Search

CN-121997979-A - Fuzzy AI model training method for data processing accelerator

CN121997979ACN 121997979 ACN121997979 ACN 121997979ACN-121997979-A

Abstract

The invention relates to the technical field of neural network architecture, in particular to a fuzzy AI model training method for a data processing accelerator, which is used for acquiring fuzzy data, sending the fuzzy data into the accelerator, tracking a first dragging point of a signal to merge an interference position set, identifying offset nodes along a path and connecting in series a construction sequence, observing a reciprocating conflict track in a channel to reorganize a signal segment, stripping a steering position and extending a linear path which is not offset, sequentially connecting the channel to the end of the flow, positioning a feature mapping segment and identifying a sample relation, and obtaining a fuzzy training result. According to the invention, the initial disturbance position is identified by capturing signal response change, the path offset node is continuously tracked and dynamic engagement is realized, a clear track sequence of a collision area is constructed to promote engagement stability, a channel propulsion sequence is clarified by the recombined guide path segment, the propulsion disorder caused by multipath staggering is avoided, logic smooth conduction is realized, mapping association is enhanced, and the integrity and path consistency of a fuzzy data guiding result are ensured.

Inventors

  • JI HE

Assignees

  • 中国传媒大学

Dates

Publication Date
20260508
Application Date
20260127

Claims (8)

