Artificial intelligence-based production line adjustment control system and method

Abstract

The application discloses a production line adjustment control system and method based on artificial intelligence, and relates to the technical field of intelligent manufacturing and artificial intelligence control; comprising a data review module, an intelligent adjustment analysis module and a control module. The data review module combines a bias learning network, a variational autoencoder and a multi-modal anomaly recognition mechanism to dynamically determine process bias and potential drift; the intelligent adjustment analysis module uses a causal graph neural network, an expert rule base and a SimFormer scorer to generate and simulate the optimal control strategy; the control module executes the strategy through an industrial communication protocol and feeds back the results in real time, forming a closed-loop optimization process. The application has high robustness, strong adaptability and intelligent decision-making ability, and is suitable for complex and variable manufacturing scenarios, significantly improving production efficiency and quality stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and artificial intelligence control, in particular to a production line adjustment control system and method based on artificial intelligence. BACKGROUND

[0002] With the development of manufacturing industry towards high-end and intelligent direction, the traditional production line control system generally has problems such as response lag, strong regulation and control rigidity, and lack of intelligent decision-making ability. Especially in the face of variable task types, complex equipment states and multi-source interference factors, the existing system mainly relies on preset threshold and fixed rules for judgment and adjustment, which is difficult to cope with nonlinear, dynamic coupling and sudden abnormal situations.

[0003] In addition, the current mainstream process parameter review and control adjustment means rely on artificial experience, single-point monitoring or simple statistical rules, which cannot effectively identify potential soft abnormalities, complex causal relationships or structured similarities between historical tasks, resulting in that the adjustment strategy cannot accurately adapt to the current production environment and equipment state, affecting product quality and production stability.

[0004] Therefore, there is an urgent need for a production line adjustment control system that can integrate artificial intelligence algorithms, causal reasoning, generative adversarial learning and multi-objective simulation evaluation to achieve high-precision identification, root cause positioning and optimal control strategy generation of production line abnormalities. SUMMARY

[0005] The purpose of the present application is to solve the problems mentioned in the background art, and to provide a production line adjustment control system and method based on artificial intelligence.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] In a first aspect, the present application provides a production line adjustment control system based on artificial intelligence, comprising a data review module, an intelligent adjustment analysis module and a control module; comprising:

[0008] The data review module receives production line information from the data acquisition module, verifies according to the set process, triggers abnormal marking, abnormal confirmation and soft abnormality;

[0009] The intelligent adjustment analysis module processes the abnormality identified by the data review module, including:

[0010] Multi-dimensional abnormal reason identification is performed, CGNN is used for causal learning on a graph structure G (V, E) to construct a causal graph, an intervention estimation method is used to model the causal strength to output a causal contribution score, a Bayesian structure search is used to strengthen causal direction identification to obtain a propagation chain path Pab to obtain a causal path length of the abnormal node, and a core root cause node is output according to the causal path length of the abnormal node and the causal contribution score;

[0011] According to the root cause node, a historical strategy vector set is extracted from an expert rule base and a similar task database, a generator is used to construct a strategy candidate vector in combination with a current state vector, the historical strategy vector set and the constructed strategy candidate vector are spliced, a joint feature vector is extracted through a multi-layer perception machine, a historical behavior encoding difference value C is compared, an approximation degree of the strategy and an efficient strategy is analyzed, and a structured strategy is output.

[0012] The SimFormer model is used to splice the state, abnormal parameters and strategy vector into an input token stream, a multi-dimensional simulation index vector is output based on an Encoder-Decoder structure, each dimension of the multi-dimensional simulation index vector is normalized, a weighted benefit function is used to calculate a standardized index value, and sorting and labeling are performed.

[0013] The strategy combination is regarded as a multi-objective resource game among multiple agents, each strategy has its own benefit function U, a non-cooperative game model is constructed, and the optimal non-conflict combination is output by solving, and the selected strategy is mapped into a standard control instruction.

[0014] The control module receives the standard control instruction, and executes the operation according to the safety mechanism and the real-time feedback mechanism.

