Search

CN-121998064-A - Vehicle-mounted real-time causal graph incremental causal discovery device and discovery method

CN121998064ACN 121998064 ACN121998064 ACN 121998064ACN-121998064-A

Abstract

The invention discloses a vehicle-mounted real-time causal graph incremental causal discovery device and a discovery method, wherein the vehicle-mounted real-time causal graph incremental causal discovery device comprises a data flow interface layer and an incremental causal discovery engine, the data flow interface layer is used for receiving newly-introduced sensor data, the incremental causal discovery engine is used for establishing a causal graph for existing sensor data to acquire a weight matrix of the causal graph, if the newly-introduced sensor data cause Laplace disturbance, the weight matrix is updated according to the newly-introduced sensor data to acquire an updated weight matrix, and the causal graph is updated based on the updated weight matrix. The invention changes the computational complexity from the traditional Significantly reduce to The causal graph generation delay is shortened from the minute level to the millisecond level.

Inventors

  • WANG HONGXIN
  • YANG NAN
  • WANG XING
  • XU RUI
  • DING GUOLIANG

Assignees

  • 武汉江夏楚能汽车技术研发有限公司

Dates

Publication Date
20260508
Application Date
20260109

Claims (10)

  1. 1. An on-board real-time causal graph incremental causal finder, comprising: A data stream interface layer for receiving newly introduced sensor data; The incremental causal discovery engine is used for establishing a causal graph for the existing sensor data, acquiring a weight matrix of the causal graph, and if the newly introduced sensor data causes Laplace disturbance, locally updating the weight matrix according to the newly introduced sensor data to obtain an updated weight matrix, and updating the causal graph based on the updated weight matrix.
  2. 2. The on-board real-time causal graph incremental causal finder of claim 1, wherein the incremental causal finding engine comprises: the causal graph establishing unit is used for establishing a causal graph for the existing sensor data, wherein the causal graph comprises a plurality of nodes and a plurality of edges connected with every two nodes, and a weight matrix of the causal graph is obtained; A selection unit for receiving candidate edges; The Laplace spectrum decomposition unit is used for establishing a Laplace matrix for all sides, and carrying out Laplace spectrum decomposition on the Laplace matrix only once to obtain a base characteristic pair of each side, wherein the base characteristic pair comprises a characteristic value and a characteristic vector; The Laplace disturbance acquisition unit is used for traversing all candidate edges according to the difference value of the existing sensor data and the newly introduced sensor data, respectively calculating the partial correlation coefficient of each candidate edge, and acquiring local Laplace disturbance according to all the partial correlation coefficients; the incremental spectrum updating unit is used for updating the base characteristic pair of each candidate edge based on the local Laplace disturbance; The weight matrix updating unit is used for acquiring a local updating weight matrix based on the updated characteristic value and the characteristic vector, and carrying out local updating on the weight matrix based on the local updating weight matrix to obtain an updated weight matrix; and a causal graph updating unit for updating the causal graph based on the updated weight matrix.
  3. 3. The vehicle-mounted real-time causal graph incremental causal finder of claim 2, wherein the edges comprise candidate edges and non-candidate edges, wherein the computing formula for updating the base feature pair of each candidate edge based on the local laplace disturbance is: Wherein, the In the event of a local laplace disturbance, And Respectively Laplace matrix Strip candidate edge The next updated eigenvalues and eigenvectors, And Respectively Laplace matrix Strip candidate edge The next updated eigenvalues and eigenvectors, And Respectively Laplace matrix Strip non-candidate edge The next updated eigenvalues and eigenvectors.
  4. 4. The on-board real-time causal graph incremental causal finder of claim 1, further comprising: an attention pruning controller comprising: the causal attention score obtaining subunit is configured to calculate a causal attention score between each pair of nodes based on a neural network attention mechanism, where a calculation formula is as follows: Wherein, the Is the first Personal node and the first A causal attention score between the individual nodes, Is the first Personal node and the first A causal strength score between the individual nodes, Is the first to A learnable attention vector associated with each node, As a learnable attention vector, Is the first Personal node and the first Causal strength scores between individual nodes; The preset score threshold value adjusting subunit is configured to, if the causal attention score between a pair of nodes reaches a preset score threshold value, mark the pair of nodes as candidate node pairs, and obtain edges between the candidate node pairs as candidate edges, where the preset score threshold value is adaptively adjusted, and a calculation formula is as follows: Wherein, the As a smoothing factor, the smoothing factor is used, , In order to find the arithmetic mean value, Is that The preset score threshold for the number of iterations, Is that A preset score threshold for the number of iterations.
  5. 5. The on-board real-time causal graph incremental causal finder of claim 4, wherein the attention pruning controller further comprises: a bitmap mask generation subunit for marking candidate edges and generating bitmap mask The calculation formula is as follows: Wherein, the The percentages are indicated by way of example, Representing the total number of nodes.
  6. 6. The on-board real-time causal graph incremental causal finder of claim 2, further comprising: A hardware processing layer, comprising: a storage format setting subunit, configured to set a customized hybrid storage format, including: Setting a row pointer for locating the physical position of the weight matrix, wherein the physical position is represented by a multi-bit fixed-length array; setting a column index to store edges with preset length and corresponding weight values, wherein the weight values are expressed by fixed point numbers; Columns where candidate edges exist are marked with bitmaps.
  7. 7. The on-board real-time causal graph incremental causal finder of claim 4, wherein the hardware processing layer further comprises: An acceleration setting subunit, configured to set a bit operation acceleration instruction sequence, including: 4-8 weight values in a local update weight matrix are updated every cycle by utilizing a channel mask function of a single-instruction multiple-data vector unit and a packing single-instruction multiple-data function based on a RISC-V architecture; And the hardware acceleration sparse matrix manager is used for realizing zero value elimination through bit exclusive OR and bit AND operation when the weight matrix is locally updated based on the locally updated weight matrix.
  8. 8. The vehicle-mounted real-time causal graph incremental causal discovery method is characterized by comprising the following steps of: receiving newly introduced sensor data; and establishing a causal graph for the existing sensor data, acquiring a weight matrix of the causal graph, updating the weight matrix according to the newly introduced sensor data if the newly introduced sensor data causes Laplace disturbance, obtaining an updated weight matrix, and updating the causal graph based on the updated weight matrix.
  9. 9. A computer readable storage medium having stored thereon a computer program, characterized in that the computer program comprises executable instructions which, when executed by a processor, implement the method of claim 8.
  10. 10. An electronic device, comprising: One or more processors; a memory for storing executable instructions for the processor that, when executed by the one or more processors, cause the one or more processors to implement the method of claim 8.

