Method and system for detecting internal defects of forging rotor shell

Abstract

The invention relates to a method and a system for detecting internal defects of a forge piece rotor shell, in particular to the field of detection of the internal defects of the forge piece rotor shell, and realizes accurate prediction of the defects of the forge piece rotor shell through multi-source data fusion and dynamic risk feature extraction. The hybrid neural network model based on sequence mining can capture the time sequence dependency relationship in the production process, and the reliability of defect probability prediction is improved; detection scheduling is modeled as a multi-agent reinforcement learning problem, a dynamic optimization detection instruction is generated, and the detection efficiency and risk control are effectively balanced; the model and the strategy are continuously fine-tuned by using an online learning mechanism, so that the system has adaptive evolution capability and can dynamically respond to the change of the production environment, and finally, the defect detection accuracy, the scheduling efficiency and the overall robustness of the system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of internal defect detection of forged rotor housings, and more specifically, to a method and system for detecting internal defects of forged rotor housings. Background Technology

[0002] Forged rotor housings, as key pressure-bearing components of large rotating machinery, are widely used in high-end equipment manufacturing fields such as energy, aviation, and heavy machinery. Their internal quality directly affects the safety and reliability of the entire machine. In modern large-scale manufacturing plants, the production of such components is usually carried out in a continuous or batch mode, with fast production line pace and high capacity. During the production process, the formation of internal defects in the rotor housing, such as inclusions, porosity, and cracks, is affected by multiple factors. Internal factors include fluctuations in the alloy composition of raw materials, instantaneous changes in forging process parameters such as temperature and pressure, and microstructure evolution during heat treatment. External factors include changes in temperature and humidity in the production environment, the tightness of the connection between upstream and downstream processes, and frequent adjustments to production plans due to dynamic fluctuations in market order demand. To ensure quality, factories need to regularly use non-destructive testing equipment such as ultrasonic and X-ray to inspect the internal defects of the rotor housing. However, in actual operation, the total amount of high-value testing equipment resources is limited and expensive, and there is also a relative shortage of qualified testing personnel. This contradiction between high-paced production and limited testing capabilities constitutes a typical manufacturing management bottleneck.

[0003] Currently, the industry commonly employs static scheduling schemes based on fixed cycles or empirical rules for scheduling and managing the detection of internal defects in forged rotor housings. These existing technologies typically rely on pre-set detection intervals or operator experience to arrange detection tasks. For example, some systems use simple linear programming to allocate detection resources at specific time points or set thresholds based on historical average defect rates for early warning. However, these methods have significant shortcomings. First, they lack dynamic perception of defect generation mechanisms and data-driven predictive capabilities, failing to effectively respond to instantaneous changes in defect rates caused by material batch changes or process parameter drift. Second, static scheduling algorithms cannot integrate real-time production line status data, such as equipment operation data. External disturbances such as production line turnover, work-in-process inventory levels, and the insertion of emergency orders cause a serious disconnect between the inspection plan and the actual production rhythm. Third, existing technical solutions usually view inspection tasks in isolation and fail to deeply integrate defect prediction, resource allocation, and production scheduling from a system optimization perspective. Fourth, traditional algorithms often fall short when facing multi-objective optimization, such as simultaneously pursuing inspection timeliness, equipment utilization, and minimizing the risk of false positives and false negatives. They are computationally complex and have poor adaptability. The core problem is the lack of an intelligent management method that can accurately predict the risk of defect occurrence based on real-time multi-source data and dynamically optimize the inspection scheduling strategy accordingly. This technical shortcoming directly restricts the overall efficiency and quality control level of the production line. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a method and system for detecting internal defects in the casing of a forged rotor, thereby resolving the issues raised in the background section.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for detecting internal defects in a forged rotor shell, specifically including the following steps: Step S1: During the production process of the forged rotor shell, based on the triggering event of a new workpiece entering the production line, multi-source data of the forged rotor shell to be inspected is collected in real time. The multi-source data includes production parameters, material spectrum, equipment status, historical defect records, inspection equipment status and order information. The multi-source data is fused and processed using a deep feature extraction network based on an attention mechanism to generate a corresponding dynamic risk feature vector for each forged rotor shell to be inspected. Step S2: Based on the dynamic risk feature vectors corresponding to multiple rotor shells of forgings to be inspected generated in step S1, arrange them in order of their production time to form a risk feature sequence. Use a hybrid neural network model of GRU and Transformer encoder to mine the risk feature sequence and output the defect occurrence probability function of each rotor shell of forging to be inspected. Step S3: Based on the defect occurrence probability function obtained in step S2 and the detection equipment status and order information obtained from the multi-source data in step S1, the detection scheduling problem is modeled as a Markov decision process, and a multi-agent near-end policy optimization algorithm is applied to optimize the detection decision with a hierarchical reward function, generating a dynamic detection scheduling instruction set to execute the detection job, thereby obtaining measured data containing defect judgment results. Step S4: During the execution of the dynamic detection scheduling instruction set generated in step S3 by the detection system, based on a preset rolling time window, the measured data generated by the detection operation and the production parameters, material spectrum, and equipment status from the current multi-source data collected in real time from the production line are collected; using an online learning mechanism, the decision strategies of the hybrid neural network model described in step S2 and the multi-agent proximal strategy optimization algorithm described in step S3 are incrementally fine-tuned based on the collected measured data and the production parameters, material spectrum, and equipment status from the current multi-source data, so as to realize the continuous self-evolution of the hybrid neural network model and the decision strategy; In a preferred embodiment, the specific process of real-time acquisition of multi-source data of the forged rotor shell to be inspected in step S1 includes: When a trigger event for a new workpiece entering the production line is detected, the multi-source data synchronous acquisition process is immediately initiated. The production parameters in the multi-source data refer to process variables directly related to the processing of the new workpiece, obtained from the manufacturing execution system, including forging temperature, pressure value, and cooling rate. The material spectrum refers to the attribute information related to the raw materials of the new workpiece extracted from the enterprise resource planning system, including steel furnace number, chemical composition ratio, supplier batch code and raw material supplier identifier; The equipment status refers to the operating indicators of the equipment when processing the new workpiece, which are read in real time from the sensor network, including machine tool vibration amplitude, current wear of the mold, equipment operating frequency, and temperature reading; The historical defect record refers to the defect information of the workpieces that have been detected in the recent production sequence, which is retrieved from the quality database. It includes the defect type, defect severity, defect location and corresponding production condition context. The status of the testing equipment refers to the resource status variables that can be used for non-destructive testing, obtained from the testing resource management unit, including the number of available ultrasonic flaw detection equipment, the health status indicators of ultrasonic flaw detection equipment, the equipment calibration cycle, and the expected idle time. The order information refers to the delivery urgency and customer level identifier of the order to which the new workpiece belongs, retrieved from the order management unit; After all multi-source data collection is completed, data preprocessing and alignment operations are performed, including cleaning, denoising and standardizing numerical data, numerical encoding of enumerated data, and aligning all data to the production time point of the new workpiece using the workpiece's unique identifier and timestamp, forming a unified multidimensional data snapshot bound to the new workpiece.

