CN-121979632-A - Industrial AI intelligent integrated machine system based on multi-model collaboration and privately-distributed deployment

CN121979632ACN 121979632 ACN121979632 ACN 121979632ACN-121979632-A

Abstract

The invention belongs to the technical field of model collaborative analysis, and discloses an industrial AI intelligent integrated machine system based on multi-model collaborative and privately deployed; the method comprises the steps of subscribing a beat trigger signal, scanning and counting, a workpiece identifier and a station identifier from a controller to generate a corresponding beat period token, generating a period data frame for binding the workpiece identifier and the station identifier, generating a multi-model reasoning task queue by analyzing the sequence of data required to be read in the period data frame by different models on a time boundary, judging overtime of different reasoning tasks in the reasoning task queue and generating a time limit evidence packet, carrying out model arbitration on the time limit evidence packet by analyzing the consistency of a control period and the workpiece identifier in the evidence packet to obtain a period decision, encoding a period decision result into a control instruction packet with the period token and an instruction sequence number, and transmitting the control instruction packet to the controller. The invention solves the problems of late and cross-cycle execution of instructions, fusion of error objects and execution of error objects.

Inventors

Shang Juncheng
WU LIUCHAO
YANG LIXIN
ZHAO ZHIHUI
SHI SHANSHAN
GUO ZHAN
MA CHUN

Assignees

河南省玄慧通智能科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260112

Claims (10)

