Large die forging press fault early warning method based on fusion model

Abstract

The invention relates to the crossing field 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 following steps: collecting multi-modal real-time data through a multi-source sensor; generating a standardized time sequence feature sequence through streaming preprocessing; performing rapid screening by a lightweight primary early warning model deployed at an edge end, and if the abnormal confidence exceeds a threshold value, triggering a secondary diagnosis process; original data fragments are transmitted to a special acceleration reasoning unit, a multi-branch heterogeneous fusion diagnosis model is operated, and a refined fault report is output; and the edge end executes an alarm or control instruction accordingly. The device comprises an edge calculation unit, a multi-source sensor array and a special accelerated reasoning unit for physical isolation. According to the invention, through a two-stage collaborative architecture, high-precision fault identification is realized while millisecond-level response is guaranteed, the calculation load and network bandwidth occupation are significantly reduced, and the data security is improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and intelligent manufacturing, specifically involving a fault early warning method for large-scale die forging presses based on a fusion model. Background Technology

[0002] Large forging presses, as core heavy equipment in high-end equipment manufacturing, are widely used in key fields such as aerospace, energy equipment, and rail transportation. Their operational status directly affects production safety and product quality. With the improvement of industrial intelligence, data-driven fault early warning technology is gradually becoming the mainstream approach to ensure equipment reliability. Currently, such early warning systems mostly rely on lightweight models deployed at the edge to meet real-time requirements, or on complex fusion models running on cloud-based high-performance computing platforms to pursue high-precision diagnosis. These two approaches have long been architecturally disconnected.

[0003] Fault early warning methods based on fusion models aim to combine multi-source sensor data (such as vibration, temperature, and hydraulic pressure) with multi-algorithms such as deep learning and physical models to accurately capture early, subtle fault characteristics. These methods typically improve the ability to identify complex and progressive faults by integrating convolutional neural networks, temporal modeling units, and expert rule bases to construct a discrimination boundary in a high-dimensional feature space. However, their high computational complexity and large memory footprint make them difficult to deploy directly on resource-constrained edge controllers.

[0004] In existing technologies, while edge-cloud collaborative architectures have been attempted to balance real-time performance and accuracy, task allocation strategies often employ static thresholds or fixed-ratio scheduling, lacking global optimization considerations for dynamic operating conditions, communication load, and equipment energy efficiency. When a press machine experiences a sudden anomaly requiring millisecond-level response, the edge device is prone to missed detections due to limited model capabilities; while uploading all data to the cloud introduces uncontrollable latency and exacerbates bandwidth pressure and energy consumption. Especially in continuous high-intensity forging operations, the contradiction between the computing power bottleneck of edge devices and the lag in cloud analysis becomes increasingly prominent, making it difficult for the early warning system to balance low-latency response and high-confidence judgment. Therefore, there is an urgent need for a collaborative mechanism that can adaptively coordinate edge and cloud computing resources to achieve efficient and accurate early warning under strict resource constraints. Summary of the Invention

[0005] This invention provides a fault early warning method for large-scale forging presses based on a fusion model, aiming to resolve the fundamental contradiction between the stringent low-latency requirements of real-time early warning and the high computational overhead of complex prediction models. In existing technologies, fault early warning for large-scale forging presses largely relies on single machine learning models deployed on edge computing nodes, such as support vector machines, shallow neural networks, or decision trees. While these models offer low inference latency, their feature representation capabilities are limited, making it difficult to effectively capture deep fault precursor patterns in the high-dimensional, nonlinear, and strongly coupled time-series data generated by multi-source heterogeneous sensors during equipment operation. To improve early warning accuracy, some solutions attempt to introduce deep learning models, such as long short-term memory networks or convolutional neural networks. However, these models have a large number of parameters and high computational complexity, failing to meet the millisecond-level real-time requirements on resource-constrained industrial edge devices, resulting in delayed early warnings and loss of the intervention window.

[0006] Another approach employs a cloud-based collaborative architecture, uploading raw data to the cloud for complex model inference and then distributing the results to the edge. However, this approach not only introduces uncontrollable network transmission latency but also places enormous pressure on the communication bandwidth of industrial sites due to the continuous uploading of massive amounts of raw sensor data, while also posing risks to data security and privacy leaks. Therefore, there is an urgent need for a novel fault early warning method that can both guarantee high-precision fault identification capabilities and strictly meet the constraints of low latency and low resource consumption at the edge.