[0006] In a preferred embodiment, the specific process of fusing multi-source data using an attention-based deep feature extraction network includes: First, each type of data in the preprocessed multi-source data is mapped to a feature sub-vector of a unified dimension through an independent embedding layer, including production parameter feature sub-vector, material spectrum feature sub-vector, equipment status feature sub-vector, historical defect record feature vector, detection equipment status feature vector, and order information feature vector; Then, all feature vectors are concatenated to form a preliminary overall feature vector. Next, a trainable scoring network is used to compute the saliency weight of each feature subvector, wherein for each feature subvector, the scoring network performs the following operations: The overall feature vector and the feature sub-vector are taken as input. Linear transformations are performed on the overall feature vector and the feature sub-vector respectively. The transformation results are added together and a first non-linear activation function is applied to obtain an intermediate feature representation. The intermediate feature representation is then multiplied by a trainable parameter vector and a second non-linear activation function is applied to obtain the initial score of the feature sub-vector. Then, the initial scores corresponding to all feature sub-vectors are subjected to exponential calculation and summation normalization to obtain the normalized attention weight of each feature sub-vector. Finally, using the calculated normalized attention weight value for each feature category, the production parameter feature vector, material spectrum feature vector, equipment status feature vector, historical defect record feature vector, detection equipment status feature vector, and order information feature vector are weighted respectively, and all weighted feature vectors are summed to generate the final dynamic risk feature vector.

[0007] In a preferred embodiment, step S2, specifically the process of arranging the risk characteristics into a sequence according to their production time, includes: The dynamic risk feature vectors of multiple rotor shells of forgings to be tested obtained from step S1 are sorted according to the timestamp of each workpiece being processed on the production line to form a continuous risk feature sequence in the time dimension, and each position in the risk feature sequence corresponds to a workpiece produced at a certain time point.

[0008] In a preferred embodiment, the specific process of mining the risk feature sequence using a hybrid neural network model employing GRU and Transformer encoders includes: First, the risk feature sequence is input as a whole into the GRU layer for processing. The GRU layer processes the dynamic risk feature vector of each workpiece in the sequence step by step through the synergistic effect of its internal update gate and reset gate, and outputs a hidden state sequence containing local temporal information. Each hidden state vector in the hidden state sequence encodes the local sequence information up to the corresponding time step. Next, the hidden state sequence is input as a whole into the Transformer encoder layer. The Transformer encoder layer uses a self-attention mechanism to calculate the association weight between the hidden state vectors of any two time steps in the sequence and outputs a feature representation sequence enhanced by global context information. This enhanced feature representation sequence consists of the enhanced feature vector of each artifact, where each enhanced feature vector incorporates global context information. Finally, the enhanced feature vector corresponding to each workpiece is input into a fully connected layer. This fully connected layer performs a linear transformation operation, which includes performing a dot product operation between the enhanced feature vector and the trainable weight vector and adding a trainable bias value. Then, a nonlinear transformation is performed through a sigmoid activation function. The result of adding the dot product operation and the bias value is mapped to a defect occurrence probability value between zero and one, thereby obtaining the defect occurrence probability function of the rotor shell of each forging to be inspected.

[0009] In a preferred embodiment, step S3, which involves modeling the detection scheduling problem as a Markov decision process, includes the following specific steps: First, based on the defect occurrence probability function of each rotor shell of the forging to be inspected obtained in step S2 and the real-time detection equipment status and order information obtained from the multi-source data in step S1, the state space of the Markov decision process is constructed. The state space is composed of the defect occurrence probability function of all work-in-process, the status of the detection equipment, the order information, and the additional environmental status; and the state vector of the state space at each time step is formed by sequentially connecting the above-mentioned defect occurrence probability function, detection equipment status, order information, and additional environmental status to form a unified dynamic environment representation. Secondly, define the action space, which is a discrete set, where each action represents a specific detection scheduling decision; Then, a hierarchical reward function is designed, which is a linear combination of three factors: efficiency reward, risk reward, and cost penalty, through weighted coefficients. Finally, the state transition model is dynamically determined by the production environment, including workpiece movement and equipment state changes.

[0010] In a preferred embodiment, the specific process of applying the multi-agent proximal policy optimization algorithm includes: Each detection device is treated as an agent, and each agent maintains a policy network and a value network. Agents collaborate through centralized training and distributed execution. During training, each agent selects an action based on its current state, and the environment returns a reward value based on a hierarchical reward function and transitions to a new state. The training objective is to update the parameters of the policy network and the value network by maximizing the expected cumulative discount reward, where the expected cumulative discount reward is obtained by summing the reward values ​​of all future time steps multiplied by the power of the discount factor. During the real-time scheduling phase, for the real-time state at each decision moment, each agent outputs an action probability distribution based on the trained policy network, and selects a specific detection and scheduling action from the probability distribution through a sampling method. The sampling actions of all intelligent agents are combined to form a dynamic detection scheduling instruction set. The dynamic detection scheduling instruction set is parsed into specific control instructions to drive the ultrasonic flaw detection equipment to perform actual detection operations, thereby obtaining measured data containing defect judgment results. The measured data includes defect type, defect size, and defect location information.

[0011] In a preferred embodiment, step S4, which involves collecting data based on a preset rolling time window, specifically includes: Within the rolling time window, two types of data are collected simultaneously: the first type is the measured data generated by the inspection operation; the second type is the current multi-source data collected in real time from the production line, which includes the production parameters, material spectrum, and equipment status of the workpieces newly entering the production line within the rolling time window. The collected measured data and the production parameters, material spectrum and equipment status in the current multi-source data are preprocessed, including data cleaning, data standardization and data alignment. The preprocessed data is then organized into a structured training sample set. Each sample contains input features and output labels. The input features are derived from the production parameters, material spectrum and equipment status in the current multi-source data, and the output labels are derived from the defect judgment results in the measured data.