  1. 1. A method for training an fuzzy AI model for a data processing accelerator, comprising the steps of: s1, acquiring fuzzy training data received in batches in input nodes, sending the fuzzy training data into a training channel inlet, tracking signal response changes, positioning the first occurrence position of response drag along the signal advancing direction, merging the first drag points of each training channel path, and obtaining a fuzzy initial interference position set; s2, calling training channel paths of interference positions in the fuzzy initial interference position set, extracting signal circulation paths formed by attribution signals in each path, tracking the connection of adjacent channel segments, identifying offset nodes, and connecting continuous offset nodes in series to construct tracks to form attribution propulsion offset number sequences; s3, observing the alternate reciprocating pushing of the attribution signal according to the channel section of the offset node in the attribution pushing offset number sequence, identifying conflict alternate fragments along the boundary of the channel section, sequentially adjacently arranging, dividing and classifying continuous paths, extracting a first fragment to calibrate a connecting channel section, and obtaining a conflict continuous segment recombination list; and S4, extracting channel segments connected with the first continuous segment in the conflict continuous segment recombination list, sorting the leading-in sequence of the first segment of the attribution signal, stripping the turning position, extending the path segments of the non-offset channel segments, and connecting the path segments in series in sequence to form a linear path sequence to obtain an attribution guiding sequence segment set.
  2. 2. The method for training the fuzzy AI model for a data processing accelerator as set forth in claim 1 wherein said fuzzy starting interference location set comprises a channel response delay point number, a batch trigger map tag, an interference primary location identifier, said home propulsion offset number sequence comprises a signal propulsion offset code, a channel segment connection sequence number, an offset node ordering flag, said conflict duration segment reassembly list comprises a path conflict segment number, a continuous conflict region identifier, a channel segment ordering list, and said home guidance sequence segment set comprises a import path segment number, a direction stable path segment identifier, a guidance sequence engagement relationship.
  3. 3. The method for training the fuzzy AI model for a data processing accelerator as set forth in claim 1, wherein the step of S1 is: S101, acquiring fuzzy training data received in batches in a buffer zone used for bearing training input in an input node, sending the fuzzy training data into a training channel inlet of a data processing accelerator, detecting a time sequence of signal response amplitude at each channel inlet, extracting numerical value differences between continuous sampling points, judging according to whether the difference value of adjacent sampling points exceeds a signal response change threshold value, summarizing position indexes of trigger judgment conditions according to channel numbers, and generating a channel response change index set; S102, analyzing a sampling sequence after corresponding positions of each index along the signal advancing direction according to the channel response change index set, recording sampling point intervals in which the change amplitude in a continuous section does not meet the response increasing trend, screening a first sampling point which accords with response delay characteristics as a dragging starting point, correspondingly recording the position numbers in each training channel, summarizing all the dragging starting point numbers in the same batch, and then executing sequence alignment operation to obtain a first dragging position number sequence; S103, tracking the circulation track of the signal along the forward path in the corresponding training channel based on the first dragging position number sequence, recording the continuous change interval of the subsequent response value of the number position in the channel section, extracting the boundary position value of the interval, performing position intersection judgment on the corresponding records in a plurality of training channels, selecting the path position meeting the coincidence condition as an interference position point, and generating a fuzzy initial interference position set.
  4. 4. The method for training the fuzzy AI model for a data processing accelerator as set forth in claim 1, wherein the step of S2 is: s201, calling training channel paths corresponding to interference positions in the fuzzy initial interference position set, extracting continuous position information of the attribution signals in each training channel path when the attribution signals are propelled, recording a circulation position sequence of the attribution signals in each channel section, arranging position points in the circulation position sequence according to time sequence, recording position coordinate values and propulsion numbers, and establishing a signal propulsion position coordinate set; S202, judging the connection of the attribution signal circulation paths in the adjacent channel segments according to the signal propulsion position coordinate set, identifying whether propulsion break points exist in the continuous sequence at the signal circulation positions between the two channel segments, recording connection position numbers for the channel segments with the propulsion break points, judging the position repeatability of the sequence formed by all the record numbers, obtaining a number point set conforming to the offset judgment, and generating a signal path offset node set; S203, calling the signal path offset node set, tracking the appearance position of each node in the original signal circulation path according to the serial number sequence, extracting the serial numbers of channel segments between adjacent offset nodes and combining the serial numbers into a connection path sequence, comparing the serial numbers of the offset nodes in each sequence, correcting the position offset, forming a continuous serial structure, and establishing the offset node connection path sequence; S204, according to the offset node connection path sequence, continuing to follow the subsequent change track of the attribution signal in each channel section, extracting the position numbers of adjacent offset nodes in the subsequent channel sections, concatenating the serial number nodes which continuously appear, removing the repeated serial number sections, and connecting all the concatenated structures according to the original path sequence to form a single path so as to generate an attribution propulsion offset serial number sequence.
  5. 5. The method for training the fuzzy AI model for a data processing accelerator as set forth in claim 1, wherein the step of S3 is: s301, positioning the moving start and stop positions of the home signal in each channel segment according to the channel segment to which the offset node identified in the home propulsion offset number sequence belongs, recording the path numbers and channel boundary positions in the signal propulsion process, screening the position sequences repeatedly crossing the boundary in the same channel segment, extracting a position number set with a round-trip track characteristic, and generating a home signal round-trip position set; S302, calling the home signal round trip position set, extracting channel segment numbers corresponding to all signal tracks in the set, splicing alternating segments according to the number sequence, tracking the advancing path of a signal in the flowing direction, executing path sequence dividing operation on a connecting sequence between continuous segments, marking the advancing position and boundary connecting relation of the segments in the path, and establishing a home signal continuous segment sequence; s303, extracting the signal segments with the first sequence according to the continuous segment sequence of the attribution signal, identifying the training channel segment numbers corresponding to the signal segments, taking the training channel segment numbers as the initial positioning channel segments of the attribution signal importing paths, recording the relative position indexes in the path sequences, and combining and summarizing the relative position indexes with the path connection relations in the continuous segment sequences to generate a conflict continuous segment reorganization list.
  6. 6. The method for training the fuzzy AI model for a data processing accelerator as set forth in claim 1, wherein the step of S4 is: S401, extracting a training channel segment connected with the first continuous segment of the attribution signal path in the conflict continuous segment recombination list, recording sequence numbers of attribution signal import positions in the training channel segment, establishing a signal propulsion sequence according to the appearance sequence of attribution signals in the channel segment, extracting segment numbers with continuous appearance sequence in the propulsion sequence, and generating an import position sequence; S402, detecting the signal flow vector direction corresponding to each path segment in the sequence according to the sequence of the leading-in position sequence, recording the position number with the direction mutation, taking the numbered point with the direction mutation as a turning mark for stripping, extracting the path segments with continuous flow consistency and keeping the original arrangement positions in the pushing sequence, and generating a unidirectional path segment set; S403, calling the unidirectional path segment set, sequentially splicing the path segments in series according to the propulsion sequence, recording the connection position of each path segment in the whole sequence after splicing, performing numbering summarization on all the path segments after connection, establishing a channel segment mapping relation, and generating an attribution guiding sequence segment set.
  7. 7. The fuzzy AI model training method for a data processing accelerator of claim 1, further comprising: S5, extracting a home signal linear path sequence in the home guide sequence fragment set, connecting the home signal linear path sequence section by section according to a training channel connection sequence, monitoring the continuous flow of the home signal at the connecting position, identifying the uninterrupted propulsion range and extending to the tail end, positioning a feature mapping fragment and identifying a sample labeling mapping relation to obtain a fuzzy training result; the fuzzy training result comprises feature fragment positioning codes, fuzzy sample annotation associated information and training output result mapping indexes.
  8. 8. The method for training the fuzzy AI model for a data processing accelerator as set forth in claim 7, wherein the step of S5 is: S501, extracting a home signal linear path sequence in the home guide sequence fragment set, sequentially reading training channel numbers connected with each path segment, combining the home guide path segments according to the channel number sequence, recording the start-stop channel boundary of each path segment, and carrying out structure connection processing on adjacent path segments in the combined path to generate a path segment connection sequence list; s502, continuously tracking signal flow states at the joint positions of the channels along the forward advancing direction of the attribution signals according to the path segment connection sequence list, screening path segments with signal state change amplitude not exceeding a continuous advancing judgment threshold value in adjacent channel segments, and combining the path segments meeting the continuity condition according to the connection sequence to obtain an attribution signal continuous advancing interval; S503, calling the attribution signal continuous pushing section, extending to a path end point along the pushing direction, performing position coverage operation on each channel section in the continuous pushing section, extracting a characteristic mapping value set of the traversed area, establishing a serial number corresponding relation between the mapping value and attribution signal sample labels, and generating a fuzzy training result.