1. Industrial AI intelligent all-in-one system based on multimode cooperation and privately-arranged is characterized by comprising: The period token generation module subscribes a beat trigger signal, a scanning count, a workpiece identifier and a station identifier from the controller to generate a corresponding beat period token; The data frame construction module is used for triggering and sampling the camera frame and the sensing sampling points based on the period token to generate a period data frame for binding the workpiece identifier and the station identifier; the multi-model task scheduling module is used for generating a multi-model reasoning task queue by analyzing the sequence of data which need to be read in the periodic data frames by different models on a time boundary; The evidence aggregation and arbitration module is used for judging overtime of different reasoning tasks in the reasoning task queue and generating a time limit evidence packet, and model arbitration is carried out on the time limit evidence packet by analyzing the consistency of a control period in the evidence packet and a workpiece identifier to obtain a period decision; The instruction packaging and issuing module encodes the periodic decision result into a control instruction packet with a periodic token and an instruction sequence number, and issues the control instruction packet to the controller.
2. The industrial AI intelligent all-in-one system based on multi-model collaborative and privately-distributed deployment of claim 1, wherein the method of generating the corresponding beat period tokens comprises: when the integrated machine receives a subscription event that the index in-place trigger bit is changed from 0 to 1, synchronously reading and storing the current beat count, the scanning count, the workpiece identifier and the station identifier and generating an event snapshot; acquiring a starting time stamp of a local monotonic clock of the integrated machine based on the event snapshot, and calculating the latest available time based on a preset time budget; Taking the starting time stamp as a time starting point of the beat control period, and taking the latest available time as the worst time delay boundary of the beat control period; Splitting the acquisition, reasoning, arbitration and issued delay budget according to the control period to generate a delay budget table; Constructing a period token according to the control period, and monitoring the control period state of the period token in real time, wherein the control period state comprises an opened, receivable result and closed state; when the control period state is closed, the corresponding period token is marked as an expired token, and the expired token is blocked from entering an arbitration and issuing channel.
3. The industrial AI intelligent all-in-one system based on multimodal collaborative and proprietary deployment of claim 2, wherein the method of generating periodic data frames binding a workpiece identity and a workstation identity comprises: according to the starting time stamp of the periodic token and the arrival time stamp of each trigger source, triggering alignment indexes are carried out, and the sampling boundary of the periodic data frame is determined; acquiring and integrating the frame number of the camera in the sampling boundary and the sensing sequence data segment to generate a frame body data set; and acquiring the workpiece identifier and the station identifier, sequentially binding the workpiece identifier and the station identifier with the periodic token, and writing the workpiece identifier and the station identifier into a data frame header.
4. The industrial AI intelligent all-in-one system based on multi-model collaborative and privately-distributed deployment of claim 3, wherein the method of determining sampling boundaries of periodic data frames comprises: respectively acquiring trigger source arrival time stamps associated with the workpiece objects in the same control period, and writing the trigger source time stamps into a trigger time stamp table; sequentially calculating relative offset for each arrival time stamp in the trigger time stamp table by taking the time starting point of the periodic token as a reference; Comparing the relative offsets of different arrival time stamps, extracting the earliest relative offset, performing difference calculation with a preset protection amount, marking the calculation result as a window starting offset, and marking a sampling starting moment at the arrival time stamp corresponding to the window starting offset; extracting the latest relative offset, adding the latest relative offset with a preset protection amount, marking the calculated result as a window end offset, and marking a sampling end moment at an arrival time stamp corresponding to the window end offset; and generating a sampling boundary according to the sampling starting time and the sampling ending time, and writing the sampling boundary into the frame head of the periodic data frame.
5. The industrial AI intelligent all-in-one system based on multimodal collaborative and proprietary deployment of claim 4, wherein the method of generating a frame body dataset comprises: after generating a period token, mapping a sampling boundary corresponding to the period token into a frame number interval and a sampling point interval in a write cache; cutting a frame number interval in an image reading cache according to a preset fixed target frame number to obtain a frame set of the frame number interval, and marking the frame set as an image block; Cutting a sampling point interval in a sensing read cache according to a preset fixed target sampling point number to obtain a sensing sequence section, and marking the sensing sequence section as a curve block; and combining the image frame set, the sensing sequence segment, the sampling boundary, the data source equipment identifier and the window starting offset to generate a frame body data set corresponding to the periodic data frame.
6. The industrial AI intelligent all-in-one system based on multimodal collaborative and privately-placed according to claim 5, wherein the method of generating a multimodal inferring task queue comprises: Analyzing and acquiring a controller identifier, a station identifier, a period number, a scanning count and a workpiece identifier from a frame head of a period data frame, and splicing to generate a task main key; Searching whether a task main key exists in a local queue manager, if so, updating the generation time of the search main key, if not, creating a task main key and generating an empty task card, and binding the empty task card with a frame head of a periodic data frame; Acquiring a frame body data set, and generating a data block according to data in the frame body data set by data type; Acquiring a multi-model list in the all-in-one machine and an input specification list of each model, reading a data block to be referenced from a frame body data set for each model, and generating a queue entry; And writing the task main key, the model identification and the reference data block into an empty task card according to the time priority sequence of the latest available moment in the queue entry to obtain an inference task, and generating an inference task queue according to different inference tasks.
7. The industrial AI intelligent all-in-one system based on multimodal collaborative and proprietary deployment of claim 6, wherein the method of model arbitration for time-limited evidence packages comprises: recording enqueuing time, starting time and finishing time of different reasoning tasks in each control period; calculating the difference value between the starting time of the reasoning task and the enqueuing time to obtain queuing time; Performing difference value calculation on the completion time and the starting time of the reasoning task to obtain calculation time; Comparing the queuing time length with a preset waiting time length threshold value, and judging that the corresponding model is overtime if the queuing time length is greater than the waiting time length threshold value; if the queuing time is not greater than the waiting time threshold, comparing the calculated time with a preset calculated time threshold, and if the calculated time is greater than the calculated time threshold, overtime the corresponding model; If the calculated time length is not greater than the calculated time length threshold, comparing the completion time with the latest available time of the reasoning task, and if the completion time exceeds the latest available time, judging that the corresponding model is not overtime; monitoring the judging result of each model in the multi-model list in real time, if the model is overtime, suspending the reasoning task to enter a reasoning task queue and judging that the corresponding model is invalid in a control period; and acquiring an effective model in the control period, and performing control decision of the control period on the effective model.
8. The industrial AI intelligent all-in-one system based on multimodal collaborative and proprietary deployment of claim 7, wherein the method of controlling decision-making of control cycles for an active model comprises: collecting a judging result, a period token, a model identifier and a workpiece identifier which are output by each model in a control period, and generating candidate evidence entries taking the period token as a key; Traversing candidate evidence items by taking the period token as a unique key, eliminating candidate evidence items with different control periods and unmatched workpiece identifiers, marking the rest candidate evidence items as effective evidence items, integrating the effective evidence items, and generating a time limit evidence packet; Selecting a control action corresponding to the effective evidence item and marking the control action as a candidate action, and constructing a candidate action decision list of the workpiece; acquiring an action code, a deadline margin, controller interlocking state information and an instruction write round trip delay threshold value of each control action in a candidate action decision list; performing executable verification on the candidate actions, and reserving the candidate actions passing the executable verification; obtaining model reliability weight, current confidence coefficient and action result cost of a model corresponding to the candidate action, carrying out weighted fusion, and recording a weighted fusion result as a desired cost value; and comparing the expected cost scores of different candidate actions, marking the candidate action corresponding to the maximum value of the expected cost scores as a decision action, and generating a periodic decision result.
9. The industrial AI intelligent all-in-one system based on multimodal collaborative and privately-placed of claim 8, wherein the method of performing the performability verification of candidate actions comprises: If the cut-off time margin is smaller than the write-in round-trip delay threshold, judging that the candidate action is not executable and is removed, and if the cut-off time margin is not smaller than the write-in round-trip delay threshold, reserving the candidate action; judging whether the candidate actions are in interlocking satisfaction on equipment where the workpiece is located according to the interlocking state information of the controller, if the interlocking is not satisfied, eliminating the candidate actions and generating an alarm according to the action codes, and if the interlocking is satisfied, reserving the candidate actions.
10. The industrial AI intelligent all-in-one system based on multi-model collaborative and privately-distributed deployment of claim 9, wherein the encoding of the periodic decision result into a control instruction packet with a periodic token and an instruction sequence number is issued to the controller before the latest available time, comprises: Encoding the periodic decisions into control instruction packets and mapping the control instruction packets into register write frames recognizable by the controller; reading the current scanning count and the interlocking state of the controller before issuing and comparing the current scanning count with the periodic token, and blocking issuing if the scanning count is inconsistent with the periodic token; Writing a control instruction packet into a controller, reading a key register, recording instruction writing time and generating a read-back abstract; if the read-back abstract is missing, the control period instruction is marked as a failure instruction, and if the instruction writing time exceeds the latest available time, the control period instruction is marked as the failure instruction; And carrying out degradation processing on the output decision action according to the failure instruction, outputting the degradation decision action and transmitting the degradation decision action to the controller.