[0012] In a preferred embodiment, the specific process of incremental fine-tuning using an online learning mechanism includes: Based on the measured data collected and preprocessed by the rolling time window, and the production parameters, material spectrum and equipment status in the current multi-source data, the decision-making strategies of the hybrid neural network model in step S2 and the multi-agent proximal strategy optimization algorithm in step S3 are collaboratively and incrementally fine-tuned using an online learning mechanism. For hybrid neural network models, the fine-tuning process aims to minimize the difference between the model's predicted defect probability and the actual defect label in the measured data. This is achieved by calculating the cross-entropy loss function, which is specifically calculated as the average loss of all training samples in the training sample set. The loss of each training sample is the sum of the product of the sample's actual defect label and the natural logarithm of the model's predicted defect probability. For the decision-making strategy of the multi-agent proximal policy optimization algorithm, the fine-tuning process aims to maximize the long-term cumulative reward by calculating the policy gradient loss function. The calculation of the policy gradient loss function is based on the average loss of all training samples in the training sample set, where the loss of each training sample is composed of the product of the advantage function value and the natural logarithm of the action probability output by the policy network. The total loss function is obtained by multiplying the first weight coefficient by the loss function of the hybrid neural network model and the second weight coefficient by the loss function of the decision strategy. Based on the total loss function, the gradient of the total loss function with respect to the model parameters and policy parameters is calculated, and the parameters are updated using a gradient descent algorithm with momentum and adaptive learning rate. The parameter update process is as follows: add the adaptive learning rate multiplied by the gradient of the loss function to the parameter value at the current time, and add the momentum coefficient multiplied by the difference between the parameter value at the current time and the parameter value at the previous time to obtain the updated parameter value.

[0013] This application also provides a detection system for internal defects in a forged rotor housing, specifically including: Data acquisition module: After receiving the trigger signal of a new workpiece entering the production line, it collects multi-source data of the forging rotor shell in real time. The multi-source data includes production parameters, material spectrum, equipment status, historical defect records, detection equipment status and order information. The collected multi-source data is sent to the feature extraction module. Feature extraction module: Based on the received multi-source data, the deep feature extraction network based on the attention mechanism is used to fuse the multi-source data and generate a corresponding dynamic risk feature vector for each forging rotor shell. The dynamic risk feature vector is then passed to the prediction module. Prediction module: Receives dynamic risk feature vectors of multiple forged rotor shells, arranges them in order of their production time to form a risk feature sequence, and uses a hybrid neural network model of GRU and Transformer encoder to mine the risk feature sequence and outputs the defect occurrence probability function of each forged rotor shell. Scheduling Decision Module: Based on the defect occurrence probability function, the status of the detection equipment, and order information, the detection scheduling problem is modeled as a Markov decision process. A multi-agent proximal policy optimization algorithm is applied to optimize the detection decision with a hierarchical reward function, generating dynamic detection scheduling instructions and sending them to the detection execution module. The detection execution module executes dynamic detection scheduling instructions to control the operation of the detection system. It collects measured data and current production data based on a preset rolling time window, and uses an online learning mechanism to incrementally fine-tune the hybrid neural network model and decision-making strategy to achieve continuous self-evolution of the hybrid neural network model and decision-making strategy.

[0014] The beneficial effects of this invention are as follows: by fusing multi-source data and extracting dynamic risk features, accurate prediction of defects in the outer shell of forged rotors is achieved; the hybrid neural network model based on sequence mining can capture the temporal dependencies in the production process, improving the reliability of defect probability prediction; by modeling the detection scheduling as a multi-agent reinforcement learning problem, dynamically optimizing detection instructions is generated, effectively balancing detection efficiency and risk control; and by continuously fine-tuning the model and strategy using an online learning mechanism, the system has adaptive evolution capabilities, can dynamically respond to changes in the production environment, and ultimately significantly improves the accuracy of defect detection, scheduling efficiency, and overall system robustness. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1 The method for detecting internal defects in a forged rotor housing, as shown, specifically includes the following steps: Step S1: During the production of the forged rotor shell, based on the triggering event of a new workpiece entering the production line, multi-source data of the forged rotor shell to be inspected is collected in real time. The multi-source data includes production parameters, material spectrum, equipment status, historical defect records, inspection equipment status and order information. The multi-source data is fused and processed using a deep feature extraction network based on the attention mechanism to generate a corresponding dynamic risk feature vector for each forged rotor shell to be inspected. Step S2: Based on the dynamic risk feature vectors corresponding to multiple rotor shells of forgings to be inspected generated in step S1, arrange them in order of their production time to form a risk feature sequence. Use a hybrid neural network model of GRU and Transformer encoder to mine the risk feature sequence and output the defect occurrence probability function of each rotor shell of forging to be inspected. Step S3: Based on the defect occurrence probability function obtained in Step S2 and the detection equipment status and order information obtained from the multi-source data in Step S1, the detection scheduling problem is modeled as a Markov decision process, and a multi-agent near-end policy optimization algorithm is applied to optimize the detection decision with a hierarchical reward function, generating a dynamic detection scheduling instruction set to execute the detection job, thereby obtaining the measured data containing the defect judgment results. Step S4: During the execution of the dynamic detection scheduling instruction set generated in step S3 by the detection system, based on the preset rolling time window, the system collects the measured data generated by the detection operation and the production parameters, material spectrum, and equipment status from the current multi-source data collected from the production line in real time. Using an online learning mechanism, the system incrementally fine-tunes the decision strategies of the hybrid neural network model in step S2 and the multi-agent proximal strategy optimization algorithm in step S3 based on the collected measured data and the production parameters, material spectrum, and equipment status from the current multi-source data, thereby realizing the continuous self-evolution of the hybrid neural network model and the decision strategy.