[0015] As a preferred embodiment of the present application, the data acquisition module is arranged at a fixed position of the production line, and is used for acquiring data of the production line, including equipment running state parameters, process parameters, voiceprint signals and production task information; the acquired data is input to the embedded data processing unit for preliminary processing to form unified structured production line information, and then transmitted to the data review module.

[0016] As a preferred embodiment of the present application, the process of triggering the abnormal mark is verified according to the set process:

[0017] According to the received production task, the corresponding standard process parameter set {u iretrieve historical task data of the same or similar task type as the current task, construct a composite feature vector: F(t); combine the historical records of the same task, the device state and the environmental factors, call the bias learning network, dynamically infer the tolerance boundary for each process parameter, perform adaptive process constraint modeling, and output the dynamic tolerance boundary value of the parameter under the current conditions; compare the real-time production line information {x i} collected with the dynamic tolerance boundary value output by the model, and determine whether the inequality constraint condition is met. If not, an abnormality flag is triggered.

[0018] As a preferred embodiment of the present application, the process of triggering abnormality confirmation according to the set flow is as follows:

[0019] The current time sequence is dynamically registered with the standard template obtained by clustering normal working condition samples using the dynamic time warping algorithm, and the DTW distance is calculated. If the distance exceeds the set threshold, it is determined that there is a shape abnormality. After extracting the feature vectors of the current sequence and the standard template sequence, the angle similarity is calculated, and the cosine similarity is used as an index. If the similarity is lower than the threshold, it is determined that there is a trend deviation. If both indicators are determined to be abnormal, the abnormality confirmation is triggered.

[0020] As a preferred embodiment of the present application, the process of triggering soft abnormality according to the set flow is as follows:

[0021] For information without abnormality, the feature encoding sequence {Z t-k+1 ,…,Z t} of continuous k time steps is input into the attention-enhanced variational autoencoder model, and the hidden variable distribution parameter {m t} is output.

[0022] According to the hidden variable distribution parameter, the deviation between the current window and the normal hidden variable distribution is measured using the KL divergence: , where q(z) is the hidden variable distribution of the current window, p(z) is the hidden variable distribution under normal state, and The hidden variable vector is extracted for each time window. If the KL divergence of continuous m time windows is higher than the set threshold τ, it is determined that there is a continuous potential drift at present. It is marked as a soft abnormality.

[0023] As a preferred embodiment of the present application, the process of executing operation according to the safety mechanism and the real-time feedback mechanism is as follows:

[0024] The standard control instruction received by the control module is in a structured format, which is mapped to the bottom layer control signal through the instruction parser, and the control signal is issued to the target device through the industrial bus;

[0025] After the control module completes the instruction issuing, real-time monitoring of the device response signal and state variable is performed; relying on the industrial edge control unit and the high-frequency data acquisition node, a set of core feedback parameters is collected within T seconds after the instruction execution; based on the data review module, the task ID and the parameter target interval are matched and verified, if the current adjustment is successful, the successful adjustment log is recorded, and the task control process is closed, and the monitoring standby state is restored; otherwise, the next round of abnormal detection, root cause positioning and strategy generation process is automatically entered.

[0026] In a second aspect, the present application provides a production line adjustment control method based on artificial intelligence, which is used to realize the above-mentioned production line adjustment control system based on artificial intelligence, and the method comprises:

[0027] S001: Real-time collection of raw data of device running state, process parameters, voiceprint signals and environmental information, all raw data into embedded edge processing unit, complete filtering denoising, normalization, time alignment and feature extraction, unified into structured feature vector;

[0028] S002: Combined with the current production task ID, extract the standard process parameters, fuse the historical experience, device state and environmental variables to construct the composite feature vector; use the deviation learning network to dynamically generate the tolerance boundary, compare the real-time data to judge whether it is out of the safe range; if the abnormal boundary or confidence contraction mechanism is triggered, further combine the dynamic time warping and cosine similarity analysis of single source signal form;

[0029] S003: When no explicit abnormality is triggered, extract the multi-period encoding sequence, input the attention mechanism enhanced variational autoencoder, and get the hidden variable distribution; the distribution deviation between the current state and the normal state is quantified by KL divergence, and the continuous deviation is greater than the threshold value, which is identified as a soft abnormality;

