feedback layer, and an incremental learning module is designed in the closed-loop feedback layer to fine-tune the decoder in the model online with the feedback data, so as to improve the classification accuracy of the decoder.

[0320] The feedback data of the fault management platform is denoted as: D fb {(F i ,pr i * ,type i * ,d i * )};

[0321] Where i is the i-th moment, pr i * represents the priority of the feedback of the fault management platform, type i * represents the type of the feedback of the fault management platform, d i * represents the description of the feedback of the fault management platform;

[0322] The learnable parameters of the encoder are denoted as: Θenc , the learnable parameters of the decoder are denoted as: Θ dec ;

[0323] The incremental learning loss function is set in the incremental learning module, and the encoder parameters Θ enc are frozen, and only the decoder Θ dec is fine-tuned; the incremental learning loss function L fb is:

[0324] L fb =λ1L p +λ2L tp +λ3L ds ;

[0325] In the formula, λ1, λ2, and λ3 are the learnable parameter weights of the decoder, which are used to focus on the description accuracy; in the embodiment, λ1=0.3, λ2=0.2, and λ3=0.5.

[0326] The priority loss function L p is a cross-entropy function: L p =-pr i * ∑ i * logP(pr i │F i ,Θ dec );

[0327] In the formula, P() represents a probability function;

[0328] The type loss function L tp is a cross-entropy function: L tp =-type i * ∑ i * logP(type i │F i ,Θ dec );

[0329] In the formula, type i is the type value output by the test question description grabbing model at the i-th moment;

[0330] The description loss function L ds is a cross-entropy function with a term constraint penalty: L ds =-∑logP(d i * │F i ,Θ dec );

[0331] The AdamW optimizer is used to update the decoder parameters:

[0332] ;

[0333] where t is the time, η = 1e-5 is the learning rate; λ = 1e-4 is the weight decay to avoid overfitting; ∇_Θ is the partial derivative of the decoder parameters Θ with respect to the loss function L, and represents the direction and rate of change of the loss function L with respect to Θ. The negative gradient direction is the direction in which the loss function L decreases most rapidly. w dec dec fb dec

[0334] S4: Constructing a test problem description grabbing model;

[0335] The test problem description grabbing model includes an input layer, an encoder, and a decoder connected in sequence.

[0336] In this method, a closed-loop process is constructed based on the problem description grabbing model, as follows.

[0337] Trigger: Abnormal signals (such as lane deviation and steering delay) are collected in real time and verified by multiple sensors to confirm their effectiveness; Upload: A problem ticket is automatically created through the JIRA API and pushed to the development team in synchronization; Analysis: The development team analyzes the results on the platform, such as: steering delay due to slow response of the ESP module; Feedback: The analysis results are fed back to the model, and the incremental learning module fine-tunes the "problem type" classifier, for example: more accurately associate the steering delay with the hardware class; Iteration: Next time the same type of problem is triggered, the model output type and description are more accurate.

[0338] For the test problem description grabbing model, public datasets are used for training and verification, and the results of optimizing effect quantification are compared. The details are as follows.

[0339] Based on public datasets, the nuScenes autonomous driving dataset and the test problem annotation subset of the Waymo OpenDataset are selected in this embodiment. Through training and verification experiments, the advantages and disadvantages of the RTCL-SK-TMT optimized model and the original SK-TMT model are compared, and the corresponding optimized effect quantification table is explained. Experimental settings: dataset selection: training set: 100,000 test scenes in nuScenes, containing sensor data, abnormal event annotations, combined with 50,000 artificially recorded problems of Waymo. Containing priority, type, description, merging into 150,000 samples, covering L2-L4 level intelligent driving test scenes. Validation set: independently extract 20,000 samples that do not participate in training, containing more than 30 test tasks and more than 100 abnormal signal types, simulating real road test scene distribution.

[0340] ​​​​​Evaluation indicators: Real-time: end-to-end processing delay (time consumption from signal input to text output); Accuracy: priority classification accuracy, problem type classification accuracy, description compliance (BLEU-4 score); Practicality: human intervention rate, i.e. the proportion of samples that need to be corrected, the success rate of platform docking, calculated as the proportion of direct import into JIRA / AzureDevOps without format errors; Robustness: high-priority problem omission rate, calculated as the proportion of real high-priority problems that are not identified by the model.