[0020] In this embodiment, the specific process of real-time acquisition of multi-source data of the forged rotor shell to be tested in step S1 includes: When a trigger event for a new workpiece entering the production line is detected, the multi-source data synchronous acquisition process is immediately initiated. The production parameters in the multi-source data refer to the process variables directly related to the processing of the new workpiece, including forging temperature, pressure value and cooling rate, obtained from the manufacturing execution system (the manufacturing execution system is used to record and store the forging temperature, pressure and cooling rate of each workpiece in real time; this application automatically retrieves the above production parameters when the workpiece enters the production line through its open API interface). Material spectrum refers to the attribute information related to the raw materials of the new workpiece extracted from the enterprise resource planning system (which stores complete material master data and sales order data; this application queries and associates the raw material batches used in the current workpiece and the priority information of their respective customer orders by connecting to the database of the enterprise resource planning system). This includes steel furnace number, chemical composition ratio, supplier batch code, and raw material supplier identifier. Equipment status refers to the operating indicators of the equipment when processing the new workpiece, which are read in real time from the sensor network (the sensor network consists of vibration accelerometers and temperature thermocouples deployed in the machining center, which collect equipment operation data in real time in a high-frequency sampling manner and transmit it through the Industrial Internet of Things protocol). These indicators include machine tool vibration amplitude, current wear of mold, equipment operating frequency and temperature reading. Historical defect records refer to the defect information of inspected workpieces in the recent production sequence queried from the quality database (which stores the final defect records of all inspected workpieces in a time-series structure and associates them with historical production parameters in the manufacturing execution system to form a 'defect-process' mapping dataset for model training). This includes defect type, defect severity, defect location, and corresponding production condition context. The production condition context includes the historical production parameters and historical environmental data corresponding to the inspected workpiece during production. The historical production parameters include forging temperature, pressure value, and cooling rate, while the historical environmental data includes workshop ambient temperature and humidity. The status of the testing equipment refers to the resource status variables that can be used for non-destructive testing obtained from the testing resource management unit (which is integrated into the scheduling system of this application, continuously monitors the online / offline status, current task queue, and estimated completion time of each ultrasonic testing equipment through a heartbeat mechanism, and calculates its health score based on the equipment maintenance records, thereby dynamically maintaining an available testing resource pool). These variables include the number of available ultrasonic flaw detection equipment, the health status indicators of ultrasonic flaw detection equipment, the equipment calibration cycle, and the estimated idle time. Order information refers to the delivery urgency and customer level identifier of the order to which the new workpiece belongs, retrieved from the order management unit (the order management unit assigns a dynamic priority coefficient to each order, which is determined by the delivery date, customer level, and delay risk. This coefficient is read by this application and used as an important input for scheduling optimization). After collecting all multi-source data, data preprocessing and alignment operations are performed, including cleaning, denoising and standardizing numerical data to unify the units, numerical encoding of enumerated data, and strictly aligning all data to the production time point of the new workpiece using the workpiece's unique identifier and timestamp, forming a unified multi-dimensional data snapshot bound to the new workpiece, providing a foundation for subsequent feature fusion. The specific process of fusing multi-source data using attention-based deep feature extraction networks includes: First, each type of data in the preprocessed multi-source data is mapped to a feature sub-vector of a unified dimension through an independent embedding layer or a fully connected layer. These feature sub-vectors include production parameter feature sub-vectors, material spectrum feature sub-vectors, equipment status feature sub-vectors, historical defect record feature sub-vectors, detection equipment status feature sub-vectors, and order information feature sub-vectors. Each feature sub-vector is converted into a numerical representation of a fixed dimension through a neural network layer. Then, all feature sub-vectors are concatenated to form a preliminary overall feature vector, which contains combined information of all data categories; Next, a trainable scoring network is used to compute the saliency weights of each feature subvector, where, for each feature subvector, the scoring network performs the following operations: Taking the overall feature vector and the feature sub-vector as input, linear transformations are performed on both. The transformation results are summed and then a first nonlinear activation function is applied to obtain an intermediate feature representation. This intermediate feature representation is then multiplied by a trainable parameter vector, and a second nonlinear activation function is applied to obtain the initial score for the feature sub-vector. Finally, the initial scores for all feature sub-vectors are exponentially calculated and normalized to obtain the normalized attention weight for each feature sub-vector. The formula for calculating the normalized attention weight is as follows: ; in, Indicates the first The attention weights for the class feature vectors, which are the output of the formula, are numerical values ​​between 0 and 1, representing the importance of the i-th class feature vector (such as "production parameters") in the overall risk assessment. All weights... The sum equals 1. and The index representing the category of the feature sub-vector. This represents the feature category whose weight is currently being calculated; This represents all feature categories traversed in the summation operation, for example, It can represent six categories: {P (production parameters), M (material profile), E (equipment status), H (historical defects), D (inspection equipment), and O (order information)}. σ represents the activation function, such as LeakyReLU, which introduces a non-linear transformation to enhance the model's expressive power. In the formula, σ is applied to the output of the tanh function to further adjust the numerical range and avoid linear limitations. exp represents the exponential function, used to exponentially process the attention scores, ensuring all scores are positive for subsequent normalization (through the Softmax operation). This is a standard step in the attention mechanism, aiming to transform the scores into a probability distribution. Let represent the transpose of a trainable parameter vector, where 'a' is a vector used to compute the attention score, and its dimension matches the output dimension of the tanh function. The dot product operation... represents the vector inner product, used to generate a scalar score. tanh represents the hyperbolic tangent activation function, which compresses the input values ​​to the interval [-1, 1]. Here, it is used to process the feature vector after linear transformation, ensuring numerical stability and nonlinear interaction. This represents a trainable parameter matrix used for linear transformations. Acting on the overall feature vector , Acting on a single feature vector These mappings map the input vector to a new dimension to compute attention scores. The dimensions of the matrix are determined by the network structure and are typically learned during training. This represents the initial overall feature vector, composed of all six types of feature sub-vectors ( It is composed of multiple parts, representing the global production context information of the current workpiece, and serves as the reference basis for attention weight calculation. Indicates the first The feature vectors corresponding to class feature sub-data (e.g., P). This represents the bias vector, used to adjust the result of the linear transformation (i.e. The bias term can improve the flexibility of the model and prevent the output from being too dependent on the input. The effect of this formula is that it no longer treats each feature independently, but makes the weight allocation of each feature sub-vector depend on the overall context composed of all features. This enables true dynamic, context-aware feature fusion. For example, when the system identifies materials from a particular supplier, it may automatically increase the weight of relevant process parameter features. Finally, using the calculated normalized attention weight values ​​for each feature category, the production parameter feature vector, material spectrum feature vector, equipment status feature vector, historical defect record feature vector, detection equipment status feature vector, and order information feature vector are weighted respectively. All weighted feature vectors are then summed to generate the final high-dimensional dense dynamic risk feature vector, which serves as a deep abstract representation of the comprehensive risk of each workpiece.