Description

Vehicle-mounted real-time causal graph incremental causal discovery device and discovery method Technical Field The invention relates to the technical field of automobile control, in particular to a vehicle-mounted real-time causal graph incremental causal discovery device and a discovery method. Background The existing causal discovery technology mainly relies on an offline batch processing mode of a cloud server, adopts classical methods such as PC algorithm, liNGAM and the like, and the time complexity is usuallyTo the point ofThe delay is calculated to reach the minute level, and the memory occupation exceeds the GB level. Such schemes cannot meet the stringent requirements of the in-vehicle device for real-time (< 20 ms) and low power consumption (< 1W). Although research has been made to shorten the causal graph generation time to the second level through GPU acceleration, there are still problems of high data transmission delay, strong network dependence, large risk of privacy leakage, and the like based on a cloud computing architecture. In addition, conventional causal graphs are stored using dense adjacency matrices, on a variable scaleWhen large, memory consumption grows in square, and is difficult to deploy on the limited resources of automobile-level embedded processors (typically <100MB of available memory). Existing sparse matrix optimization techniques can adopt CSR format or bitmap representation, but lack hardware-level optimization for causal graph dynamic update characteristics, resulting in frequent weight update operations still requiredThe secondary calculation becomes a real-time bottleneck. Therefore, there is a need for a causal graph generator for an in-vehicle environment, implemented on embedded hardwareIncremental updating of complexity while satisfying hard constraints of memory, latency and power consumption. Disclosure of Invention Aiming at the problems in the prior art, the invention provides a vehicle-mounted real-time causal graph incremental causal discovery device and a discovery method, which are used for solving the problem that the computational complexity is reduced from the prior artSignificantly reduce toThe causal graph generation delay is shortened from the minute level to the millisecond level. Other features and advantages of the application will be apparent from the following detailed description, or may be learned by the practice of the application. According to a first aspect of the present application there is provided a vehicle-mounted real-time causal graph incremental causal finder comprising: A data stream interface layer for receiving newly introduced sensor data; The incremental causal discovery engine is used for establishing a causal graph for the existing sensor data, acquiring a weight matrix of the causal graph, and if the newly introduced sensor data causes Laplace disturbance, locally updating the weight matrix according to the newly introduced sensor data to obtain an updated weight matrix, and updating the causal graph based on the updated weight matrix. In some embodiments of the application, based on the foregoing, the incremental cause and effect discovery engine comprises: the causal graph establishing unit is used for establishing a causal graph for the existing sensor data, wherein the causal graph comprises a plurality of nodes and a plurality of edges connected with every two nodes, and a weight matrix of the causal graph is obtained; A selection unit for receiving candidate edges; The Laplace spectrum decomposition unit is used for establishing a Laplace matrix for all sides, and carrying out Laplace spectrum decomposition on the Laplace matrix only once to obtain a base characteristic pair of each side, wherein the base characteristic pair comprises a characteristic value and a characteristic vector; The Laplace disturbance acquisition unit is used for traversing all candidate edges according to the difference value of the existing sensor data and the newly introduced sensor data, respectively calculating the partial correlation coefficient of each candidate edge, and acquiring local Laplace disturbance according to all the partial correlation coefficients; the incremental spectrum updating unit is used for updating the base characteristic pair of each candidate edge based on the local Laplace disturbance; The weight matrix updating unit is used for acquiring a local updating weight matrix based on the updated characteristic value and the characteristic vector, and carrying out local updating on the weight matrix based on the local updating weight matrix to obtain an updated weight matrix; and a causal graph updating unit for updating the causal graph based on the updated weight matrix. In some embodiments of the present application, based on the foregoing solution, the edges include candidate edges and non-candidate edges, and based on the local laplacian perturbation, a calculation formula for updating the base feature pair of e