[0030] S004: After the abnormality is confirmed, a state variable causal graph is constructed, a causal path is identified by applying a causal graph neural network, and a core influencing factor is locked; combined with the expert rule base and the historical similar case base, a plurality of strategy candidates are generated; the behavior code is built by a multilayer perception mechanism, the similarity between the current strategy and the historical efficient strategy is compared, and a priority strategy set is output;

[0031] S005: The candidate strategy is input into the SimFormer simulation model, and the current system state, abnormal parameters and strategy vector are dynamically simulated; based on the Encoder-Decoder architecture, the effects of each strategy under multi-dimensional indicators are extracted, a multi-dimensional score vector is output and standardized and weighted, and the optimal strategy is sorted out;

[0032] S006: Introduce resource constraints and benefit functions, solve Nash equilibrium to obtain the optimal non-conflict strategy combination; finally, the strategy combination is translated into bottom control instructions, which are issued to the equipment through the industrial bus for execution, and the execution effect is monitored in real time; if the adjustment target is not reached, it will automatically re-enter the abnormal detection process.

[0033] Compared with the prior art, the beneficial effects of the present application are:

[0034] 1、The present application constructs a multi-modal feature vector, introduces a bias learning network and an attention mechanism, dynamically infers the parameter tolerance boundary based on the current production task and equipment environment, so as to improve process adaptability and data reliability.

[0035] 2、The present application combines DTW, cosine similarity, attention-enhanced VAE and KL divergence analysis mechanism, which can not only detect explicit abnormalities, but also identify potential soft abnormalities and system state drift, realize abnormal forward identification and early warning intervention.

[0036] 3、The present application is based on CGNN causal graph modeling and Bayesian structure search, combined with expert knowledge base and historical similar task data, and realizes individualized and efficient automatic generation of control strategies through multilayer perceptron evaluation of strategy adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to facilitate the understanding of those skilled in the art, the present application will be further described below in conjunction with the drawings.

[0038] Figure 1 The present application is a module connection diagram;

[0039] Figure 2 The present application is a method step flow chart. DETAILED DESCRIPTION

[0040] The technical solutions of the present application will be described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0041] It should be understood that the terms "include" and "contain" used in the specification and claims of the present disclosure indicate the presence of the described features, whole, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or their sets.

[0042] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. As used in this disclosure and the claims, "a," "an," and "the" are intended to include both singular and plural forms, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this disclosure and the claims, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0043] Referring to Figure 1 As shown in the drawings, in one aspect, the present application provides an artificial intelligence-based production line adjustment control system, comprising: a data acquisition module, a data review module, an intelligent adjustment analysis module and a control module;

[0044] The data acquisition module is deployed at a key position of the production line and is used to collect data of the production line, including but not limited to: equipment operating state parameters, process parameters, voiceprint signals and production task information, etc. The collected data is first input to an embedded data processing unit for preliminary processing, which includes denoising, filtering, standardization, time synchronization and key feature extraction, etc. Then, a unified structured production line information is formed and transmitted to the data review module.

[0045] After receiving the production line information from the data acquisition module, the data review module processes it according to a pre-set verification process, which includes:

[0046] S1: According to the received production task, the corresponding standard process parameter set {u i} is extracted from the database according to the task number; then the historical task data of the same or similar type as the current task is retrieved to construct a composite feature vector: F(t) = [h i (t), s i (t), e i (t)], wherein: h i (t) is the deviation distribution of the i-th process parameter in the historical similar task; s i (t) represents the diagnostic features of the current equipment operating state (such as temperature, vibration, current, etc.); e i (t) represents the current environmental condition features (such as humidity, cleanliness, noise level, etc.);

[0047] Combined with the historical similar task records, equipment state and environmental factors, a deviation learning network (DLN) is called, which adopts a multi-layer neural network structure and has an attention mechanism built-in to weight the action strength of different influencing factors. The deviation learning network dynamically infers the tolerance boundary for each process parameter to obtain an adaptive process constraint model, and outputs the dynamic tolerance boundary value of the parameter under the current conditions ; compare the real-time collected production line information {x i} with the dynamic tolerance boundary value output by the above model, and judge whether the inequality constraint condition is met: If all process parameters meet the above inequality constraint condition, it indicates that the current production state is within the acceptable process fluctuation range; otherwise, an abnormality flag is triggered; it should be noted that if the historical data exceeds the training domain of the current state, the DLN model activates the uncertainty measurement unit to output the confidence interval, and automatically tightens the tolerance boundary value to improve safety;