[0341] Referring to Table 3 below, where the original SK-TMT model represents the SK-TMT framework problem description extraction model, and the RTCL-SK-TMT optimization model represents the problem description extraction model constructed based on the method.

[0342] Table 3 performance comparison example

[0343]

[0344] S5: Before the test starts, the signal access module of the input layer in the test problem description extraction model is docked to the vehicle bus of the vehicle being tested.

[0345] S6: After the test starts, the vehicle being tested drives in the actual road scene according to the test scene setting and performs the test task; the signal access module extracts vehicle data signals and test problem associated abnormal signals in real time;

[0346] Once an abnormal signal occurs, the signal access module filters the received signal based on the scene knowledge graph SKG, sends the signal to the encoder to extract the features for problem description, obtains the fused features F, and then sends F to the decoder;

[0347] The decoder outputs a structured problem description sentence, which is synchronized and submitted to the fault management platform to complete the submission of the problem description related to the abnormal signal.

[0348] The core advantages of the RTCL-SK-TMT constructed in this application include: real-time, through lightweight architecture and quantization technology, to meet the millisecond-level response requirement of the edge end of the real vehicle; automation, based on term constraints and logical verification to eliminate human intervention, combined with platform protocol adaptation to realize the closed loop of uploading as soon as it occurs; robustness, multi-sensor verification and incremental learning, which greatly reduces the high-priority problem omission rate and continuously optimizes with feedback.

[0349] Universality of public data set verification: the experiment reproduces the optimization effect on the public data set, indicating that the improvement of RTCL-SK-TMT does not depend on specific private data and has industry promotion value, especially suitable for intelligent driving test scenarios that require rapid iteration.

[0350] The following is an example of AEB (automatic emergency braking) anomaly detection in a high-speed NOA scenario, illustrating a complete process of this method.

[0351] The test vehicle is driving at 110 km / h on a straight highway in sunny weather in the "high-speed NOA (navigation-assisted driving)" task, with scene characteristics s = [0, 0, 110], where s weather = 0 = sunny, s road = 0 = highway, s speed = 110 km / h, real-time collection of 10-dimensional sensor signals D = 10, including radar front vehicle distance, camera recognition distance, vehicle speed, brake signal, etc., with a sampling frequency of 100 Hz, continuous collection for 1 second, T = 100 time steps. An example of sensor correlation signals collected based on the vehicle bus is shown in Table 4 below.

[0352] Table 4: Example of correlation signals

[0353]

[0354] I. Real-time filtering and encoding of correlation signals. The original correlation signal X sig ∈ R 100×10 , the radar front vehicle distance signal at time steps 50-60 jumps: [150, 151, 152, 151, 200, 154, 153, 152, 151, 150, 155] (t = 50 to t = 60), where it suddenly increases to 200 meters (anomaly) at t = 55, while other signals such as vehicle speed and brake pressure remain stable, with vehicle speed maintaining 110 km / h and brake pressure 0 bar.

[0355] The dynamic noise filtering process is as follows.

[0356] Read the preset sliding window size ω = 5 of the radar signal, calculate the window variance σ 2 55 at time step 55: the window contains radar values [152, 151, 200, 154, 153] at t = 53-57, the mean value = 162, and the variance σ 2 55 ≈ (1 / 5) [(152-162) 2 +…+(153-162) 2 ] ≈ 428.

[0357] Read the scene knowledge graph SKG, which shows that the stability threshold θ skg = 5 for radar signals in high-speed scenarios, because σ 2 55 = 428 > 5, it is determined to be noise. Use linear interpolation to complete: take t = 50 (150 meters) and t = 60 (155 meters), and the interpolated value is X"sig,55 = 152.5 meters. The final output filtered signal X" sig ∈ R 100×50 The radar noise point at t = 55 is corrected, and other signals remain unchanged.

[0358] II. The process of determining whether real-time expansion of the scene knowledge graph SKG is needed is as follows.

