Fuzzy AI model training method for data processing accelerator

Abstract

The invention relates to the technical field of neural network system structures, in particular to a fuzzy AI model training method for a data processing accelerator, which comprises the following steps: 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 serially connecting to construct a sequence; and observing round-trip conflict tracks in the channels to recombine signal segments, stripping steering positions, extending non-offset linear paths, connecting the channels to the end of the process in sequence, positioning feature mapping segments, and identifying sample relationships to obtain a fuzzy training result. According to the method, the initial disturbance position is recognized by capturing signal response changes, path offset nodes are continuously tracked, dynamic connection is achieved, a clear track sequence of a conflict area is constructed to improve the connection stability, guide path fragments are recombined to define the channel propulsion sequence, and propulsion disorder caused by multi-path staggering is avoided; logic smooth conduction is achieved, mapping association is enhanced, and it is ensured that a fuzzy data guiding result has integrity and path consistency.

Description

Technical Field

[0001] This invention relates to the field of neural network architecture technology, and in particular to a method for training fuzzy AI models for data processing accelerators. Background Technology

[0002] The field of neural network architecture technology encompasses the design and implementation methods of information processing systems based on artificial neural network theory. The core content of this technology includes neuron model design, network topology construction, connection weight adjustment strategies, forward propagation mechanisms, and error backpropagation mechanisms. Neural network architecture not only involves the construction and optimization of typical structures such as single-layer perceptrons, multi-layer feedforward networks, convolutional neural networks, and recurrent neural networks, but also covers specific design methods at the hardware implementation level regarding network parameter mapping, computational resource scheduling, memory management, and parallel computing strategies. This field plays a crucial role in supporting computational acceleration and energy efficiency optimization for diverse artificial intelligence tasks and is widely applied in various scenarios such as image recognition, speech processing, natural language processing, and control systems.

[0003] Among them, the fuzzy AI model training method for data processing accelerators refers to the process of iteratively adjusting neural network parameters during the training phase for inaccurate or fuzzy data input scenarios, combined with the operating characteristics of the data processing accelerator. This is achieved by setting fuzzy data classification rules, constructing a normalized weight mapping system, and defining a nonlinear error tolerance threshold. The method mainly covers the hierarchical processing mechanism of fuzzy input data, the dynamic setting strategy of the membership function, the optimization method of the fuzzy objective function, and the calculation of the fuzzy error interval in the gradient descent path. By embedding the matching relationship between fuzzy feature parameters and weights into the data processing accelerator structure, neural network model training for fuzzy characteristic data is realized. The training process is usually controlled by parameter mapping reconstruction, step size adjustment rule setting, and statistical modeling of data fuzziness.

[0004] Existing technologies lack a real-time capture mechanism for channel response status when processing fuzzy input. In training channels, they easily overlook response lag and directional deviation during the flow of fuzzy signals along the path. Conflicting segments generated by multiple signal alternations within a channel are not explicitly identified, leading to segmentation disorder or discontinuous progress between signal paths. When multi-channel data is advanced in parallel, the path connection logic is easily misjudged as fluctuations within the error range. There is a lack of ability to correct for the signal advancement order and path structure. Position mapping deviation and path tracking confusion may occur in the training output. In complex signal scenarios, it is easy to cause unclear result boundaries and distorted feature attribution. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a fuzzy AI model training method for data processing accelerators. The technical solution is as follows: A method for training fuzzy AI models for data processing accelerators includes the following steps: S1: Obtain the batch of fuzzy training data received from the input node, send it into the training channel entrance, track the signal response changes, locate the first occurrence of the response drag along the signal forward direction, merge the first drag points of each training channel path, and obtain the set of fuzzy initial interference positions. S2: Call the training channel path of the interference position in the set of fuzzy initial interference positions, extract the signal flow path formed by the belonging signal in each path, trace the connection of adjacent channel segments, identify the offset node, connect the continuous offset nodes to construct the trajectory, and form the belonging advance offset number sequence. S3: Based on the channel segment to which the offset node in the assigned advance offset number sequence belongs, observe the alternating round-trip advance of the assigned signal, identify conflicting alternating segments by following the channel segment boundary, arrange them in order and classify them into continuous paths, extract the first segment to mark the connecting channel segment, and obtain the conflict duration segment reorganization list. S4: Extract the channel segment connected to the first continuous segment in the conflict duration reorganization list, organize the import order of the first segment of the home signal, strip the turning position, extend the path segment of the unoffset channel segment, and connect them in sequence into a linear path sequence to obtain the home guidance sequence segment set.

[0006] As a further aspect of the present invention, the fuzzy initial interference location set includes channel response delay point number, batch trigger mapping label, and interference initial appearance location identifier; the attribution propulsion offset number sequence includes signal propulsion offset code, channel segment connection sequence number, and offset node sorting mark; the conflict duration segment reorganization list includes path conflict segment number, continuous conflict area identifier, and channel segment arrangement sequence table; and the attribution guidance sequence fragment set includes import path fragment sequence number, directional stable path segment identifier, and guidance sequence connection relationship.

[0007] As a further aspect of the present invention, the step of obtaining S1 is as follows: S101: Obtain the batch received fuzzy training data in the buffer of the input node used to carry the training input, send the fuzzy training data into the training channel entrance of the data processing accelerator, detect the time series of the signal response amplitude at each channel entrance, extract the numerical difference between consecutive sampling points, judge whether the difference between adjacent sampling points exceeds the signal response change threshold, summarize the position index of the trigger judgment condition according to the channel number, and generate the channel response change index set. S102: Based on the channel response change index set, analyze the sampling sequence after the corresponding position of each index along the signal forward direction, record the sampling point intervals in which the change amplitude in the continuous segment does not meet the response growth trend, select the first sampling point that meets the response delay characteristics as the drag start point, record the corresponding position number in each training channel, summarize all drag start point numbers in the same batch and then perform sequence alignment operation to obtain the first drag position number sequence. S103: Based on the initial drag position number sequence, track the flow trajectory of the signal along the forward path in the corresponding training channel, record the continuous change interval of the subsequent response value of the number position in the channel segment, extract the interval boundary position value, and perform position intersection judgment on the corresponding records in multiple training channels. Select the path position that meets the overlap condition as the interference position point and generate a fuzzy initial interference position set.