[0048] S2: Obtain the data corresponding to the abnormality flag, and perform double similarity check on the key single-source signals (such as temperature change sequence, current fluctuation signal):

[0049] First, use the dynamic time warping (DTW) algorithm to dynamically align the current time sequence with the standard template obtained by clustering normal working condition samples, and calculate the DTW distance; if the distance exceeds the set threshold, it is determined that there is a shape abnormality; second, after extracting the feature vectors of the current sequence and the standard template sequence, the angle similarity is calculated, and the cosine similarity is used as an index; if the similarity is lower than the threshold, it is determined that there is a trend deviation; if both indicators are determined to be abnormal, trigger the abnormality confirmation and enter the intelligent adjustment analysis process;

[0050] S3: For information that does not appear to be abnormal, input the feature encoding sequence {Z t-k+1 ,…,Z t} of the continuous k time steps to the attention-enhanced variational autoencoder (VAE) model, and the output hidden variable distribution parameters {m t} are output;

[0051] According to the hidden variable distribution parameters, use the KL divergence to measure the deviation between the current window and the normal hidden variable distribution: , where q(z) is the hidden variable distribution of the current window (the statistical result of the current state encoding), p(z) is the hidden variable distribution under normal state (usually obtained by statistical analysis of historical normal data), and extract the hidden variable vector for each time window; if the KL divergence of the continuous m time windows is higher than the set threshold τ, it indicates that there is a continuous potential drift; mark it as a soft abnormality and intervene before it is completely out of control;

[0052] The intelligent adjustment analysis module processes the abnormality identified by the data review module and generates a control strategy, the specific process is as follows:

[0053] Step 1: Perform multi-dimensional abnormality reason identification, construct a heterogeneous graph structure G(V, E) for process parameters, device feedback, and environmental variables, where each node v irepresents a state variable;

[0054] Causal learning using CGNN on graph structure G(V,E) and building causal graph; modeling causal strength by using intervention estimation method (such as do-calculus), outputting causal contribution score; using Bayesian structure search to strengthen causal direction identification, obtaining propagation chain path Pab: representing how the anomaly propagates from A variable to B variable; obtaining the causal path length of the anomaly node from the propagation chain path Pab; outputting the core root cause node according to the causal path length of the anomaly node and the causal contribution score (attention score in CGNN);

[0055] Step two: according to the root cause node corresponding variable, searching the expert rule base, extracting the matching strategy template, such as feeding delay, then executing: increasing the spindle speed by 10%, combining the similar task historical database, matching the adjustment records executed in similar scenes, inputting the historical data: abnormal category label, root cause node parameter and current environment state vector S, and outputting the historical multiple structured strategy candidate vectors a through the generator; judging whether the candidate strategy is close to the historical efficient strategy: concatenating the current system state, including task type, process condition, environmental disturbance factor and other context information through the generator to output the current multiple structured strategy candidate vectors d, extracting the joint feature vector h through the multi-layer perception (MLP) as the behavior code of the strategy in the current state; the difference between the strategy current state behavior code and the historical state behavior code is output through a mapping function, C∈[0,1], when the difference value C tends to 0, it means that the strategy is close to the historical efficient strategy; otherwise, it means that the strategy is far away from the historical efficient strategy, and the expert rule base is searched again to extract the matching strategy template;

[0056] The structured strategy is represented as: adjustment object and parameter; expected target index (such as error reduction of 5%); potential side effect (such as energy consumption increase), recommended applicable condition;

[0057] Step three: simulation verification is performed on the final strategy;

[0058] Through the SimFormer scorer, the current system state, abnormal parameter and candidate strategy vector are concatenated into a time series input as the SimFormer input token stream;

[0059] When the input token stream is input, the influence weight of each parameter on the simulation index is extracted based on the Encoder-Decoder structure, and a multi-dimensional simulation index vector F is output, including: product quality improvement amplitude, energy consumption change, stability score, execution success probability, potential risk level, etc.