[0359] The nodes already in the initial structure of SKG are: v task = "high-speed NOA", v scene = "high-speed sunny straight road s = [0, 0, 110]", v abnorm = "radar anomaly" "brake failure". Edge weight: v task → v abnorm = "brake failure" e = 0.8, indicating that the risk of brake failure is high in high-speed NOA; v task → v abnorm = "radar anomaly" e = 0.7, indicating that radar is the core perception source of AEB.

[0360] Real-time scene clustering. The current scene feature s = [0, 0, 110] is calculated with the Gaussian similarity of the existing scene node v scene,1 = [0, 0, 110] in SKG. The squared Euclidean distance is: |s-v scene,1 |2 2 = 0, the similarity sim = exp(0 / (2 / 10 2 )) = 1, which is greater than the threshold τ = 0.6, indicating that it is a known scene and no new node needs to be added, and the association rule of "high-speed sunny straight road" in SKG is followed. In this embodiment, the radar threshold θ skg = 5.

[0361] III. The following is calculated in the encoder.

[0362] Local temporal features are extracted based on depth separable convolution.

[0363] Depth convolution: X" sig (100 x 10) is convolved with a kernel W d ∈ R 3×1×10 Channel-wise convolution, output X depth ∈ R 100×10 , which fuses the local features of the adjacent 3 time steps, for example, the convolution kernel of the radar channel is [0.2, 0.5, 0.3], which fuses the corrected radar values [151, 152.5, 154] at t = 53-55 to extract local trend features.

[0364] Point-wise convolution: a 1 x 1 convolution kernel W p ∈ R 1×1×10×64; H=128, hence H / 2=64 dimension reduction, output X point ∈R 100×64 ; Cross-channel feature fusion of 10 signals, such as radar distance and brake signal correlation.

[0365] Local attention-based extraction of scene and signal correlation features.

[0366] Scene embedding s embed ∈R 100×64 ; Scene feature s=[0,0,110] is encoded into a 64-dimensional vector and copied to 100 time steps, aligned with X point . Attention Q=X point W q , W q ∈R 64×64 Query vector based only on signal features; K=[X point ;s embed ]W k , V=[X point ;s embed ]W v , W q , W k , W v ∈R 64×64 Key-value vector that fuses signals and scenes.

[0367] Local window k att =5, mask M only allows correlation of 2 steps before and after t=55, output attention feature Att(Q,K,V)∈R 100×64 ; Focus on "high-speed scene, radar anomaly, brake signal response".

[0368] Final fusion feature F=Concat(X point ,Att)∈R 100×128 , where 128 dimensions =64-dimensional local time series features +64-dimensional scene correlation features.

[0369] Four, the decoder performs multi-task logic constraint output, the specific operation is as follows.

[0370] Initial output and logic check. The decoder outputs based on F:

[0371] Time t=“2025-11-04-10:00:00”, log path p=“ / logs / 20251104 / segment_100.csv”, priority pr=2 (high), type type=0 (perception), description d=“radar signal anomaly jump”.

[0372] Compliance check: valid(pr=2, type=0)=0 (SKG rule: high priority only allows control class 2 or hardware class 3, sensing class 0 cannot be high priority).

[0373] Attention backtracking correction: feature F i In the middle, attention weight a i High time step shows "brake signal not triggered with radar anomaly" (control class feature), priority prediction probability P(pr=2│F i )=0.9 highest, corresponding feature shows "brake pressure still 0 bar when radar anomaly (AEB not triggered)", which is a control class feature. Therefore, pr=2 is corrected, and the type is corrected to type=2 (control class), at which time valid(2,2)=1 (compliant), and the corrected description: d="brake control signal not triggered with radar anomaly in high-speed NOA scenario (AEB failure risk)".

[0374] Platform protocol mapping: interface JIRA platform fields Φ={summary, priority, created}:

[0375] Description d corresponds to summary, formatted as "brake control signal not triggered with radar anomaly in high-speed NOA scenario"; priority pr=2 corresponds to priority, formatted as "High"; time t corresponds to created, formatted as "2025-11-04T10:00:00Z". p corresponds to "attachment": " / logs / 20251104 / segment_100.csv"; type=2 corresponds to "issuetype": "control anomaly".