[0008] As a further aspect of the present invention, the step of obtaining S2 is as follows: S201: Call the training channel path corresponding to each interference position in the set of fuzzy initial interference positions, extract the continuous position information of the signal advancing in each training channel path, record the sequence of the flowing position of the signal in each channel segment, arrange each position point in the sequence of flowing position in chronological order and record the position coordinate value and advancement number, and establish a signal advancement position coordinate set. S202: Based on the signal propulsion position coordinate set, the connection judgment of the signal flow path in adjacent channel segments is performed, and it is identified whether there is a propulsion breakpoint in the continuous sequence of the signal flow position between two channel segments. For the path segment with the propulsion breakpoint, the connection position number is recorded. The position repetition of the sequence composed of all recorded numbers is judged, and the set of number points that meet the offset judgment is obtained, and the signal path offset node set is generated. S203: Call the signal path offset node set, track the position of each node in the original signal flow path according to the number order, extract the channel segment number between adjacent offset nodes and combine them into a connection path sequence, perform number comparison and position offset correction on the offset nodes in each sequence, form a continuous serial structure, and establish an offset node connection path sequence. S204: Based on the offset node connection path sequence, continue to follow the subsequent change trajectory of the home signal in each channel segment, extract the position number of the adjacent offset node in the subsequent channel segment, connect the consecutively appearing numbered nodes and remove duplicate numbered segments, connect all the connected structures according to the original path order to form a single path, and generate the home advancement offset number sequence.

[0009] As a further aspect of the present invention, the step of obtaining S3 is as follows: S301: Based on the channel segment to which the offset node identified in the assigned advance offset number sequence belongs, locate the start and end positions of the assigned signal in each channel segment, record the path number and channel boundary position during the signal advance process, filter the position sequence that repeatedly crosses the boundary in the same channel segment, and extract the set of position numbers with round-trip trajectory characteristics to generate the round-trip position set of the assigned signal. S302: Call the set of return and return positions of the home signal, extract the channel segment numbers corresponding to all signal trajectories in the set, splice the alternating segments according to the number order, track the advancement path of the signal in the flow direction, perform path order division operation on the connection sequence between continuous segments, mark the advancement position and boundary connection relationship of each segment in the path, and establish a continuous segment sequence of home signal. S303: Based on the sequence of continuous segments of the attribution signal, extract the signal segment that ranks first in the sequence, identify the training channel segment number corresponding to the signal segment, use it as the starting positioning channel segment of the attribution signal import path, record the relative position index in the path sequence, and combine and summarize it with the path connection relationship in the continuous segment sequence to generate a list of conflict persistence segments to be reorganized.

[0010] As a further aspect of the present invention, the step of obtaining S4 is as follows: S401: Extract the training channel segment connected to the first continuous segment of the attribution signal path in the conflict duration reorganization list, record the sequential number of the attribution signal import position in the training channel segment, establish a signal advancement sequence according to the order of appearance of the attribution signal in the channel segment, extract the segment numbers of consecutive appearance in the advancement sequence, and generate an import position sequence. S402: Based on the imported position sequence, detect the signal flow direction vector direction corresponding to each path segment in the sequence, record the position number where the direction change occurs, peel off the numbered points with direction changes as turning marks, extract the continuous path segments with consistent flow direction and retain their original arrangement position in the advancement sequence, and generate a set of unidirectional path segments. S403: Call the set of unidirectional path segments, connect and splice each path segment in sequence according to the advancement order, record the continuation position of each path segment in the overall sequence after splicing, summarize the numbering of all connected path segments and establish channel segment mapping relationship, and generate a set of belonging guidance sequence segments.

[0011] As a further aspect of the present invention, the method further includes: S5: Extract the linear path sequence of the attribution signal from the attribution guidance sequence fragment set, connect them segment by segment according to the training channel connection order, monitor the continuous flow of the attribution signal at the connection point, identify the uninterrupted advancement range and extend it to the end, locate the feature mapping fragment and mark the sample labeling mapping relationship to obtain the fuzzy training result. The fuzzy training results include feature fragment location codes, fuzzy sample annotation association information, and training output result mapping indexes.

[0012] As a further aspect of the present invention, the step of obtaining S5 is as follows: S501: Extract the linear path sequence of the home signal from the home guidance sequence fragment set, read the training channel number connected to each path segment in sequence, combine the home guidance path segments according to the channel number order, record the start and end channel boundaries of each path segment, and perform structural continuation processing on adjacent channel segments in the combined path to generate a path segment connection order list. S502: According to the path segment connection order list, continuously track the signal flow status at the channel connection position along the forward advancement direction of the home signal, filter the path segments in which the signal status change amplitude in adjacent channel segments does not exceed the continuous advancement judgment threshold, and combine the path segments that meet the continuity condition according to the connection order to obtain the continuous advancement interval of the home signal. S503: Call the continuously advancing interval of the home signal, extend it along the advancing direction to the end of the path, perform position coverage operation on each channel segment in the continuously advancing interval, extract the feature mapping value set of the traversed area, establish the number correspondence between the mapping value and the home signal sample label, and generate fuzzy training results.

[0013] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, by capturing the continuously changing nodes of fuzzy signal responses in the training channels, the initial disturbance position that triggers propulsion anomalies is identified, enabling continuous tracking and dynamic connection of offset nodes in the signal path. A continuous and clearly segmented signal trajectory sequence is constructed within the path conflict area, improving the stable expression of path connection relationships between channels. By recombining the guiding path segments of the assigned signal, the channel propulsion order is clarified, avoiding the propulsion disorder caused by the disordered intersection of fuzzy signals between multiple paths. This achieves smooth transmission of path logic in the training process, enhances the identifiable association between signals and output mappings, and ensures that the results guided by fuzzy data during the training process have completeness and path consistency. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the process of obtaining the fuzzy initial interference position set according to the present invention. Figure 3 This is a flowchart illustrating the process of obtaining the propulsion offset number sequence of this invention. Figure 4 This is a flowchart illustrating the process of obtaining the conflict persistence segment reorganization list in this invention. Figure 5This is a flowchart illustrating the process of obtaining the hierarchical sequence fragment set of this invention. Figure 6 This is a flowchart illustrating the process of obtaining fuzzy training results according to the present invention. Detailed Implementation