Description

Industrial AI intelligent integrated machine system based on multi-model collaboration and privately-distributed deployment Technical Field The invention relates to the technical field of model collaborative analysis, in particular to an industrial AI intelligent integrated machine system based on multi-model collaborative and proprietary deployment. Background The industrial AI intelligent integrated machine is generally deployed in a factory private OT network, is used for accessing multi-source data such as cameras, displacement, torque, vacuum, temperature and the like at a position close to a production line, and is used for locally completing tasks such as defect detection, assembly in-place judgment, process abnormality identification and the like. Taking the automobile part turntable indexing type multi-station assembly and online detection production line as an example, a PLC (programmable logic controller) controls indexing, press fitting, screwing and rejecting mechanisms by using fixed beats, and an all-in-one machine needs to process multi-station data in each beat and outputs control conclusion which can be used for rejecting, stopping or parameter issuing, so that a closed loop of quality detection and process control is formed. In the prior all-in-one machine, the defects in terms of time sequence and certainty still commonly exist in a multi-model cooperative control link deployed by privately, such as lack of definite time sequence constraint and verifiable worst time delay boundary of a control period reference, data acquisition triggering, reasoning completion, arbitration decision-instruction issuing of a full link, so that AI reasoning and PLC control fixed beats are difficult to be aligned strictly, further, multiple-camera frames and sensing sampling are often driven by different trigger sources, if unified control period reference is not established by scanning counting or beat triggering and an associable period identifier is formed, reasoning input can be mixed across periods, multi-model sharing calculation force and asynchronous queue scheduling lead to reasoning completion time fluctuation, if latest available time and expiration discarding mechanism are lacked, a late result can still participate in arbitration and triggering actions, and when the arbitration conclusion is generated near the period boundary and the issuing link has network or protocol jitter, the control instruction can be written into the PLC/control and executed in the next time, so that the risk of instruction delay, error period execution or error object execution in the same control period appears, and multi-station tracing of the quality of a turntable is influenced. Disclosure of Invention In order to overcome the defects in the prior art and achieve the purposes, the invention provides the following technical scheme that the industrial AI intelligent integrated machine system based on multi-model collaboration and privately deployed comprises: subscribing a beat trigger signal, scanning counting, a workpiece identifier and a station identifier from a controller to generate a corresponding beat period token; triggering and sampling the camera frame and the sensing sampling points based on the periodic token to generate a periodic data frame for binding the workpiece identifier and the station identifier; generating a multi-model reasoning task queue by analyzing the sequence of data required to be read in the periodic data frames by different models on a time boundary; Performing overtime judgment on different reasoning tasks in the reasoning task queue, generating a time limit evidence packet, and performing model arbitration on the time limit evidence packet by analyzing the consistency of a control period in the evidence packet and a workpiece identifier to obtain a period decision; And encoding the cycle decision result into a control instruction packet with a cycle token and an instruction sequence number, and issuing the control instruction packet to the controller. Preferably, the method for generating the corresponding beat period token includes: when the integrated machine receives a subscription event that the index in-place trigger bit is changed from 0 to 1, synchronously reading and storing the current beat count, the scanning count, the workpiece identifier and the station identifier and generating an event snapshot; acquiring a starting time stamp of a local monotonic clock of the integrated machine based on the event snapshot, and calculating the latest available time based on a preset time budget; Taking the starting time stamp as a time starting point of the beat control period, and taking the latest available time as the worst time delay boundary of the beat control period; Splitting the acquisition, reasoning, arbitration and issued delay budget according to the control period to generate a delay budget table; Constructing a period token according to the control period, and monitorin