[0376] Five, closed-loop feedback layer for incremental learning.

[0377] Suppose the development team confirms the feedback data D fb : F i is the fusion feature described above, pr i * =2, type i * =2, d i * ="brake control signal delayed by 200ms (not triggered with radar anomaly) in high-speed NOA scenario", and the developer supplements the details of "delayed by 200ms".

[0378] Loss function optimization: freeze encoder parameters Θ enc , fine-tune decoder Θ dec , loss L fb =0.3L p +0.2Ltp +0.5L ds , focus description accuracy.

[0379] Priority loss Lp: -logP(pr=2│F i )=-log(0.9)≈0.105, indicating that the original prediction is more accurate, and the loss is small;

[0380] Type loss L tp : -logP(type=2│F i )=-log(0.8)≈0.223, indicating that the type correction reduces the loss;

[0381] Description loss L ds : The original description does not mention "delay 200ms", and the penalty term makes -logP(d i * |F i )=0.5, indicating that the loss is large, which promotes the model to learn the detailed description;

[0382] Total loss L fb =0.3×0.105+0.2×0.223+0.5×0.5≈0.031+0.045+0.25=0.326.

[0383] Parameter update: use AdamW optimizer η=1e-5, λ w =1e-4 to update Θ dec , so that the accuracy of the description of "brake delay" is improved by 32.6% next time.

[0384] Through the whole process, the system accurately identifies "AEB control abnormality in high-speed NOA scene", outputs high-priority reports in JIRA format, and avoids single radar false positives through multi-sensor verification. Subsequently, through incremental learning, the model improves the accuracy of the description of "brake delay", forming a closed loop of "data collection, anomaly detection, and feedback optimization". After 3 iterations, the false positive rate of AEB anomaly detection is reduced from 5% to 1.2%.

[0385] In this embodiment, the model is fine-tuned online with feedback data, about 50-100 per day, and each fine-tuning takes ≤10 minutes, which improves the problem type classification accuracy by 1%-2% per week, solving the problem of model performance degradation caused by "scene drift".

[0386] After using the technical scheme, the test question description grabbing model based on the TCL-SK-TMT model is constructed, the model input contains three types of heterogeneous data: time sequence correlation signal, discrete test task name and event type abnormal signal, the self-attention mechanism of the transformer can naturally model the dependence relationship of cross-modal features, such as the correlation between the "high-speed AEB test" task and the "radar point cloud loss" abnormal signal, compared with the CNN which is only good at local features or the RNN whose time sequence dependence modeling ability is weaker than the attention mechanism, the model in the application has more advantages in processing complex problems. When reporting a bug, tasks such as priority classification, problem type classification and problem description generation need to be completed at the same time, the decoder of the transformer can realize multi-task joint optimization through different head branches, avoid the fragmentation of single model training, and the shared encoder features can improve the cooperativity between tasks. The intelligent driving test scene has strong field, such as the abnormality of "LCC lane centering" task is mostly related to perception / control, the flexible architecture of the transformer supports embedding the scene knowledge graph, solving the generalization problem in the small sample scene, which is not realized by traditional machine learning models such as SVM, random forest and the like. The RTCL-SK-TMT model solves the problems of real-time and multiple scenes. The end-to-end processing delay, manual intervention rate, platform docking success rate, problem type classification accuracy and high priority problem missing rate are greatly optimized.