[0015] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0016] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0017] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0018] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] Please see Figure 1 This invention provides a technical solution: a fuzzy AI model training method for data processing accelerators, comprising the following steps: S1: Obtain the batch of fuzzy training data received in the buffer of the input node used to carry the training input, send the fuzzy training data into the training channel entrance of the data processing accelerator, track the continuous changes of the signal response excited by the fuzzy training data at the entrance of each training channel, locate the first occurrence of the response drag along the direction of signal movement, continue to follow the signal flow trajectory in the channel where the first drag position is located, merge and mark the first drag positions on multiple training channel paths in the same batch, and obtain the set of fuzzy initial interference positions; S2: Call the training channel path corresponding to each interference position in the set of fuzzy initial interference positions, extract the signal flow path formed by the assigned signal during the advancement of each path, track the connection between adjacent channel segments of the assigned signal in the signal flow path, identify the position nodes of the assigned signal that have shifted due to feature fuzziness according to the training advancement order, connect each shift node in the signal flow path in the order of appearance to construct, continue to follow the changes of the assigned signal in the channel segment after the shift node, connect the head and tail of the consecutive shift nodes in the same signal flow path to form a complete trajectory, and form the assigned advancement shift number sequence; S3: Based on the channel segment to which the offset node belongs in the offset number sequence, locate the positional relationship of the home signal moving within the channel segment of each offset node during the training process. Enter the channel segment to which the offset node belongs to observe the back-and-forth trajectory of the home signal between alternating paths. Continuously follow the boundary position of the channel segment crossed by the back-and-forth trajectory of the home signal. Identify the alternating segments of the home signal that appear multiple times in each channel segment and exhibit path conflicts. Arrange the alternating segments of the home signal adjacent to each other according to the channel order. Unfold the continuous path of the home signal along the training flow direction. Perform sequence division and classification on the continuous segments of the home signal according to the direction of the home signal path. Extract the segment with the first position in the continuous path of the home signal and mark the training channel segment connected to the segment as the starting location basis for the home import path. Obtain the list of conflict-continuous segments to be reassembled. S4: Extract the training channel segment connected to the first continuous segment of the attribution signal path in the conflict duration reorganization list, organize the attribution signal import sequence continuity relationship formed by the first segment of the attribution signal path during the training input process, strip the signal turning position that has shifted direction during the attribution signal import process, extend the path segments formed by the attribution signal in the continuous channel segments that have not shifted direction, and connect the path segments into linear paths in sequence according to the attribution signal import order. Collect the corrected attribution signal linear path sequence to obtain the attribution guidance sequence segment set. S5: Extract the linear path sequence of the attribution signal from the attribution guidance sequence fragment set. Based on the attribution guidance path segments covered by the path sequence, connect them segment by segment according to the connection order of the training channel. Follow the forward extension order of the attribution signal along the attribution guidance path to perform the path continuation process. Track the changes in the continuous flow state of the attribution signal at the channel connection position. Identify the continuous advancement range of the attribution signal in the continuous channel segment without interruption. Extend along the advancement direction of the attribution signal to the end segment of the training process. Locate the feature mapping segments traversed by the extension range of the attribution guidance path and label the corresponding fuzzy training sample annotation mapping relationship to obtain the fuzzy training result.

[0021] The fuzzy initial interference location set includes channel response delay point number, batch trigger mapping label, and interference initial location identifier; the assigned propulsion offset number sequence includes signal propulsion offset code, channel segment connection sequence number, and offset node sorting mark; the conflict duration segment reorganization list includes path conflict segment number, continuous conflict area identifier, and channel segment arrangement sequence table; the assigned guidance order segment set includes import path segment sequence number, directional stable path segment identifier, and guidance order connection relationship; and the fuzzy training results include feature segment positioning code, fuzzy sample annotation association information, and training output result mapping index.

[0022] Please see Figure 2 The steps to obtain S1 are as follows: S101: Obtain the batch received fuzzy training data in the buffer of the input node used to carry the training input, send the fuzzy training data into the training channel entrance of the data processing accelerator, detect the time series of the signal response amplitude at each channel entrance, extract the numerical difference between consecutive sampling points, judge whether the difference between adjacent sampling points exceeds the signal response change threshold, summarize the position index of the trigger judgment condition according to the channel number, and generate the channel response change index set. The system acquires batches of fuzzy training data received from the buffer in the input node, which carries the training input. For user behavior log data or low-resolution multimedia content data with high-dimensional sparse features from internet platforms, the execution logic uses the DMA (Direct Memory Access) controller to locate the physical address of the DRAM memory buffer carrying the data. It reads data packets of a preset batch size (e.g., 256 or 512 samples). Each feature vector in this batch of data packets is mapped to the corresponding physical input pin or logic input port of the data processing accelerator (such as a GPU or TPU) via the PCIe bus. The accelerator's internal hardware performance counter starts a high-frequency sampling clock to discretize the voltage signal (analog domain) or floating-point activation value (digital domain) at the entrance of each training channel. The sampling frequency is set to twice the accelerator's main frequency to comply with the Nyquist sampling theorem and prevent aliasing. This constructs the time-series dataset for each channel. (The dimension is voltage V or a dimensionless normalized value), traverse each time step in the time series. The accelerator's tensor core is invoked to execute the value of the current sample point. Compared with the value of the previous sampling point The difference operation calculates the absolute difference between the two. Simultaneously, it invokes the signal response change threshold pre-stored in the configuration register. The threshold The setting is based on the background noise variance of historical training data from the internet platform. Three times (for example, if the background noise level is 0.02, then the threshold is set to 0.06, which is in line with...) (Criteria), the calculated absolute difference With threshold Perform a numerical comparison; if Then determine the current time step. A significant signal jump occurred; immediately capture the current time step. As a trigger index, the trigger index is written to a FIFO (First In First Out) temporary index storage queue. The above difference calculation and threshold comparison process is repeated for all parallel training channels (e.g., 1024 channels). The index values ​​that meet the trigger conditions in all channels are classified and archived according to the channel number (Channel_ID_0 to Channel_ID_1023). Empty set channels are removed. The remaining channel numbers containing the trigger index and their corresponding index lists are structurally integrated to generate a channel response change index set.