[0021] In this embodiment, the specific process of arranging the risk characteristics into a sequence according to their production time in step S2 includes: The dynamic risk feature vectors of multiple rotor shells of forgings to be inspected obtained from step S1 are strictly sorted according to the timestamp of each workpiece being processed on the production line to form a continuous risk feature sequence in the time dimension. Each position in the risk feature sequence corresponds to a workpiece produced at a certain time point. The length of the risk feature sequence is determined by the number of workpieces for which defect probability prediction is currently required. When constructing the risk feature sequence, a time window is maintained to ensure that the workpieces in the risk feature sequence have continuity in production time, and to avoid the break in the time sequence relationship due to excessive time intervals. For multiple workpieces with the same production time, a secondary sorting will be performed according to the order in which they enter the inspection process to ensure the uniqueness of each position in the risk feature sequence. After sorting, each dynamic risk feature vector in the risk feature sequence is stored and organized in chronological order to form a structured sequence data object, which will serve as the direct input to the hybrid neural network model. The specific process of mining risk feature sequences using a hybrid neural network model of GRU and Transformer encoder includes: First, the risk feature sequence is input as a whole into the GRU layer for processing. The GRU (Gated Recurrent Unit) layer processes the dynamic risk feature vector of each workpiece in the sequence step by step through the synergistic effect of its internal update gate and reset gate, captures the short-term local dependencies between adjacent workpieces in the sequence, and outputs a hidden state sequence containing local temporal information. Each hidden state vector in the hidden state sequence encodes the local sequence information up to the corresponding time step. Next, the hidden state sequence is fed as a whole into the Transformer encoder layer. The Transformer encoder layer uses a self-attention mechanism to calculate the association weights between the hidden state vectors of any two time steps in the sequence, thereby capturing long-term global dependency patterns across multiple artifacts and outputting a feature representation sequence enhanced by global context information. This enhanced feature representation sequence consists of the enhanced feature vectors of each artifact, where each enhanced feature vector incorporates global context information. Finally, the enhanced feature vector corresponding to each workpiece is input into a fully connected layer. This fully connected layer performs a linear transformation operation, including performing a dot product operation between the enhanced feature vector and the trainable weight vector and adding a trainable bias value. Then, a nonlinear transformation is performed through a sigmoid activation function. The result of the dot product operation and the addition of the bias value is mapped to a defect occurrence probability value between zero and one, thus obtaining the defect occurrence probability function of the rotor shell of each forging to be inspected. Among them, the trainable weight vector and the trainable bias value are dynamically learned through the model training process to adapt to the defect risk mode under different production contexts. By performing a linear transformation operation and using a sigmoid activation function to convert the linear combination of the enhanced feature vectors into a defect probability, the nonlinear relationship between features and probabilities is reflected, ensuring the rationality and interpretability of the probability value.

[0022] In this embodiment, it is particularly important to explain the specific process in step S3 of modeling the detection scheduling problem as a Markov decision process, which includes: First, based on the defect occurrence probability function of the rotor shell of each forging to be inspected obtained in step S2 and the real-time inspection equipment status and order information obtained from the multi-source data in step S1, the state space of the Markov decision process is constructed. The state space is composed of the defect occurrence probability function of all work-in-process, the status of inspection equipment, order information, and additional environmental status. The defect occurrence probability function is directly derived from the output of step S2. The inspection equipment status includes the number of available ultrasonic testing equipment, the health status indicators of ultrasonic testing equipment, the equipment calibration cycle, and the estimated idle time. The order information includes the urgency of order delivery and the customer level identifier. The additional environmental status includes the workpiece's position on the production line and the timestamp. Furthermore, the state vector of the state space at each time step is formed by sequentially connecting the aforementioned defect occurrence probability function, inspection equipment status, order information, and additional environmental status to form a unified dynamic environment representation. Secondly, define the action space, which is a discrete set. Each action represents a specific detection scheduling decision, such as assigning a specified workpiece to a specified ultrasonic flaw detection device. The size of the action space depends on the number of workpieces to be inspected and the available detection equipment resources. Then, a hierarchical reward function is designed, which is a linear combination of three factors: efficiency reward, risk reward, and cost penalty, through weighted coefficients. The efficiency bonus is calculated based on the progress of the detection task, which encourages reducing equipment idle time and total detection time. The risk reward is calculated by multiplying the probability function of defect occurrence of each workpiece to be inspected by the negative exponential function of its cumulative delay time, and summing such products for all workpieces. The cumulative delay time is calculated from the time the workpiece is inspectable, the base of the exponential function is the natural constant, and the exponent is the product of the negative loss coefficient and the cumulative delay time. This dynamically adjusts the risk weight to encourage priority inspection of high-risk workpieces with short delay times. Cost penalties are based on equipment switching and energy costs to optimize resource utilization; The hierarchical reward function balances the importance of each sub-item by adjusting the efficiency reward weight coefficient, risk reward weight coefficient, and cost penalty weight coefficient, ensuring that scheduling decisions take into account efficiency, risk, and economy. Finally, the state transition model is dynamically determined by the production environment, including workpiece movement and equipment state changes. Due to the complexity of the environment, the state transition probability is approximated by collecting historical state transition data and using machine learning methods. The specific process of applying the multi-agent proximal policy optimization algorithm includes: Each detection device is considered an agent, and multiple agents collaborate within a Markov decision process framework to optimize the global reward. Each agent maintains a policy network and a value network. The policy network takes the current state as input and outputs the probability distribution of each possible action in that state. The internal parameters of the policy network are learned through training. The value network takes the current state as input and outputs an estimate of the long-term expected reward for that state. The internal parameters of the value network are also learned through training. Agents collaborate through centralized training and distributed execution. During training, each agent selects an action based on the current state, and the environment returns a reward value based on a hierarchical reward function and transitions to a new state. The training objective is to update the parameters of the policy network and the value network by maximizing the expected cumulative discounted reward. The expected cumulative discounted reward is obtained by summing the reward values ​​of all future time steps multiplied by powers of the discount factor. The discount factor is used to adjust the current value of future rewards, and its value is between zero and one. The multi-agent proximal policy optimization algorithm uses proximal pruning to ensure training stability and avoid excessive policy updates; During the real-time scheduling phase, for the real-time state at each decision moment, each agent outputs an action probability distribution based on the trained policy network, and selects a specific detection and scheduling action from the probability distribution through a sampling method. The sampling actions of all intelligent agents are combined to form a dynamic detection scheduling instruction set. The dynamic detection scheduling instruction set is parsed into specific control instructions, such as assigning a specified ultrasonic flaw detection equipment to a specified workpiece for detection; thereby driving the ultrasonic flaw detection equipment to perform actual detection operations, thereby obtaining measured data containing defect judgment results. The measured data includes defect type, defect size, and defect location information.