Claims

1. A test method for intelligent assisted driving test, characterized in that, Comprising the following steps: S1: constructing an input layer to receive test signals; The input layer comprises: a signal access module, a signal quality dynamic filtering unit and a scene knowledge graph SKG; The signal access module is connected to the vehicle bus, and real-time test associated signal data is collected and sent to the signal quality dynamic filtering unit; the test associated signal data includes: associated signal and abnormal signal, the abnormal signal includes: fault code and system alarm; the associated signal is time series data, denoted as: X sig ∈R T×D , wherein T is the time step, and D is the signal dimension; The scene knowledge graph SKG describes the relationship between test tasks, test scenes and abnormal types in the test problem; The scene knowledge graph SKG is recorded based on a directed graph G=(V,E); V = {v task ,v scene ,v abnorm}, wherein v task is a test task, v scene is a test scenario, and v abnorm is an exception type; The test task is the test content performed by the preset vehicle under test; the test scene describes the precondition for the development of the test task; and the abnormal type describes the defect type that needs to be reported in the test; Each node includes attribute information including a signal stability threshold θ skg ; Edge E is an association weight, indicating the association strength between the two nodes connected by the edge, and the value of edge E is in the range of 0~1; The signal quality dynamic filtering unit detects and dynamically filters sensor noise and transient interference signals through sliding window variance; during filtering, the signal stability threshold θ defined in the scene knowledge graph SKG is used as the basis skg The filtering operation is implemented for each signal; the signal quality filtering unit outputs the filtered correlation signals and sends them to the encoder; S2: constructing an encoder for extracting features from input signals; The encoder performs lightweight timing attention and parallel computing, outputs fused features F, and sends them to the decoder; S3: constructing a decoder to generate test problem description sentences; S4: constructing a test problem description grabbing model; The test problem description grabbing model comprises: an input layer, an encoder and a decoder connected in turn; S5: Before the test starts, the signal access module of the input layer in the test problem description grabbing model is connected to the vehicle bus of the vehicle under test; S6: After the test starts, the vehicle under test drives in the actual road scene according to the test scene setting and performs the test task; the signal access module real-time grabs vehicle data signals and test problem associated abnormal signals; Once an abnormal signal occurs, the signal access module filters the received signal based on the scene knowledge graph SKG, sends it to the encoder to extract features for problem description, obtains fused features F, and then sends F to the decoder; The decoder outputs a structured problem description sentence, synchronously submits the problem description sentence to the fault management platform, and completes the submission of the problem description related to the abnormal signal; The test problem description grabbing model further comprises a closed-loop feedback layer, which is arranged between the decoder and the fault management platform; the closed-loop feedback layer and the fault management platform establish a bidirectional communication link; the structured problem description sentence output by the decoder is sent to the fault management platform through the closed-loop feedback layer; the analysis result returned by the fault management platform is sent to the closed-loop feedback layer as feedback data, and is automatically labeled as a fine-tuning sample of the model in the closed-loop feedback layer; an incremental learning module is designed in the closed-loop feedback layer, and the decoder in the model is fine-tuned online with feedback data to improve the classification accuracy of the decoder; Let the feedback data from the fault management platform be denoted as: D fb {(F i ,pr i * ,type i * ,d i * )}; where i is the i-th time instant, pr i * represents the modified priority fed back by the fault management platform, type i * represents the modified type fed back by the fault management platform, d i * represents the modified description fed back by the fault management platform; Let the learnable parameters of the encoder be denoted by: Θ enc Let the learnable parameters of the decoder be denoted by: Θ dec ; An incremental learning loss function is set in the incremental learning module, and the encoder parameter Θ is frozen enc , and only the decoder Θ is fine-tuned dec ; the incremental learning loss function L fb is: L fb =λ1L p +λ2 L tp +λ3 L ds ; In the formula, λ1, λ2, λ3 are learnable parameter weights of the decoder; Priority loss function L p is a cross-entropy function: L p =-pr i * ∑ i * log P (pr i │F i ,Θ dec ); where P() represents a probability function; pr i to test the problem description of the pr value of the i-th moment output by the model output Type loss function L tp is the cross-entropy function: L tp =-type i * ∑ i * log P (type i │F i ,Θ dec ) In the formula, type i to test the type value of the i-th moment output of the problem description capture model output; Describing the loss function L ds is the cross-entropy function: L ds = -∑logP (d i * | F i , Θ dec ); where d i is the d value output by the problem description scraping model at the i-th time instant of the test An AdamW optimizer is used to update the decoder parameters: ； where t is the time, η is the learning rate, λ is the weight decay, and w is the partial derivative of the decoder parameters Θ dec with respect to the input x. dec with respect to the input x. 2.The test method for intelligent assisted driving test of claim 1, wherein: In step S1, the dynamic filtering in the signal quality dynamic filtering unit removes noise through sliding window variance detection, and the window size ω: ； ； where σ t 2 denotes the variance at time step t, X″ sig is the filtered correlation signal, X sig,t is the correlation signal corresponding to time step t, θ skg is a preset signal stability threshold in the scene knowledge graph SKG, X sig,i denotes the time series data X sig is the i-th data point in the sliding window, is the mean value of the signal within the sliding window; Interp() is a linear interpolation function. 3.The test method for intelligent assisted driving test of claim 1, wherein: The input layer further comprises an SKG dynamic expansion module, which performs real-time clustering on the current test scene through an online DBSCAN algorithm based on the original static SKG, and automatically generates a new scene node and temporarily adds it to the SKG atlas if no matching scene in the SKG is found; specifically comprising the following steps: a1: real-time scene clustering: compute similarity to all existing scene nodes with DBSCAN for current scene features s = [S weather , S road , S speed ] ； where v scene，i denotes the i-th scene node in the SKG picture, and σ is a Gaussian kernel parameter. a2: find the maximum similarity max(sim), and the scene corresponding to max(sim) is recorded as Vmax; a3: compare max(sim) and the new scene judgment threshold τ; If the maximum similarity max(sim) < τ, a new node v is generated scene,new = s, the weight of the edge E is initialized as e temp = 0.3; Otherwise, the current scene feature s is classified as the existing scene Vmax; where e temp is a temporary weight. 4.The test method for intelligent assisted driving test of claim 1, wherein: In step S2, the decoder specifically comprises the following operations: For the high-frequency change correlation signal, the local time sequence features are extracted by using the depth separable convolution; for the low-frequency but key scene signal, the long-range dependence is captured by using the local attention; during the feature extraction process, the numerical precision is compressed and adjusted; The encoder fuses the features extracted from the input signal and outputs the fused features F, which are sent to the decoder.