[0060] According to the output multi-dimensional simulation index vector F, each dimension of the simulation index is converted to a 0-1 interval using a normalization function, and a weighted benefit function is used to obtain a multi-dimensional simulation index standard value. According to the size of the multi-dimensional simulation index standard value, sorting is performed and a sorting label is marked.

[0061] Step four: combine the strategy as a multi-objective resource game between multiple agents;

[0062] Each strategy has its own benefit function U (based on SimFormer output multi-dimensional simulation index vector F, resource usage and side effects), a non-cooperative game model is constructed, and the constraint conditions are resource limitations, parameter conflict matrix, etc. The solution outputs the optimal non-conflict combination:

[0063] The Nash equilibrium solution is used to output the optimal strategy set zy, and the selected strategy is mapped to the standard control instruction.

[0064] After the control module receives the standard control instruction, it executes the operation according to the preset safety mechanism and real-time feedback mechanism:

[0065] The standard control instruction received by the control module is in a structured format, and the control content is mapped to the bottom control signal (such as PLC code segment, Modbus instruction or OPC UA communication instruction) through the instruction parser, and the control signal is issued to the target device relying on the industrial bus (such as CAN, EtherCAT, PROFINET);

[0066] After the control module completes the instruction issuance, it listens to the device response signal and state variable (such as temperature, pressure, speed and power, etc.); relying on the industrial edge control unit and high-frequency data acquisition node, the core feedback parameter set is collected within T seconds after the instruction execution; based on the data review module, task ID and parameter target interval, the collected data is matched and verified (such as whether the set temperature is reached, whether it is within the allowed fluctuation range), if the current adjustment is successful, the successful adjustment log is recorded, and the task control process is closed, and the monitoring standby state is restored; otherwise, it automatically enters the next round of abnormal detection → root cause positioning → strategy generation process, that is, the process of data review module and intelligent adjustment analysis module.

[0067] Please refer to Figure 2 As shown in the figure, the other aspect provides an artificial intelligence-based production line adjustment control method, comprising:

[0068] S001: Deploy multi-modal perception units at key positions on the production line to collect device operating parameters, process data, voice signals and environmental information in real time. All raw data are first input into the embedded processing unit to perform filtering and denoising, standardization, time alignment, feature extraction and other preprocessing operations, and are converted into structured production line information in a unified format.

[0069] S002: After receiving the structured production information, the standard process parameter set is extracted according to the current task number, and a composite feature vector integrating historical experience, equipment state and environmental characteristics is constructed. A deviation learning network (DLN) is introduced to dynamically generate an adaptive tolerance model, and the real-time parameters are compared with it. If it is found that the parameters exceed the dynamic boundary or trigger the confidence interval contraction mechanism, it is marked as abnormal; further combined with dynamic time warping (DTW) and cosine similarity calculation, the abnormal morphology of single source signal is verified, and the accurate labeling of abnormality is completed.

[0070] S003: For the case where no explicit abnormality is triggered, the feature encoding sequence within the time window is constructed and input into the attention-enhanced variational autoencoder (VAE) model to obtain the current latent variable distribution. Then the KL divergence is used to quantify the deviation of the current state from the normal state distribution, and if the deviation of multiple consecutive time windows is higher than the set threshold, it is identified as a soft abnormality, indicating a potential performance drift risk, which provides a basis for early warning intervention.

[0071] S004: After the abnormality is confirmed, a state variable causal graph is constructed, and a causal path reasoning is performed based on the causal graph neural network (CGNN) to identify the core root cause variable that triggered the abnormality. A preliminary strategy candidate set is generated by combining the expert rule base and the historical similar task database. A joint feature representation of historical behavior encoding and current strategy behavior encoding is constructed through a multi-layer perceptron (MLP), and the strategy difference value is compared to select the priority candidate strategy close to the efficient historical strategy.

[0072] S005: The generated candidate strategy is dynamically scored through the SimFormer simulation model, which inputs include system state, abnormal parameters and strategy vector, etc. Based on the Encoder-Decoder structure, the influence intensity of the strategy in different dimensions (such as product quality, energy consumption, stability, etc.) is extracted, and a multi-dimensional simulation index vector F is output. After standardization and weighted sorting, the strategy scheme with the optimal multi-dimensional benefit index is selected.