[0023] S102: Based on the channel response change index set, analyze the sampling sequence after the corresponding position of each index along the signal forward direction, record the sampling point intervals in the continuous segment where the change amplitude does not meet the response growth trend, select the first sampling point that meets the response delay characteristics as the drag start point, record the corresponding position number in each training channel, summarize all drag start point numbers in the same batch and then perform sequence alignment operation to obtain the first drag position number sequence. Based on the channel response change index set, call each trigger index position recorded in the index set. After extracting this position along the positive time axis, continue... sampling points (e.g.) A window is constructed for analysis. First-order difference gradient analysis is performed on the sampled sequence within the window, and the local slope formed by adjacent sampled points is calculated point by point. (in It is a unit clock cycle, with dimensions of (or dimensionless), introducing a benchmark value for responding to growth trends. The baseline value is set to 80% of the minimum derivative of the ideal activation function (such as ReLU) in the linear region (e.g., set to 0.8). The analysis window is continuously scanned, and once a continuous... sampling points (e.g.) Local slope Continuously below the baseline value for the response growth trend If the sampled value does not reach the upper limit of the saturation region (e.g., does not reach 1.0), then it is determined that the response growth in that interval is weak, which is called signal stagnation. This continuous range is then locked in. Time index of the first sampling point in a set of sampling points As the starting point of the drag, index this time. The corresponding training channel physical number is associated and stored. All activated training channels in the current batch are traversed, and the gradient scan and starting point locking operations described above are performed independently for each channel. A discrete set of drag starting point numbers is obtained. Using the first clock cycle of the batch data as a unified time zero point, the drag starting point of each channel is calculated. The absolute offsets relative to the zero point of time are sorted in ascending order according to their numerical values, and the time axis is normalized and calibrated. The inherent transmission delay constant of the system is subtracted to eliminate the inherent transmission delay caused by the different physical line lengths of different channels. The calibrated drag start positions are arranged according to the channel logical topology order to obtain the initial drag position number sequence.

[0024] S103: Based on the initial drag position number sequence, track the flow trajectory of the signal along the forward path in the corresponding training channel, record the continuous change interval of the subsequent response value of the number position in the channel segment, extract the interval boundary position value, and perform position intersection judgment on the corresponding records in multiple training channels. Select the path position that meets the overlap condition as the interference position point and generate a fuzzy initial interference position set. Based on the initial drag position number sequence, each drag start position recorded in the sequence is selected as the tracking start anchor point, and subsequent data are read along the forward propagation path in the corresponding training channel. time steps (e.g.) The signal response value is used to calculate the second derivative (representing curvature) or variance (representing volatility) of the signal value within that range. This identifies regions where the signal value undergoes non-monotonic changes or high-frequency oscillations, and defines the continuous change range where the signal attempts to recover or becomes distorted after lag. Specifically, this involves finding the interval. If the number of times the sign of the difference between any two adjacent points within the interval changes (i.e., the derivative changes sign) exceeds a preset frequency (e.g., 3 times, indicating the presence of high-frequency noise), then the starting boundary of the interval is extracted. and termination boundary As the interval boundary value, for multiple associated training channels that process the same set of internet user features (such as the click history and browsing history of the same user) in parallel, their corresponding continuous variation intervals are extracted, and the intersection operation of the interval sets is performed. That is, the interval tree algorithm is used to determine the interval of channel A. The interval with channel B Does an overlapping region exist? Calculate the length of the overlapping interval. ; like If the overlapping area is determined to be a common interference domain shared by multiple channels, then the center position index of the overlapping area is selected. As key interference points, all logically related channel combinations are traversed, and all center position indices that meet the overlap condition are collected. After removing duplicates, they are mapped back to the hierarchical structure coordinates of the neural network to generate a set of fuzzy initial interference positions.

[0025] Please see Figure 3 The steps to obtain S2 are as follows: S201: Call the training channel path corresponding to each interference position in the fuzzy initial interference position set, extract the continuous position information of the signal advancing in each training channel path, record the sequence of the flowing position of the signal in each channel segment, arrange each position point in the sequence of flowing position in chronological order and record the position coordinate value and advancement number, and establish a signal advancement position coordinate set. The training channel paths corresponding to each interference position in the set of fuzzy initial interference positions are invoked. For each training channel of the neural network model in the Internet platform (e.g., the connection path from the Embedding layer to the feature cross layer in a deep learning recommendation model), the Hook mechanism of the deep learning framework (such as TensorFlow or PyTorch) is used to read the activation state of the signal belonging to that channel at discrete time steps. The channel is divided into several continuous micro-segments (e.g., one micro-segment every 10 clock cycles). The signal sampling points in each micro-segment are traversed, and the positions of all sampling points whose activation values ​​exceed the silence threshold (e.g., set to 0.01 to exclude dead zone neurons) are recorded to form the original signal point cloud. According to the sampling timestamp... Sort all point cloud data in ascending order and assign each sorted sampling point a unique incrementing integer advance number. Simultaneously, the three-dimensional coordinates of the sampling point in the neural network topology are extracted. (representing hierarchy depth, in-layer row index, and in-layer column index respectively), construct a quadruple data entry containing timestamp, topological coordinates, and advancement number. For example, the activation of the neuron in the 12th row and 8th column of layer 5 at the 100th time step is recorded as follows: All quadruple data entries from all channels are merged into a unified in-memory database (such as Redis) table to establish a set of signal propulsion position coordinates.

[0026] S202: Based on the signal advance position coordinate set, perform connection judgment on the signal flow path belonging to the adjacent channel segment, identify whether there is an advance breakpoint in the continuous sequence of signal flow positions between two channel segments, record the connection position number for the path segment with the advance breakpoint, perform position repetition judgment on the sequence formed by all recorded numbers, obtain the set of number points that meet the offset judgment, and generate the signal path offset node set. Based on the signal propulsion position coordinate set, retrieve two adjacent channel segments (e.g., the first) from the database. Layer and First Calculate the timestamp difference between signal points in adjacent layers, corresponding to the coordinates of all signal propagation positions for each layer. Set a continuity determination threshold (For example, set to 5 clock cycles, determined based on hardware pipeline latency), if If there is a transmission delay or signal loss between these two points, it is determined that there is a propulsion breakpoint, and the predecessor node (the first node) at the breakpoint is extracted. The position number of the layer node) Record all identified breakpoint numbers to form a candidate breakpoint list. For each number in the list, count the frequency with which it is marked as a breakpoint in different batches of training data. Introducing a position repeatability determination threshold ; This threshold is set to 20% of the current total number of training batches (e.g., if the total number of batches is 1000, the threshold is 200), and numerical comparisons are performed. If the condition is met, then the location is confirmed as a systematic signal path offset point, rather than an occasional breakpoint caused by random noise (if not met, it is considered random packet loss and ignored). All points that meet the condition will be... Store the data in an array in ascending order to generate a set of signal path offset nodes.