5. The test method for intelligent assisted driving test according to claim 4, characterized in that: In step S2, a Mobile-Transformer model is used to construct an encoder to output a fusion feature F∈R T×D , and the following steps are implemented in the encoder: b1: filter the correlation signal X" sig First, the single channel of the input signal is independently convolved to extract the single-channel local timing feature X depth Lightweight model: ； where ★ is a convolution operation, W d is a channel-wise convolution weight, b d is a channel-wise convolution bias, ReLU is an activation function; k is a kernel size; b2: after the point-by-point convolution dimension reduction, the output results of the depth convolution are cross-channel fused, the dimension adjustment is realized, the multi-channel features are fused, and the channel number is compressed from D to H / 2, H being the feature dimension: ； In the formula, X point is the reduced dimension time sequence feature, W p is the point-by-point convolution weight, and b p is the point-by-point convolution bias. b3: using local attention, capturing scene and signal correlation, for scene feature s and signal feature X point , compute attention weights within local window; Q = X point W q , K = [X point ; s embed ] W k , V = [X point ; s embed ] W v ; where s embed ∈R T×H / 2 is a scene embedding, W q is a mapping matrix; and k is a mapping matrix; and v ∈R H / 2×H / 2 is a mapping matrix; and The correlation features Att(Q, K, V) of the scene and the signal are extracted: ； In the formula, M ∈ R T×T is a local mask; b4: output the fused features F after splicing; F = Concat(X point , Att(Q, K, V)). 6.The test method for intelligent assisted driving test of claim 1, wherein: In step S3, the decoder is internally provided with a rule library and a text converter; The text converter analyzes the input fused features F and outputs a structured problem description sentence; The fields in the problem description sentence include time t, slice log path p, priority pr, type type and description d; The rule library defines the relationship between the priority pr and the type type, and judges the compliance of the relationship between pr and type based on the compliance function valid(pr, type); For the problem description sentence with compliant relationship between pr and type, the decoder synchronizes the problem description sentence to the fault management platform; For the data with non-compliant relationship between pr and type, the problem description sentence is updated after backtracking and correction, and the updated problem description sentence is used for subsequent calculation.