[0073] S006: The top-ranked strategy is considered as a non-cooperative game participated by multiple agents. By constructing the payoff function and resource constraint conditions, the Nash equilibrium is solved to obtain the optimal non-conflict strategy combination. The system translates this strategy combination into bottom-level control instructions and issues them to the target equipment to execute the adjustment task. At the same time, the feedback state is monitored to verify whether the adjustment target is achieved. If the target is not achieved, the system automatically enters a new round of abnormality detection and intelligent adjustment process, realizing closed-loop adaptive optimization control.

[0074] The preferred embodiments of the application disclosed above are only to facilitate the elucidation of the application. The preferred embodiments do not describe all the details of the application and limit the application to the specific embodiments. Obviously, many modifications and variations can be made in light of the teachings above. The description is chosen and described in order to provide the best illustration of the application and its practical application to those skilled in the art and to enable those skilled in the art to best utilize the application. The application is limited only by the claims and their full scope and equivalents.

Claims

1. An artificial intelligence-based production line adjustment control system, comprising a data review module, an intelligent adjustment analysis module, and a control module; characterized in that: The data review module receives production line information from the data acquisition module, verifies it according to the set process, and triggers anomaly marking, anomaly confirmation, and soft anomaly. The intelligent adjustment and analysis module processes anomalies identified by the data review module, including: To identify multidimensional anomalies, CGNN is used to perform causal learning on the graph structure G(V,E) to construct a causal graph. Then, an intervention estimation method is used to model the causal strength and output the causal contribution score. Bayesian structure search is used to enhance the identification of causal directionality and obtain the propagation chain path Pab to obtain the causal path length of the anomaly node. Based on the causal path length and causal contribution score of the anomaly node, the core root cause node is output. Based on the root cause node, retrieve the expert rule base and similar task database to extract the historical policy vector set; combine the current state vector and use a generator to construct policy candidate vectors; concatenate the historical policy vector set and the constructed policy candidate vectors, extract the joint feature vector through a multilayer perceptron, compare the historical behavior encoding difference value C∈[0,1], analyze the closeness between the policy and the efficient policy, and output the structured policy. Using the SimFormer model, state, anomaly parameters, and policy vectors are concatenated into an input token stream; a multi-dimensional simulation index vector is output based on the Encoder-Decoder structure; each dimension of the multi-dimensional simulation index vector is normalized, a weighted profit function is used to calculate standardized index values, and then the values ​​are sorted and labeled. The strategy combination is viewed as a multi-objective resource game among multiple agents; each strategy has its own payoff function U; a non-cooperative game model is constructed, the optimal non-conflict combination is solved, and the selected strategy is mapped to a standard control instruction. After receiving standard control commands, the control module executes the operation based on the safety mechanism and real-time feedback mechanism.

2. The production line adjustment control system based on artificial intelligence according to claim 1, characterized in that, It also includes a data acquisition module, which is deployed at a fixed location on the production line to collect data from the production line, including equipment operating status parameters, process parameters, voiceprint signals, and production task information. The collected data is input to the embedded data processing unit for preliminary processing to form a unified structured production line information, and then transmitted to the data review module.

3. The production line adjustment control system based on artificial intelligence according to claim 1, characterized in that, The verification process, which follows the established procedure, triggers the exception flag as follows: Upon receiving a production task, the corresponding set of standard process parameters {u} is retrieved from the database based on the task number. i }; Retrieve historical task data that is the same or similar to the current task type, and construct a composite feature vector: F(t); Combine historical records of similar tasks, equipment status, and environmental factors, call the bias learning network, dynamically infer the tolerance boundary for each process parameter, perform an adaptive process constraint model, and output the dynamic tolerance boundary value of the parameter under the current conditions; Combine the real-time collected production line information {x i The dynamic tolerance boundary value output by the model is compared to determine whether the inequality constraint conditions are met. If not, an anomaly flag is triggered.