[0027] S203: Call the signal path offset node set, track the position of each node in the original signal flow path according to the number order, extract the channel segment number between adjacent offset nodes and combine them into a connection path sequence, perform number comparison and position offset adjustment on the offset nodes in each sequence, form a continuous serial structure, and establish the offset node connection path sequence. Call the signal path offset node set and read the offset node number of each node in the set in sequence. By reverse-searching the original signal propagation position coordinate set, the corresponding quadruple data is determined, thus identifying its specific coordinates within the neural network hierarchy. Identify the channel segment number to which the coordinate belongs. ; Calculate the offset of two adjacent nodes and Hierarchical distance between: ,like Directly record that these two nodes are adjacent and connected. Then, extract the channel segment numbers of all intermediate levels between the two, insert them into the path list in hierarchical order, construct a complete connection path sequence, and for each node in the sequence, calculate the Euclidean distance between its actual activation position in the current batch and the preset ideal path position (determined by the model initialization weights). ,like If the offset vector is greater than the position correction threshold (e.g., set to a distance of 2 neurons), then record the offset vector. Correct the coordinates of the node in the path sequence to To eliminate positional errors caused by random disturbances, a doubly linked list structure is established for the corrected nodes in sequence, and a sequence of offset node connection paths is established.

[0028] S204: Based on the offset node connection path sequence, continue to follow the subsequent change trajectory of the home signal in each channel segment, extract the position number of the adjacent offset node in the subsequent channel segment, connect the consecutively appearing numbered nodes and remove the duplicate numbered segments, connect all the connected structures according to the original path order to form a single path, and generate the home advancement offset number sequence. Based on the offset node connection path sequence, starting from the last corrected offset node in the sequence, signal tracking continues in subsequent training channel segments (e.g., the front of fully connected layers or output layers), with the signal tracking window size set to [value missing]. (For example, 50 time steps) Scan the activation trajectory of the assigned signal within the window, identify nodes that deviate from the preset centerline (i.e., the centroid of the activation value deviates from the center of the feature map) in the trajectory, and extract the advancement numbers of these subsequent nodes. The process checks whether these numbers already exist in the previous sequence. If they do, they are removed to prevent deadlock. If they do not exist, they are appended to the end of the current path sequence. Simultaneously, it checks whether the continuously appended node numbers form a consecutive integer sequence (e.g., 1005, 1006, 1007). If so, they are merged into a single number range. Otherwise, maintain the independent numbering format, and follow the natural flow order of the original signal from the input layer to the output layer. Connect all processed numbering segments and independent numbering nodes with logical connectors to form a logical link that describes the complete distortion trajectory of the signal, and generate the assigned propulsion offset numbering sequence.

[0029] Please see Figure 4 The steps to obtain S3 are as follows: S301: Based on the channel segment to which the offset node identified in the offset number sequence belongs, locate the start and end positions of the home signal in each channel segment, record the path number and channel boundary position during the signal propulsion process, filter the position sequence that repeatedly crosses the boundary in the same channel segment, and extract the set of position numbers with round-trip trajectory characteristics to generate the round-trip position set of the home signal. Based on the channel segment to which the offset node belongs in the assigned offset number sequence, for each feature extraction channel segment (such as the connection region between convolutional and pooling layers) in the Internet platform recommendation model, read each offset node recorded in the sequence. The system retrieves the associated channel segment ID, accesses the virtual memory address mapping table (Page Table) of that channel segment, and determines the starting physical address of the signal within that segment. and termination physical address The signal trajectory tracing program is started, and the real-time read and write positions of the associated signals in the memory address space are recorded using a memory bus monitoring tool, with clock cycles as the step size. Construct a set of path points ,in and The set is traversed based on the timestamps of the signal entering and leaving the channel segment. ; Calculate two consecutive location points and Displacement vector between Identify all traversal starting addresses or termination address Events at the border, recording the moment of crossing. And the direction of crossing (positive for entering, negative for leaving); Count the total number of times the boundary is crossed within the same channel segment. ,like This indicates that the signal oscillated or backflowed at the boundary (e.g., due to data re-fetching caused by a cache miss). Further, time segments between these crossing points are extracted, and the displacement vector within each segment is calculated. If the displacement vectors and directions of two consecutive segments are opposite (e.g., moving forward first and then backward), then it is determined that the region has a round-trip trajectory feature, and the position numbers corresponding to these feature segments are extracted. (Such as memory block number or neuron index), which is stored in a dynamic array to generate a set of return and return positions of the attribution signal.

[0030] S302: Call the set of return and return positions of the home signal, extract the channel segment numbers corresponding to all signal trajectories in the set, splice the alternating segments according to the number order, track the advance path of the signal in the flow direction, perform path order division operation on the connection sequence between continuous segments, mark the advance position and boundary connection relationship of each segment in the path, and establish the continuous segment sequence of the home signal. Call the set of return and return positions of the home signal, and iterate through each return and return position number in the set. The training channel segment number to which it belongs is obtained by looking up a table. (For example, the 3rd convolutional channel), create an empty path list. According to the channel segment number The numerical values ​​(corresponding to network layer depth) are used to sort the signal segments corresponding to all round-trip positions in ascending order. For each sorted segment... Read the signal flow start time contained therein. and termination time Check two adjacent segments and Time continuity, calculating time gaps ,like If the two segments are less than a preset continuity threshold (e.g., set to 2 clock cycles), then they are determined to belong to the same continuous flow process and are included in the linked list. China establishes from point to The pointers are linked, and the boundary type at the link is recorded (e.g., "intra-layer join" or "cross-layer join"). For each breakpoint exceeding the threshold, a new split marker is inserted, dividing the linked list into multiple independent subsequences. For each subsequence, its relative percentage of progress throughout the entire training process is calculated. The percentage is marked on the head node of the subsequence to establish a continuous segment sequence of the attribution signal.

