Search

CN-122020452-A - Large die forging press fault early warning method based on fusion model

CN122020452ACN 122020452 ACN122020452 ACN 122020452ACN-122020452-A

Abstract

The invention relates to the field of intersection of artificial intelligence and intelligent manufacturing, and discloses a large die forging press fault early warning method based on a fusion model. The method comprises the steps of collecting multi-mode real-time data through a multi-source sensor, generating a standardized time sequence characteristic sequence through stream preprocessing, carrying out quick screening through a lightweight primary early warning model deployed at an edge end, triggering a secondary diagnosis flow if abnormal confidence exceeds a threshold value, transmitting an original data fragment to a special acceleration reasoning unit, operating a multi-branch heterogeneous fusion diagnosis model, outputting a refined fault report, and executing an alarm or control instruction by the edge end according to the information. The invention comprises an edge computing unit, a multi-source sensor array and a physically isolated special acceleration reasoning unit. According to the invention, through a two-stage cooperative architecture, high-precision fault identification is realized while millisecond response is ensured, the calculation load and network bandwidth occupation are obviously reduced, and the data security is improved.

Inventors

  • WANG LAIFU
  • WANG TINGXIAN

Assignees

  • 来富汽车配件(嘉善)有限公司

Dates

Publication Date
20260512
Application Date
20251231

Claims (10)