7. The test method for intelligent assisted driving test according to claim 6, characterized in that: In step S3, the text converter specifically comprises the following steps: c1: abnormal time step positioning: The feature activation value act of the computed fusion feature F at each time step i , take the time step with the largest act i as the anomaly core time step i that is strongly related to the anomaly signal max ; ； In the formula, F i,k is the eigenvalue of the kth dimension corresponding to the ith time step of F c2: timestamp mapping: The calculation formula of the time t of the abnormal signal is: t = t sys + i max x sampling interval; In the formula, t sys is a system timestamp, indicating the start time of the current test; the sampling interval is the time length corresponding to each time step, in seconds; c3: convert t to the time format required by the fault management platform by using the Format() function; c4: generate the slice log path p; Confirm the preset log slice time value ts in the system; According to the time t, record the signal log of each ts second before and after t to generate the slice log file associated with the abnormal signal, and generate the slice log file name according to the preset format; read the pre-stored pre-log storage path template in the system, and together with the slice log file, constitute the log path p; c5: generate the priority pr; A priority prediction branch is constructed based on a two-layer connection network, input of the priority prediction branch is time step average features of the fusion features F, and output is three probability values corresponding to probabilities of pr=0, 1 and 2; the maximum probability pr is taken as a prediction value; ； In the formula, P() represents a probability function; c5: generating a type type; Extracting scene and anomaly associated features F scene-abnorm : F scene-abnorm = Linear(Att(Q, K, V)), F scene-abnorm ∈ R T×32 ; In the formula, Linear is a linear layer, which is a basic neural network operation for mapping input features to a target dimension; A type classification branch is constructed based on a fully connected network with softmax, the input of the type classification branch is F scene-abnorm , the maximum activation feature F abnorm -max at the time step F abnorm -max =max T i-1 F scene-abnorm,i , F abnorm –max ∈R 32 ; The type classification branch outputs Num probability values, Num>1, corresponding to prediction probabilities of type=0~Num-1; The maximum probability type is taken as a type prediction value type; c6: generating a description d; The fusion features F are mapped to semantic labels, including: Scene tag: from s embed Extraction; Signal anomaly label: from X point Extract; A priority label is mapped from pr; And a type label is mapped from type; The semantic labels are output in a fixed format: [scene label][signal anomaly label][priority / type label]. 8.The test method for intelligent assisted driving test of claim 7, wherein: In step S3, the backtracking correction method includes the following steps: d1: logical verification based on the compliance function valid(pr, type); ； If valid=1, it indicates that the relationship between pr and type is compliant; If valid=0, it indicates that the relationship between pr and type is not compliant, and then d2 is executed; d2: priority correction through attention backtracking; ； In the formula, pr' is the corrected priority, a i is the attention weight of the i-th time step feature F i in the fusion feature, P(pr'|F i ) is the probability of predicting the priority as (pr'|F i ) based on the i-th feature F i , and pr' is the candidate priority. d3: type correction; If pr=2 after correction: type is forced to be adjusted to type″={2, 3} with the higher probability; If pr=1 after correction: type is forced to be adjusted to type″={4}; The original pr and type are replaced by the corrected pr″ and type″. 9.The test method for intelligent assisted driving test of claim 7, wherein: In step S3, the problem description sentence is synchronized to the fault management platform, specifically including the following operations: e1: a fault management platform field set is denoted as Φ={φ1, φ2, …, φ5}; In which, φ1=created creation time; φ2=attachment log attachment path; φ3=priority priority; φ4=issueType problem type; φ5=summary; e2: the field output by the test problem description grabbing model is denoted as: Ψ={ψ1, ψ2, …, ψ5}; In which, ψ1=the time t; ψ2=log path p; ψ3=priority pr; ψ4=type type; ψ5=description d; e3: a mapping function is constructed; ； In the formula, Map p is a preset mapping table of the platform P, Format is a format conversion function, and Default indicates direct use. φ j represents the jth fault management platform field, ψ k represents the kth problem description statement field.

Images