[0031] S303: Based on the sequence of continuous segments of the assigned signal, extract the signal segment at the top of the sorted order, identify the training channel segment number corresponding to the signal segment, use it as the starting positioning channel segment of the path of the assigned signal, record the relative position index in the path sequence, and combine and summarize it with the path connection relationship in the continuous segment sequence to generate a list of conflict persistence segments to be reorganized. Based on the continuous sequence of attribution signals, access the sequence storage space, read the head node information of the first sub-sequence, and extract its relative progress percentage. And the associated signal segment content, analyze the training channel segment number to which the first segment belongs. (For example, Input_Layer_0), which is locked as the reference anchor point for signal input. The entire continuous segment sequence is traversed, and a globally unique index value is assigned to each segment. (Incrementing from 0), while simultaneously reading the pointer connections established in S302 for each segment (such as Next_Ptr, Prev_Ptr), constructing a structure containing... Structured data entries, in which Indicates the connection type, and will connect the first and second segments. As the primary key, all subsequent fragments of structured data entries are used as associated attributes and combined into a hash table or adjacency list structure for regions with path conflicts (i.e., round-trip characteristics). Its duration is specifically noted. and all conflicting segments according to Sort in descending order to generate a list of conflict-prone segments to be reorganized.

[0032] Please see Figure 5 The steps to obtain S4 are as follows: S401: Extract the training channel segment connected to the first consecutive segment of the attribution signal path in the conflict duration reorganization list, record the sequential number of the attribution signal entry position in the training channel segment, establish a signal advancement sequence according to the order of occurrence of the attribution signal in the channel segment, extract the segment numbers of consecutive occurrence in the advancement sequence, and generate an entry position sequence. Extract the training channel segment connected to the first consecutive segment of the attribution signal path in the conflict persistence segment reorganization list. For the first layer neural network channel in the Internet platform responsible for processing the initial embedding of image pixels or text vectors, read the channel segment ID marked as "starting location" (e.g., Channel_Seg_001) in the list, access the input port log register of that channel segment, obtain all import event records of the attribution signal in the most recent training batch, and extract the physical address offset from each record. and write timestamp According to the timestamp of writing All imported events are sorted in ascending order based on their order of appearance, and each sorted event is assigned a monotonically increasing sequence number. (Starting from 1, for example, 1, 2, ..., N); Constructing signal propulsion sequences Traverse the sequence and check two adjacent events. and Are the sequential numbers consecutive (i.e.) At the same time, check whether its physical address offset meets the contiguous storage space adjacency condition: (For example ,in (This refers to the size of the memory block, such as 64 bytes, conforming to the cache line alignment standard). If the above two continuity conditions are met, then these two numbers will be grouped into the same continuous segment group. Repeat this check until the entire sequence has been traversed, grouping all the divided consecutive segments. According to its internal minimum Sort the values ​​twice by size, extract the start and end numbers of each group, and form a format as follows: The list of intervals is used to generate an import position sequence.

[0033] S402: Based on the import position sequence, detect the signal flow direction vector direction corresponding to each path segment in the sequence, record the position number where the direction change occurs, use the number point with the direction change as a turning mark for peeling, extract the continuous path segments with consistent flow direction and retain their original arrangement position in the advancement sequence, and generate a set of unidirectional path segments. Based on the import position order sequence, retrieve each interval in the sequence one by one. For the corresponding path segment, obtain the signal flow vector within the segment in the neural network feature space. For each segment, calculate the difference in feature vectors between its first and last positions. This vector represents the actual propagation direction of the signal within that segment, and is used to calculate the propagation direction of two adjacent segments. and Cosine value of the angle between the flow vectors (Range [-1, 1]), set the directional consistency threshold (For example, setting it to 0.9 corresponds to an angle of approximately 25 degrees), perform a numerical comparison: if Then it is determined that in the fragment and A sudden change in direction occurs at the connection point, which typically corresponds to an abnormal gradient update direction caused by noisy internet data. Record the segment number following the point of this change. As a turning marker, a culling operation is performed, removing all segments marked as turning points from the original sequence (for the gaps created by removal, logical pointers between the preceding and following segments are bridged), retaining the remaining segments that have not undergone mutation, and maintaining their relative order in the original advancing sequence (e.g., if the original sequence is ABCD, and C is the mutation point, then ABD is retained), re-indexing the retained segments to form a new ordered list, generating a set of unidirectional path segments.

[0034] S403: Call the unidirectional path segment set, connect and splice each path segment in sequence according to the advancement order, record the continuation position of each path segment in the overall sequence after splicing, summarize the numbering of all connected path segments and establish channel segment mapping relationship, and generate the belonging guidance sequence segment set; Call the collection of one-way path fragments to initialize an empty global path linked list. Following the new index order determined in S402, each unidirectional path segment in the set is read sequentially. Extract the starting node coordinates of the segment. and the coordinates of the termination node ,Will Data structures are appended to the linked list At the end of the list, during the append process, record the index of the current segment's continuation position in the linked list. (Incrementing from 0), and establish a bidirectional pointer connection between the current segment and the previous segment. For each connected segment, query its original channel segment mapping table to obtain the physical channel ID corresponding to the segment (e.g., a specific BankID in the GPU memory or the core number of the TPU). Build key-value pairs All key-value pairs are stored in a B+ tree or hash map structure to ensure that physical hardware resources can be quickly looked up through linked list indexes. At the same time, the total length of the entire concatenated path is calculated. Verify whether the length matches the theoretical expected value of the model depth (e.g., the preset depth is 50 layers, with an allowable error). If the verification passes, the serial structure is confirmed to be valid, and a set of hierarchical ordered fragments is generated.