  1. 1. The utility model provides a large-scale die forging press fault early warning method based on a fusion model, which is characterized by comprising the following steps: synchronously acquiring multi-mode real-time state data in the running process of equipment through a multi-source sensor array arranged on a key part of a large-scale forging press; executing preset stream data preprocessing operation on the multi-mode real-time state data to generate a standardized time sequence characteristic sequence; Inputting the standardized time sequence characteristic sequence into a lightweight primary early warning model which is pre-constructed and deployed in an edge computing unit, executing first-round rapid fault screening by the model, and outputting primary early warning confidence representing the abnormal probability of the current running state of the equipment; Judging whether the primary early warning confidence coefficient exceeds a preset first threshold value; if the current early warning period is not exceeded, judging that the equipment runs normally, and ending the current early warning period; If the data channel is in excess, triggering a secondary depth diagnosis process, and transmitting the cached original multi-mode real-time state data fragments in the same time period to a special acceleration reasoning unit which is physically isolated from the edge computing unit but logically cooperated with the edge computing unit through a safe isolated data channel; Loading and running a secondary fusion diagnosis model with a more complex structure on the special acceleration reasoning unit, wherein the model carries out fine-grained cross-modal feature extraction and space-time correlation analysis on an input original multi-modal real-time state data segment, and outputs a fine diagnosis report containing specific fault types, fault positions and fault severity; and transmitting the refined diagnosis report back to an edge computing unit, and executing corresponding local alarm, data record or sending an emergency stop instruction to an upper computer system by the edge computing unit according to the content of the report.
  2. 2. The method for early warning faults of the large die forging press based on the fusion model according to claim 1 is characterized in that the multi-source sensor array at least comprises a vibration acceleration sensor, a pressure and flow sensor, a displacement sensor, a temperature sensor and an oil quality sensor, wherein the vibration acceleration sensor is arranged on a main transmission shaft, the pressure and flow sensor is arranged on a hydraulic system, the displacement sensor is arranged on a sliding block guide rail, the temperature sensor is arranged on a motor winding, the oil quality sensor is arranged on a lubricating pipeline, the sampling frequency of the vibration acceleration sensor is not lower than 10 kilohertz, the sampling frequency of the pressure and flow sensor is not lower than 1000 hertz, the sampling frequency of the other sensors is not lower than 100 hertz, and all the sensors are connected with an edge computing unit in a hard real-time communication mode through an industrial Ethernet bus and are provided with a hardware level time stamp synchronization mechanism to ensure that time alignment errors of all data are smaller than 1 microsecond.
  3. 3. The fusion model-based large die forging press fault early warning method according to claim 2 is characterized in that the streaming data preprocessing operation specifically comprises the steps of carrying out sliding window interception on an original data stream from each sensor, wherein the window length is 5 seconds, the sliding step length is 500 milliseconds, carrying out zero-mean normalization processing on data in the window, carrying out statistics on historical data collected by equipment in a factory empty test stage on mean and standard deviation parameters, carrying out downsampling on normalized data, uniformly resampling all sensor data to a reference frequency of 200 hertz, and finally splicing the resampled multichannel data into a two-dimensional matrix as the standardized time sequence feature sequence.
  4. 4. The large die forging press fault early warning method based on the fusion model according to claim 3, wherein the light-weight primary early warning model is a deep separable convolutional neural network, the number of network layers is not more than eight, the total parameter quantity is controlled within 50 ten thousand, the input of the model is the standardized time sequence characteristic sequence, the output layer is a full-connection layer of single neurons and is matched with a Sigmoid activation function to directly output the primary early warning confidence, the model is subjected to end-to-end training by using an offline historical data set covering normal working conditions and various early slight fault modes of equipment before deployment, key discrimination characteristics are migrated and learned from a teacher model with better performance and complex structure through a knowledge distillation technology so as to maintain a higher abnormal detection rate under a very simple network structure, and the first threshold is set to be 0.7.
  5. 5. The method for early warning faults of large die forging presses based on fusion models according to claim 4 is characterized in that the special acceleration reasoning unit is an embedded graphic processor module independent of a main control edge computing unit, is provided with a special tensor computing core and a high-bandwidth memory, performs high-speed data interaction with the edge computing unit through a PCIe third generation interface, is physically isolated from a factory control network, is activated only when triggered, and is in a deep sleep state in normal times to reduce power consumption.
  6. 6. The large die forging press fault early warning method based on the fusion model according to claim 5 is characterized in that the secondary fusion diagnosis model is a multi-branch heterogeneous fusion network and comprises 3 parallel feature extraction branches, wherein a first branch is a one-dimensional convolutional neural network and is used for extracting local time sequence modes inside sensor signals, a second branch is a graph injection force network, key components of the press are modeled as nodes of the graph, physical connection and force transmission relation among the components are modeled as edges of the graph, sensor data are used as node features and are used for learning dynamic fault propagation association among the components, a third branch is a two-way gating circulation unit network and is used for capturing long-distance global time sequence dependence, output feature vectors of the 3 branches are dynamically integrated through an adaptive feature weighting fusion module, the module is used for automatically distributing weights of the feature of each branch according to the characteristics of current input data, and the fused feature vectors are sent into a full-connection classifier and finally output the refined diagnosis report.
  7. 7. The large die forging press fault early warning method based on the fusion model according to claim 6 is characterized in that in the graph injection force network, key components comprise a main motor, a reduction gearbox, a crankshaft, a connecting rod, a sliding block, a workbench, a main hydraulic cylinder and a servo valve, 8 nodes are arranged in total, a static adjacent matrix is formed by physical connection, a force transmission path and a hydraulic coupling relation among the components, sensor data are mapped into initial feature vectors of corresponding nodes, and a graph injection force layer dynamically aggregates neighbor information by calculating importance weights among the nodes to learn propagation paths and influence ranges of faults in the component network.
  8. 8. The fusion model-based large die forging press fault early warning method is characterized in that the self-adaptive feature weighted fusion module is composed of 2 full-connection layers and a Softmax function, weight coefficients of 3 branches are dynamically generated according to global statistical characteristics of current input data, and the global statistical characteristics comprise energy entropy and kurtosis coefficients.
  9. 9. The fusion model-based large die forging press fault early warning method according to claim 8, wherein after the refined diagnosis report is returned to an edge calculation unit, the edge calculation unit can upload complete context information of the early warning event, including primary early warning confidence level, trigger time, hash abstract of an original data fragment and diagnosis report, asynchronously upload the complete context information to a remote equipment health management cloud platform after packaging and encryption, and the complete context information is used for subsequent online model updating and iterative optimization of a fault knowledge base.
  10. 10. The fusion model-based large die forging press fault early warning method according to claim 9, wherein the hash digest is generated by adopting an SHA-256 algorithm, the packing encryption is performed by adopting an AES-256 algorithm, and the uploading operation is performed through a unidirectional data outlet of a factory firewall, so that the local real-time control task is not affected.