[0023] In this embodiment, the specific process of collecting data based on a preset rolling time window in step S4 is explained as follows: Throughout the execution of the dynamic detection scheduling instruction generated in step S3, the duration of the rolling time window is preset according to the production rhythm, for example, set to 24 hours. The rolling time window covers all production activities within the preset time period calculated backward from the current moment. Within the rolling time window, two types of data are collected simultaneously. The first type is the measured data generated by the detection operation. This measured data comes directly from the defect judgment results obtained by the detection system after executing the dynamic detection scheduling instruction in step S3 and performing actual detection on the workpiece. This includes defect type, defect size, and defect location information. The defect type covers inclusions, porosity, and cracks. The defect size is quantified in millimeters or micrometers. The defect location is represented by the workpiece's local coordinate system or global coordinate system. The second type is the current multi-source data collected in real time from the production line. This current multi-source data includes the production parameters, material spectrum, and equipment status corresponding to the workpiece newly entering the production line within the rolling time window. The production parameters include forging temperature, pressure value, and cooling rate. The material spectrum includes steel furnace number, chemical composition ratio, supplier batch code, and raw material supplier identification. The equipment status includes machine tool vibration amplitude, current die wear, equipment operating frequency, and temperature reading. The collected measured data and the production parameters, material spectrum, and equipment status from the current multi-source data are preprocessed, including data cleaning to remove outliers, data standardization to unify the units of measurement, and data alignment to ensure that the various types of data for the same workpiece are consistent in terms of timestamps. The preprocessed data is then organized into a structured training sample set, with each sample containing input features and output labels. The input features are derived from the production parameters, material spectrum, and equipment status from the current multi-source data, and the output labels are derived from the defect judgment results in the measured data, thus providing high-quality input for the subsequent online learning mechanism. The specific process of incremental fine-tuning using online learning mechanisms includes: Based on the measured data collected and preprocessed by the rolling time window, and the production parameters, material spectrum and equipment status in the current multi-source data, the decision-making strategies of the hybrid neural network model in step S2 and the multi-agent proximal strategy optimization algorithm in step S3 are collaboratively and incrementally fine-tuned using an online learning mechanism. For hybrid neural network models, the fine-tuning process aims to minimize the difference between the model's predicted defect probability and the actual defect label in the measured data. This is achieved by calculating the cross-entropy loss function, which is specifically calculated as follows: based on the average loss of all training samples in the training sample set. The loss of each training sample is the sum of the product of the sample's actual defect label and the natural logarithm of the model's predicted defect probability. The actual defect label takes the value of zero or one, representing no defect and defective, respectively. The model's predicted defect probability is a real number between zero and one. The number of training samples is determined by the amount of effective data within the rolling time window. For the decision-making strategy of the multi-agent proximal policy optimization algorithm, the fine-tuning process aims to maximize the long-term cumulative reward by calculating the policy gradient loss function. The calculation of the policy gradient loss function is based on the average loss of all training samples in the training sample set. The loss of each training sample is composed of the product of the advantage function value and the natural logarithm of the action probability output by the policy network. The advantage function value represents the relative value of performing a given action in a given state, and the action probability is the selection probability output by the policy network based on the given state. The number of training samples is consistent with the number of training samples used in the hybrid neural network model. The total loss function is obtained by multiplying the first weight coefficient by the loss function of the hybrid neural network model and the second weight coefficient by the loss function of the decision policy. Both the first and second weight coefficients are adjusted through the validation set to balance prediction accuracy and policy performance. Based on the total loss function, the gradient of the total loss function with respect to the model parameters and policy parameters is calculated. The parameters are then updated using a gradient descent algorithm with momentum and adaptive learning rate. The parameter update process is as follows: the current parameter value is added to the gradient of the loss function multiplied by the adaptive learning rate, and the momentum coefficient is added to the difference between the current parameter value and the previous parameter value to obtain the updated parameter value. The adaptive learning rate gradually decays with the time step, and the decay method is controlled by the decay coefficient. The momentum coefficient takes a real number between zero and one. The parameter update is incremental, and the gradient is calculated and the parameters are updated only based on the new data within the rolling time window each time, avoiding retraining the entire model.