[0035] Please see Figure 6 The steps to obtain S5 are as follows: S501: Extract the linear path sequence of the home signal from the home guidance sequence fragment set, read the training channel number connected to each path segment in sequence, combine the home guidance path segments according to the channel number order, record the start and end channel boundaries of each path segment, and perform structural continuation processing on adjacent channel segments in the combined path to generate a path segment connection order list. Extract the linear path sequence of the attribution signal from the attribution-oriented sequence fragment set. For the modified feature extraction path in the deep learning recommendation system of the Internet platform, traverse the data structure of each path node in the sequence and read the physical number of the training channel stored therein. (For example, Conv_Layer_3_Channel_15), based on the hierarchical topology of the neural network (such as the hierarchical connection rules of the ResNet structure), all extracted channel numbers are processed according to... Arrange and combine them in ascending order to construct an ordered array of channel segments. For two adjacent channel segments in the array and Extract respectively Output port address and Input port address Record these two addresses as their respective start and end channel boundaries, perform a structural continuity check, and calculate the address offset. ,like If the memory alignment step is equal to the preset value (e.g., 32 bytes), confirm that the physical connection is valid. If the value is not zero and not equal to the step size (e.g., in cross-card communication scenarios), the DMA (Direct Memory Access) descriptor is called to establish a virtual mapping connection between the two addresses. This mapping relationship is then written into the connection configuration table. All verified and confirmed channel segment metadata (including number, boundary address, and connection type) is encapsulated into a structure and stored in a linked list in sequence to generate a path segment connection order list.

[0036] S502: Based on the path segment connection order list, continuously track the signal flow status at the channel connection position along the forward advancement direction of the assigned signal, filter the path segments in which the signal status change amplitude in adjacent channel segments does not exceed the continuous advancement judgment threshold, and combine the path segments that meet the continuity condition according to the connection order to obtain the continuous advancement interval of the assigned signal. Based on the path segment connection order list, the signal status monitoring thread is activated. Following the forward progression direction of the assigned signal in the list (i.e., from the input layer to the output layer), the probe reads in real-time the connection points of each channel (i.e.,...) and (between) signal activation values Calculate the connection points between two adjacent points. and The ratio of signal energy change between ; in For time window The average energy within the range (physically defined as the root mean square of the signal strength) is used to introduce a continuous propagation determination threshold. This threshold is set to 1.5 times the average energy decay rate of historically stable training samples on the internet platform (for example, if the historical decay rate is 0.1, then...). Perform a numerical comparison: if This indicates that no significant gradient vanishing or exploding occurred during signal transmission, satisfying the continuity condition. The corresponding path segment is marked as "active," and its index is added to the current continuous interval. ,like If signal transmission is interrupted or abnormally attenuated (in case of abnormality, trigger a local rollback of model parameters or a dynamic adjustment mechanism for the learning rate), the current interval will immediately end. The record is started, and a new interval record is opened. This process is repeated until the entire list is traversed. All continuous paths marked as "active" and whose length exceeds the minimum layer limit (such as 3 layers) are combined to obtain the continuous advancement interval of the belonging signal.

[0037] S503: Call the continuous advancement interval of the home signal, extend it along the advancement direction to the end of the path, perform position coverage operation on each channel segment in the continuous advancement interval, extract the feature mapping value set of the traversed area, establish the number correspondence between the mapping value and the home signal sample label, and generate fuzzy training results; The system continuously advances the interval based on the attribution signal, locks the end position of that interval in the neural network architecture (i.e., the node before the last fully connected layer or Softmax layer), and reads the output tensor vector at that position at the end of training. This vector corresponds to the predicted probability distribution of a specific user behavior (such as clicking or purchasing) on ​​an internet platform. Simultaneously, it retrieves the feature map regions covered by all intermediate channel segments within this interval, calculates the activation heatmap for each region, and extracts the coordinates of pixels with activation values ​​higher than a significance threshold (e.g., 0.7, normalized value) from the heatmap to form a set of feature mapping values. Read the real business labels of the current fuzzy samples from the original training data label library. (e.g., "high-potential users"), establish the output vector Feature Mapping Set With real labels A ternary association mapping table is used to assign a unique hash index to each mapping relationship. The index is then written to the results database (such as MySQL or HBase) as the basis for subsequent model inference and weight correction, generating fuzzy training results.

[0038] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for training a fuzzy AI model for a data processing accelerator, characterized in that, Includes the following steps: S1: Obtain the batch of fuzzy training data received from the input node, send it into the training channel entrance, track the signal response changes, locate the first occurrence of the response drag along the signal forward direction, merge the first drag points of each training channel path, and obtain the set of fuzzy initial interference positions. S2: Call the training channel path of the interference position in the set of fuzzy initial interference positions, extract the signal flow path formed by the belonging signal in each path, trace the connection of adjacent channel segments, identify the offset node, connect the continuous offset nodes to construct the trajectory, and form the belonging advance offset number sequence. S3: Based on the channel segment to which the offset node in the assigned advance offset number sequence belongs, observe the alternating round-trip advance of the assigned signal, identify conflicting alternating segments by following the channel segment boundary, arrange them in order and classify them into continuous paths, extract the first segment to mark the connecting channel segment, and obtain the conflict duration segment reorganization list. S4: Extract the channel segment connected to the first continuous segment in the conflict duration reorganization list, organize the import order of the first segment of the home signal, strip the turning position, extend the path segment of the unoffset channel segment, and connect them in sequence into a linear path sequence to obtain the home guidance sequence segment set.

2. The fuzzy AI model training method for data processing accelerators according to claim 1, characterized in that: The set of fuzzy initial interference locations includes channel response delay point numbers, batch trigger mapping labels, and interference initial location identifiers. The attribution propulsion offset number sequence includes signal propulsion offset codes, channel segment connection numbers, and offset node sorting marks. The conflict duration reorganization list includes path conflict segment numbers, continuous conflict area identifiers, and channel segment arrangement sequence tables. The attribution guidance sequence fragment set includes import path fragment numbers, direction-stable path segment identifiers, and guidance sequence connection relationships.