Description

Large die forging press fault early warning method based on fusion model Technical Field The invention belongs to the field of intersection of artificial intelligence and intelligent manufacturing, and particularly relates to a large die forging press fault early warning method based on a fusion model. Background The large die forging press is used as core heavy equipment in high-end equipment manufacture, is widely applied to the key fields of aerospace, energy equipment, rail transit and the like, and the running state of the large die forging press is directly related to production safety and product quality. With the improvement of the industrial intelligence level, a fault early warning technology based on data driving gradually becomes a main stream direction for guaranteeing the reliability of equipment. Currently, the early warning system depends on a lightweight model deployed at the edge side to meet the real-time requirement, or a complex fusion model is operated by a cloud high-performance computing platform to pursue high-precision diagnosis, and the two models are in a fracture state in a long-term structure. The fault early warning method based on the fusion model aims at combining multisource sensing data (such as vibration, temperature and hydraulic pressure) with a multi-element algorithm such as deep learning and a physical model, and the like, so as to accurately capture early weak fault characteristics. The method generally builds a discrimination boundary under a high-dimensional feature space by integrating a convolutional neural network, a time sequence modeling unit and an expert rule base, so that the recognition capability of composite and progressive faults is improved. However, the method has high computational complexity and large memory occupation, and is difficult to directly deploy on the edge controller with limited resources. In the prior art, although an edge-cloud cooperative architecture is tried to balance instantaneity and accuracy, a task allocation strategy mostly adopts static threshold or fixed proportion scheduling, and global optimization consideration on dynamic working conditions, communication loads and equipment energy efficiency is lacked. When the press is suddenly abnormal and needs millisecond-level response, the edge end is easy to generate missing report due to limited model capacity, and uncontrollable delay is introduced when all data is uploaded to the cloud end, so that bandwidth pressure and energy consumption cost are increased. Especially in the continuous high-strength forging operation scene, the contradiction between the calculation bottleneck of the edge equipment and the cloud analysis lag is more remarkable, so that the early warning system is difficult to consider low-delay response and high-confidence judgment. Therefore, a cooperative mechanism capable of adaptively coordinating the edge and cloud computing resources and realizing efficient and accurate early warning under strict resource constraint is needed. Disclosure of Invention The invention provides a large die forging press fault early warning method based on a fusion model, and aims to solve the fundamental contradiction between the severe requirement of real-time early warning on low delay and the high calculation cost of a complex prediction model. In the prior art, fault early warning of large forging presses is mostly dependent on a single machine learning model deployed on edge computing nodes, such as a support vector machine, a shallow neural network, or a decision tree. Although the model has lower reasoning delay, the characteristic expression capability is limited, and deep fault precursor modes in high-dimensional, nonlinear and strongly-coupled time sequence data generated by the multi-source heterogeneous sensor in the running process of equipment are difficult to effectively capture. In order to improve early warning precision, part of schemes attempt to introduce deep learning models, such as long-term memory networks or convolutional neural networks, but the models have huge parameters and high calculation complexity, and real-time requirements of millisecond-level response cannot be met on industrial edge equipment with limited resources, so that early warning is delayed and an intervention window is lost. And the other scheme adopts a cloud collaborative architecture, the original data is uploaded to the cloud for complex model reasoning, and then the result is issued to the edge. However, the scheme not only introduces uncontrollable network transmission delay, but also continuously uploads massive original sensing data to form huge pressure on the communication bandwidth of the industrial field, and meanwhile, the risks of data security and privacy disclosure exist. Therefore, a novel fault early warning method capable of guaranteeing high-precision fault identification capability and strictly meeting the constraint of low delay and low resource consumption