[0024] Example 2 This embodiment provides, for example Figure 2 The system shown is for detecting internal defects in a forged rotor housing, specifically including: Data acquisition module: After receiving the trigger signal of a new workpiece entering the production line, it collects multi-source data of the forging rotor shell in real time. The multi-source data includes production parameters, material spectrum, equipment status, historical defect records, inspection equipment status and order information. The collected multi-source data is sent to the feature extraction module. Feature extraction module: Based on the received multi-source data, the deep feature extraction network based on the attention mechanism is used to fuse the multi-source data and generate a corresponding dynamic risk feature vector for each forging rotor shell. The dynamic risk feature vector is then passed to the prediction module. Prediction module: Receives dynamic risk feature vectors of multiple forged rotor shells, arranges them in order of their production time to form a risk feature sequence, and uses a hybrid neural network model of GRU and Transformer encoder to mine the risk feature sequence and outputs the defect occurrence probability function of each forged rotor shell. Scheduling Decision Module: Based on the defect occurrence probability function, the status of the detection equipment, and order information, the detection scheduling problem is modeled as a Markov decision process. A multi-agent proximal policy optimization algorithm is applied to optimize the detection decision with a hierarchical reward function, generating dynamic detection scheduling instructions and sending them to the detection execution module. The detection execution module executes dynamic detection scheduling instructions to control the operation of the detection system. It collects measured data and current production data based on a preset rolling time window, and uses an online learning mechanism to incrementally fine-tune the hybrid neural network model and decision-making strategy to achieve continuous self-evolution of the hybrid neural network model and decision-making strategy.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method of detecting internal defects in a forged rotor shell, characterized by, Specifically comprising the following steps: Step S1, in the production process of the forged rotor shell, based on the triggering event of a new workpiece entering the production line, real-time acquisition of multi-source data of the forged rotor shell to be detected, the multi-source data including production parameters, material pedigree, equipment state, historical defect records, detection equipment state and order information, using a deep feature extraction network based on an attention mechanism to fuse the multi-source data, and generating a corresponding dynamic risk feature vector for each forged rotor shell to be detected; Step S2, based on the dynamic risk feature vectors corresponding to the plurality of forged rotor shells to be detected generated in step S1, arranging the dynamic risk feature vectors in the order of their production time to form a risk feature sequence, using a hybrid neural network model of GRU and Transformer encoder to mine the risk feature sequence, and outputting a defect occurrence probability function of each forged rotor shell to be detected; Step S3, according to the defect occurrence probability function obtained in step S2 and the detection equipment state and order information obtained from the multi-source data in step S1, modeling the detection scheduling problem as a Markov decision process, and applying a multi-agent proximal policy optimization algorithm to optimize the detection decision with a hierarchical reward function, generating a dynamic detection scheduling instruction set to execute the detection job, and thus obtaining measured data containing defect judgment results; Step S4, in the process of executing the dynamic detection scheduling instruction set generated in step S3 by the detection system, based on a preset rolling time window, acquiring the measured data generated by the detection job and the production parameters, material pedigree and equipment state in the current multi-source data acquired from the production line in real time; using an online learning mechanism, based on the measured data and the production parameters, material pedigree and equipment state in the current multi-source data, incrementally fine-tuning the decision strategy of the hybrid neural network model in step S2 and the multi-agent proximal policy optimization algorithm in step S3, and realizing continuous self-evolution of the hybrid neural network model and the decision strategy.

2. A method of detecting internal defects in a forged rotor shell according to claim 1, characterized in that: In step S1, the specific process of real-time acquisition of multi-source data of the forged rotor shell to be detected includes: When the triggering event of a new workpiece entering the production line is detected, immediately start the multi-source data synchronous acquisition process; The production parameters in the multi-source data refer to the process variables directly related to the processing of the new workpiece obtained from the manufacturing execution system, including forging temperature, pressure value and cooling rate; The material pedigree refers to the attribute information related to the raw materials of the new workpiece extracted from the enterprise resource planning system, including steel furnace number, chemical composition ratio, supplier batch code and raw material supplier identification; The equipment state refers to the running index of the equipment when processing the new workpiece read in real time from the sensor network, including machine tool vibration amplitude, current die wear, equipment running frequency and temperature reading; The historical defect record refers to the defect information of the detected workpiece in the recent production sequence queried from the quality database, including defect type, defect severity, defect location and corresponding production condition context; The detection equipment state is a resource state variable available for nondestructive detection obtained from a detection resource management unit, including the number of ultrasonic flaw detection equipment available, the health state indicator of the ultrasonic flaw detection equipment, the equipment calibration period, and the predicted idle time. The order information is the delivery urgency and customer level identifier of the order to which the new workpiece belongs, retrieved from an order management unit. After the collection of all multi-source data is completed, data preprocessing and alignment operations are performed, including cleaning and denoising and standardization processing of numerical data, numerical encoding of enumeration data, and alignment of all data to the production time point of the new workpiece using the unique workpiece identifier and timestamp, forming a unified multi-dimensional data snapshot bound to the new workpiece.

3. A method of detecting internal defects in a forged rotor shell according to claim 2, characterized in that: The specific process of using the deep feature extraction network based on the attention mechanism to fuse the multi-source data includes: First, each type of preprocessed multi-source data is mapped to a unified dimension feature subvector through an independent embedding layer, including production parameter feature subvector, material pedigree feature subvector, equipment state feature subvector, historical defect record feature subvector, detection equipment state feature subvector, and order information feature subvector. Then, all feature subvectors are spliced to form a preliminary overall feature vector. Next, a trainable scoring network is used to calculate the significance weight of each feature subvector. For each feature subvector, the scoring network performs the following operations: Take the overall feature vector and the feature subvector as input, perform linear transformation on the overall feature vector and the feature subvector respectively, add the transformation results, apply a first nonlinear activation function to get an intermediate feature representation, then perform dot product operation with a trainable parameter vector and apply a second nonlinear activation function to get the initial score of the feature subvector, then perform exponential calculation and summation normalization on the initial scores of all feature subvectors to get the normalized attention weight of each feature subvector. Finally, use the calculated normalized attention weight value of each feature category to weight the production parameter feature subvector, material pedigree feature subvector, equipment state feature subvector, historical defect record feature subvector, detection equipment state feature subvector, and order information feature subvector, and add all weighted feature subvectors to generate the final dynamic risk feature vector.

4. A method of detecting internal defects in a forged rotor shell according to claim 3, characterized in that: In step S2, the specific process of arranging the risk feature sequence in the order of production time includes: The dynamic risk feature vectors of the multiple forgings to be detected rotor housings obtained in step S1 are sorted according to the time stamp of each workpiece completing processing on the production line to form a continuous risk feature sequence in the time dimension, and each position in the risk feature sequence corresponds to a workpiece produced at a time point.

5. A method of detecting internal defects in a forged rotor shell according to claim 4, characterized in that: The specific process of using the hybrid neural network model of GRU and Transformer encoder to mine the risk feature sequence includes: Firstly, the risk feature sequence is input as a whole into a GRU layer for processing, the GRU layer processes the dynamic risk feature vector of each workpiece in the sequence step by step through the synergistic effect of its internal update gate and reset gate, and outputs a hidden state sequence containing local time sequence information, wherein each hidden state vector in the hidden state sequence encodes the local sequence information up to the corresponding time step; Then, the hidden state sequence is input as a whole into a Transformer encoder layer, the Transformer encoder layer calculates the correlation weight between the hidden state vectors of any two time steps in the sequence using the self-attention mechanism, and outputs an enhanced feature representation sequence enhanced by global context information, the enhanced feature representation sequence is composed of an enhanced feature vector of each workpiece, wherein each enhanced feature vector fuses global context information; Finally, the enhanced feature vector corresponding to each workpiece is input into a fully connected layer, the fully connected layer performs linear transformation operations, including dot product operation of the enhanced feature vector and a trainable weight vector and adding a trainable bias value, and then non-linear transformation through a S-shaped activation function, mapping the result of the dot product operation and the bias value added to a defect occurrence probability value between zero and one, thereby obtaining a defect occurrence probability function of each workpiece rotor shell to be detected.