4. The production line adjustment control system based on artificial intelligence according to claim 3, characterized in that, The verification process, which follows the established procedure, triggers an exception confirmation as follows: The dynamic time warping algorithm is used to dynamically register the current time series with the standard template obtained by clustering samples from normal working conditions, and calculate their DTW distance; if the distance exceeds the set threshold, it is determined that there is a morphological abnormality; after extracting feature vectors from the current sequence and the standard template sequence, the angle similarity between them is calculated, and cosine similarity is used as the index. If the similarity is below the threshold, a trend deviation is determined; if both indicators are determined to be abnormal, an anomaly confirmation is triggered.

5. The production line adjustment control system based on artificial intelligence according to claim 4, characterized in that, The verification process, which follows the established procedure, triggers a soft exception as follows: For information that does not show any anomalies, the feature encoding sequence {Z} for k consecutive time steps is... t-k+1 ,…,Z t After being input into the attention-enhanced variational autoencoder model, the output latent variable distribution parameters {m} are obtained. t }; The KL divergence measure is used to measure the deviation of the current window from the normal latent variable distribution based on the latent variable distribution parameters: Where q(z) is the latent variable distribution of the current window, and p(z) is the latent variable distribution under normal conditions. Extract latent variable vectors for each time window; if the KL divergence of m consecutive time windows is higher than the set threshold τ, it is determined that there is a continuous potential drift and it is marked as a soft anomaly.

6. The production line adjustment control system based on artificial intelligence according to claim 1, characterized in that, Based on the security mechanism and real-time feedback mechanism, the operation process is as follows: The standard control commands received by the control module are in a structured format. The command parser maps the control content into low-level control signals and sends the control signals to the target device via the industrial bus. After the control module issues the command, it monitors the device's response signals and status variables in real time. Relying on the industrial edge control unit and high-frequency data acquisition node, it collects the core feedback parameter set within T seconds after the command is executed. Based on the data review module connection, the task ID and parameter target range are matched and verified. If the adjustment is successful, the successful adjustment log is recorded, the task control process is closed, and the system is restored to the monitoring standby state. Otherwise, it automatically enters the next round of anomaly detection, root cause localization, and strategy generation process.

7. A production line adjustment control method based on artificial intelligence, characterized in that... Implementing an artificial intelligence-based production line adjustment control system as described in any one of claims 1-6, the method comprises: S001: Real-time acquisition of raw data on equipment operating status, process parameters, acoustic signals and environmental information, inputting all raw data into the embedded edge processing unit to complete filtering, noise reduction, normalization, time alignment and feature extraction, and uniformly converting them into structured feature vectors; S002: Combine the current production task ID, extract standard process parameters, and integrate historical experience, equipment status and environmental variables to construct a composite feature vector; use the bias learning network to dynamically generate tolerance boundaries, and compare real-time data to determine whether it exceeds the safe range; if an abnormal boundary or confidence contraction mechanism is triggered, further combine dynamic time warping and cosine similarity analysis of the single-source signal morphology. S003: When no obvious anomaly is triggered, extract the multi-time period encoding sequence, input it into the variational autoencoder with attention mechanism enhancement, and obtain the latent variable distribution; measure the distribution deviation between the current state and the normal state through KL divergence, and identify soft anomalies if the continuous deviation exceeds the threshold. S004: After anomaly confirmation, construct a causal graph of state variables, apply a causal graph neural network to identify causal paths, and lock in core influencing factors; combine expert rule base and historical similar case base to generate multiple policy candidates; construct behavior encoding through multilayer perceptron, compare the similarity between the current policy and historical efficient policies, and output a set of preferred policies; S005: Input the candidate strategies into the SimFormer simulation model, and perform dynamic simulation by combining the current system state, abnormal parameters and strategy vectors; extract the effects of each strategy under multi-dimensional indicators based on the Encoder-Decoder architecture, output multi-dimensional score vectors and perform standardized weighting, and rank them to obtain the optimal strategy; S006: Introduce resource constraints and revenue functions to solve the Nash equilibrium and obtain the optimal non-conflict strategy combination; finally, translate this strategy combination into low-level control commands, send them to the equipment for execution through the industrial bus, and monitor the execution effect in real time; if the adjustment target is not achieved, automatically re-enter the anomaly detection process.

Images