Description

Fuzzy AI model training method for data processing accelerator Technical Field The invention relates to the technical field of neural network architectures, in particular to a fuzzy AI model training method for a data processing accelerator. Background The technical field of the neural network architecture comprises a design and implementation method for constructing an information processing system based on an artificial neural network theory, the core content of the technical field comprises neuron model design, network topology construction, connection weight adjustment strategy, forward propagation mechanism, error back propagation mechanism and the like, the neural network architecture not only relates to construction and optimization of typical structures such as a single-layer perceptron, a multi-layer feedforward network, a convolutional neural network, a cyclic neural network and the like, but also covers specific design modes of network parameter mapping, calculation resource scheduling, storage management and parallel calculation strategy in a hardware implementation level, plays an important role in supporting calculation acceleration and energy efficiency optimization of diversified artificial intelligent tasks, and is widely applied to a plurality of scenes such as image recognition, voice processing, natural language processing and control systems. The fuzzy AI model training method for the data processing accelerator aims at the condition of inaccurate or fuzzy data input in a training stage, combines the operation characteristics of the data processing accelerator, completes the iterative adjustment process of the neural network parameters by setting a fuzzy data classification rule, constructing a normalized weight mapping system and defining a nonlinear error tolerance threshold, mainly covers the hierarchical processing mechanism of fuzzy input data, the dynamic setting strategy of a attribution function, the optimization mode of a fuzzy objective function and the accommodation calculation of a fuzzy error interval in a gradient descent path, and realizes the training of the neural network model aiming at fuzzy characteristic data by embedding the matching relation between fuzzy characteristic parameters and weights in the data processing accelerator structure, and generally completes the training process control of the model by a parameter mapping reconstruction, step length adjustment rule setting and a statistical modeling mode of data ambiguity. In the prior art, when fuzzy input is processed, a real-time capturing mechanism for channel response states is lacking, response delay and direction deviation conditions in the process of circulating a fuzzy signal along a path are easy to ignore in a training channel, conflict fragments generated by multiple signal alternation in the channel are not explicitly identified, so that the problem of segment disorder or discontinuous propelling among signal paths is caused, when multi-channel data parallel propelling is involved, path connection logic is easy to be misjudged as a fluctuation phenomenon in an error range, correction capability for a signal propelling sequence and a path structure is lacking, position mapping deviation and path tracking confusion can be generated in training output, and result boundary disorder and characteristic attribution distortion are easy to be caused in a complex signal scene. Disclosure of Invention In order to solve the technical problems in the prior art, the embodiment of the invention provides a fuzzy AI model training method for a data processing accelerator. The technical scheme is as follows: A fuzzy AI model training method for a data processing accelerator, comprising the steps of: s1, acquiring fuzzy training data received in batches in input nodes, sending the fuzzy training data into a training channel inlet, tracking signal response changes, positioning the first occurrence position of response drag along the signal advancing direction, merging the first drag points of each training channel path, and obtaining a fuzzy initial interference position set; s2, calling training channel paths of interference positions in the fuzzy initial interference position set, extracting signal circulation paths formed by attribution signals in each path, tracking the connection of adjacent channel segments, identifying offset nodes, and connecting continuous offset nodes in series to construct tracks to form attribution propulsion offset number sequences; s3, observing the alternate reciprocating pushing of the attribution signal according to the channel section of the offset node in the attribution pushing offset number sequence, identifying conflict alternate fragments along the boundary of the channel section, sequentially adjacently arranging, dividing and classifying continuous paths, extracting a first fragment to calibrate a connecting channel section, and obtaining a conflict continuous segment recombin