3. The fuzzy AI model training method for data processing accelerators according to claim 1, characterized in that: The steps for obtaining S1 are as follows: S101: Obtain the batch received fuzzy training data in the buffer of the input node used to carry the training input, send the fuzzy training data into the training channel entrance of the data processing accelerator, detect the time series of the signal response amplitude at each channel entrance, extract the numerical difference between consecutive sampling points, judge whether the difference between adjacent sampling points exceeds the signal response change threshold, summarize the position index of the trigger judgment condition according to the channel number, and generate the channel response change index set. S102: Based on the channel response change index set, analyze the sampling sequence after the corresponding position of each index along the signal forward direction, record the sampling point intervals in which the change amplitude in the continuous segment does not meet the response growth trend, select the first sampling point that meets the response delay characteristics as the drag start point, record the corresponding position number in each training channel, summarize all drag start point numbers in the same batch and then perform sequence alignment operation to obtain the first drag position number sequence. S103: Based on the initial drag position number sequence, track the flow trajectory of the signal along the forward path in the corresponding training channel, record the continuous change interval of the subsequent response value of the number position in the channel segment, extract the interval boundary position value, and perform position intersection judgment on the corresponding records in multiple training channels. Select the path position that meets the overlap condition as the interference position point and generate a fuzzy initial interference position set.

4. The fuzzy AI model training method for data processing accelerators according to claim 1, characterized in that: The steps for obtaining S2 are as follows: S201: Call the training channel path corresponding to each interference position in the set of fuzzy initial interference positions, extract the continuous position information of the signal advancing in each training channel path, record the sequence of the flowing position of the signal in each channel segment, arrange each position point in the sequence of flowing position in chronological order and record the position coordinate value and advancement number, and establish a signal advancement position coordinate set. S202: Based on the signal propulsion position coordinate set, the connection judgment of the signal flow path in adjacent channel segments is performed, and it is identified whether there is a propulsion breakpoint in the continuous sequence of the signal flow position between two channel segments. For the path segment with the propulsion breakpoint, the connection position number is recorded. The position repetition of the sequence composed of all recorded numbers is judged, and the set of number points that meet the offset judgment is obtained, and the signal path offset node set is generated. S203: Call the signal path offset node set, track the position of each node in the original signal flow path according to the number order, extract the channel segment number between adjacent offset nodes and combine them into a connection path sequence, perform number comparison and position offset correction on the offset nodes in each sequence, form a continuous serial structure, and establish an offset node connection path sequence. S204: Based on the offset node connection path sequence, continue to follow the subsequent change trajectory of the home signal in each channel segment, extract the position number of the adjacent offset node in the subsequent channel segment, connect the consecutively appearing numbered nodes and remove duplicate numbered segments, connect all the connected structures according to the original path order to form a single path, and generate the home advancement offset number sequence.

5. The fuzzy AI model training method for data processing accelerators according to claim 1, characterized in that: The steps for obtaining S3 are as follows: S301: Based on the channel segment to which the offset node identified in the assigned advance offset number sequence belongs, locate the start and end positions of the assigned signal in each channel segment, record the path number and channel boundary position during the signal advance process, filter the position sequence that repeatedly crosses the boundary in the same channel segment, and extract the set of position numbers with round-trip trajectory characteristics to generate the round-trip position set of the assigned signal. S302: Call the set of return and return positions of the home signal, extract the channel segment numbers corresponding to all signal trajectories in the set, splice the alternating segments according to the number order, track the advancement path of the signal in the flow direction, perform path order division operation on the connection sequence between continuous segments, mark the advancement position and boundary connection relationship of each segment in the path, and establish a continuous segment sequence of home signal. S303: Based on the sequence of continuous segments of the attribution signal, extract the signal segment that ranks first in the sequence, identify the training channel segment number corresponding to the signal segment, use it as the starting positioning channel segment of the attribution signal import path, record the relative position index in the path sequence, and combine and summarize it with the path connection relationship in the continuous segment sequence to generate a list of conflict persistence segments to be reorganized.

6. The fuzzy AI model training method for data processing accelerators according to claim 1, characterized in that: The steps for obtaining S4 are as follows: S401: Extract the training channel segment connected to the first continuous segment of the attribution signal path in the conflict duration reorganization list, record the sequential number of the attribution signal import position in the training channel segment, establish a signal advancement sequence according to the order of appearance of the attribution signal in the channel segment, extract the segment numbers of consecutive appearance in the advancement sequence, and generate an import position sequence. S402: Based on the imported position sequence, detect the signal flow direction vector direction corresponding to each path segment in the sequence, record the position number where the direction change occurs, peel off the numbered points with direction changes as turning marks, extract the continuous path segments with consistent flow direction and retain their original arrangement position in the advancement sequence, and generate a set of unidirectional path segments. S403: Call the set of unidirectional path segments, connect and splice each path segment in sequence according to the advancement order, record the continuation position of each path segment in the overall sequence after splicing, summarize the numbering of all connected path segments and establish channel segment mapping relationship, and generate a set of belonging guidance sequence segments.

7. The fuzzy AI model training method for data processing accelerators according to claim 1, characterized in that: The method further includes: S5: Extract the linear path sequence of the attribution signal from the attribution guidance sequence fragment set, connect them segment by segment according to the training channel connection order, monitor the continuous flow of the attribution signal at the connection point, identify the uninterrupted advancement range and extend it to the end, locate the feature mapping fragment and mark the sample labeling mapping relationship to obtain the fuzzy training result. The fuzzy training results include feature fragment location codes, fuzzy sample annotation association information, and training output result mapping indexes.

8. The fuzzy AI model training method for data processing accelerators according to claim 7, characterized in that: The steps for obtaining S5 are as follows: S501: Extract the linear path sequence of the home signal from the home guidance sequence fragment set, read the training channel number connected to each path segment in sequence, combine the home guidance path segments according to the channel number order, record the start and end channel boundaries of each path segment, and perform structural continuation processing on adjacent channel segments in the combined path to generate a path segment connection order list. S502: According to the path segment connection order list, continuously track the signal flow status at the channel connection position along the forward advancement direction of the home signal, filter the path segments in which the signal status change amplitude in adjacent channel segments does not exceed the continuous advancement judgment threshold, and combine the path segments that meet the continuity condition according to the connection order to obtain the continuous advancement interval of the home signal. S503: Call the continuously advancing interval of the home signal, extend it along the advancing direction to the end of the path, perform position coverage operation on each channel segment in the continuously advancing interval, extract the feature mapping value set of the traversed area, establish the number correspondence between the mapping value and the home signal sample label, and generate fuzzy training results.

Images