6. A method of detecting internal defects in a forged rotor shell according to claim 5, characterized in that: The specific process of modeling the detection scheduling problem as a Markov decision process in step S3 includes: Firstly, based on the defect occurrence probability function of each workpiece rotor shell to be detected obtained in step S2 and the detection equipment state and order information obtained in real time from the multi-source data in step S1, a state space of the Markov decision process is constructed; The state space is composed of the defect occurrence probability function of all work-in-process, the detection equipment state, the order information and the additional environmental state; and the state vector of the state space at each time step is formed by sequentially connecting the above-mentioned defect occurrence probability function, detection equipment state, order information and additional environmental state to form a unified dynamic environment representation; Secondly, the action space is defined, the action space is a discrete set, and each action represents a specific detection scheduling decision; Then, a hierarchical reward function is designed, the reward function is linearly combined by efficiency reward, risk reward and cost penalty through weight coefficients; Finally, the state transition model is determined by the dynamic production environment, including workpiece movement and equipment state change.

7. A method of detecting internal defects in a forged rotor shell according to claim 6, characterized in that: The specific process of applying the multi-agent proximal policy optimization algorithm includes: Each detection equipment is regarded as an agent, each agent maintains a policy network and a value network, and the agents cooperate through centralized training and distributed execution, in the training process, each agent selects an action according to the current state, the environment returns a reward value based on the hierarchical reward function and moves to a new state; the training goal is to update the parameters of the policy network and the value network by maximizing the expected cumulative discounted reward, wherein the expected cumulative discounted reward is obtained by summing the rewards of all future time steps multiplied by the power of the discount factor; In the real-time scheduling stage, for the real-time state of each decision moment, each agent outputs an action probability distribution based on the trained policy network, and selects a specific detection and scheduling action from the probability distribution through a sampling method; The combination of the sampled actions of all agents forms a dynamic detection and scheduling instruction set, which is parsed into specific control instructions to drive the ultrasonic flaw detection equipment to perform actual detection work, thereby obtaining measured data containing defect judgment results, including defect type, defect size and defect position information.

8. A method of detecting internal defects in a forged rotor shell according to claim 7, characterized in that: In step S4, the specific process of collecting data based on the preset rolling time window includes: Within the rolling time window, two types of data are synchronously collected, the first type is measured data generated by detection work; the second type is current multi-source data collected from the production line in real time, which includes production parameters, material pedigree and equipment status corresponding to workpieces newly entering the production line within the rolling time window; The collected measured data and production parameters, material pedigree and equipment status in the current multi-source data are preprocessed, including data cleaning, data standardization and data alignment, and the preprocessed data is organized into a structured training sample set, each sample containing input features and output labels, the input features are derived from production parameters, material pedigree and equipment status in the current multi-source data, and the output labels are derived from defect judgment results in the measured data.

9. A method of detecting internal defects in a forged rotor shell according to claim 8, characterized in that: The specific process of using the online learning mechanism for incremental fine-tuning includes: Based on the measured data and production parameters, material pedigree and equipment status in the current multi-source data collected and preprocessed based on the rolling time window, the online learning mechanism is used to cooperatively fine-tune the hybrid neural network model in step S2 and the decision strategy of the multi-agent proximal policy optimization algorithm in step S3; For the hybrid neural network model, the fine-tuning process aims to minimize the difference between the model's predicted defect probability and the true defect label in the measured data, by calculating the cross-entropy loss function, the specific calculation method of which is: based on the average loss of all training samples in the training sample set, where the loss of each training sample is composed of the sum of the natural logarithm product of the true defect label and the model's predicted defect probability of that sample; For the decision strategy of the multi-agent proximal policy optimization algorithm, the fine-tuning process aims to maximize the long-term cumulative reward, by calculating the policy gradient loss function, which is calculated based on the average loss of all training samples in the training sample set, where the loss of each training sample is composed of the product of the advantage function value and the natural logarithm of the action probability output by the policy network; The total loss function is obtained by weighting the sum of the loss function of the hybrid neural network model multiplied by the first weight coefficient and the loss function of the decision strategy multiplied by the second weight coefficient; Based on the total loss function, the gradient of the total loss function with respect to the model parameters and the policy parameters is calculated, and a gradient descent algorithm with momentum and adaptive learning rate is used for parameter updating, wherein the parameter updating process is: adding the adaptive learning rate multiplied by the gradient of the loss function to the current parameter value, and adding the momentum coefficient multiplied by the difference between the current parameter value and the previous parameter value, to obtain the updated parameter value.

10. A system for detecting internal defects in a forged rotor shell for use in a method for detecting internal defects in a forged rotor shell as claimed in any one of claims 1 to 9, characterized in that: Specifically, the following steps are included: The data acquisition module: after receiving the trigger signal of the new workpiece entering the production line, real-time acquisition of the multi-source data of the forging rotor shell is performed, the multi-source data includes production parameters, material pedigree, equipment state, historical defect record, detection equipment state and order information, and the acquired multi-source data is sent to the feature extraction module; The feature extraction module: based on the received multi-source data, the deep feature extraction network based on attention mechanism is used to fuse and process the multi-source data, and the corresponding dynamic risk feature vector of each forging rotor shell is generated, and the dynamic risk feature vector is transmitted to the prediction module; The prediction module: receives the dynamic risk feature vectors of multiple forging rotor shells, arranges them in the order of production time to form a risk feature sequence, uses a hybrid neural network model of GRU and Transformer encoder to mine the risk feature sequence, and outputs the defect occurrence probability function of each forging rotor shell; The scheduling decision module: according to the defect occurrence probability function, the detection equipment state and the order information, the detection scheduling problem is modeled as a Markov decision process, the multi-agent proximal policy optimization algorithm is applied, the detection decision is optimized with a hierarchical reward function, the dynamic detection scheduling instruction is generated and sent to the detection execution module; The detection execution module: executes the dynamic detection scheduling instruction to control the detection system operation, acquires the measured data and current production data based on the preset rolling time window, and uses the online learning mechanism to incrementally fine-tune the hybrid neural network model and the decision strategy, so as to realize the continuous self-evolution of the hybrid neural network model and the decision strategy.

Images