[0007] Furthermore, the fault early warning method for a large forging press based on a fusion model includes the following steps: synchronously collecting multimodal real-time status data during equipment operation using a multi-source sensor array deployed on key components of the large forging press; performing a preset streaming data preprocessing operation on the multimodal real-time status data to generate a standardized time-series feature sequence; inputting the standardized time-series feature sequence into a lightweight primary early warning model pre-built and deployed on an edge computing unit, which performs a first round of rapid fault screening and outputs a primary early warning confidence level characterizing the probability of an abnormal current operating state of the equipment; determining whether the primary early warning confidence level exceeds a preset first threshold; if it does not exceed the threshold, determining that the equipment is operating normally and ending the current early warning cycle; if... If the threshold is exceeded, a secondary deep diagnostic process is triggered. This process transmits cached original multimodal real-time status data fragments from the same time period to a dedicated accelerated inference unit that is physically isolated from but logically coordinated with the edge computing unit via a secure, isolated data channel. On this dedicated accelerated inference unit, a more complex secondary fusion diagnostic model is loaded and run. This model performs fine-grained cross-modal feature extraction and spatiotemporal correlation analysis on the input original multimodal real-time status data fragments, outputting a refined diagnostic report containing specific fault types, fault locations, and fault severity. This refined diagnostic report is then sent back to the edge computing unit, which, based on the report's content, executes corresponding local alarms, records data, or sends an emergency shutdown command to the host computer system.

[0008] Furthermore, the multi-source sensor array includes at least a vibration acceleration sensor mounted on the main drive shaft, pressure and flow sensors mounted on the hydraulic system, a displacement sensor mounted on the slider guide rail, a temperature sensor mounted on the motor windings, and an oil quality sensor mounted on the lubrication pipeline. The sampling frequency of the vibration acceleration sensor is not less than 10 kHz, the sampling frequency of the pressure and flow sensor is not less than 1000 Hz, and the sampling frequency of the other sensors is not less than 100 Hz. All sensors establish a hard real-time communication connection with the edge computing unit through an industrial Ethernet bus and are configured with a hardware-level timestamp synchronization mechanism to ensure that the time alignment error of each data source is less than 1 microsecond.

[0009] Furthermore, the streaming data preprocessing operation specifically includes: performing a sliding window truncation on the raw data streams from each sensor, with a window length of 5 seconds and a sliding step size of 500 milliseconds; performing zero-mean normalization on the data within the window, with the mean and standard deviation parameters obtained from historical data collected during the device's factory no-load testing phase; downsampling the normalized data, uniformly resampling all sensor data to a reference frequency of 200 Hz; and finally, concatenating the resampled multi-channel data into a two-dimensional matrix as the standardized time-series feature sequence.

[0010] Furthermore, the lightweight primary early warning model is a deep separable convolutional neural network with no more than eight layers and a total parameter count controlled within 500,000. The model's input is the standardized temporal feature sequence, and its output layer is a fully connected layer with a single neuron, using a sigmoid activation function to directly output the primary early warning confidence level. Before deployment, the model is trained end-to-end using an offline historical dataset covering normal equipment operating conditions and various early minor fault modes. Knowledge distillation technology is used to transfer key discriminative features from a more performant but structurally complex teacher model to maintain a high anomaly detection rate under a minimally simplistic network structure. The first threshold is set to 0.7.

[0011] Furthermore, the dedicated accelerated inference unit is an embedded graphics processor module independent of the main control edge computing unit. It has a dedicated tensor computing core and high-bandwidth memory, and interacts with the edge computing unit at high speed via a PCIe third-generation interface. This accelerated inference unit is physically isolated from the factory control network and only activates when triggered, otherwise remaining in a deep sleep state to reduce power consumption.

[0012] Furthermore, the secondary fusion diagnostic model is a multi-branch heterogeneous fusion network containing three parallel feature extraction branches: the first branch is a one-dimensional convolutional neural network used to extract local temporal patterns within each sensor signal; the second branch is a graph attention network, which models the key components of the press as nodes in a graph, and the physical connections and force transmission relationships between components as edges, using sensor data as node features to learn dynamic fault propagation correlations between components; the third branch is a bidirectional gated recurrent unit network used to capture long-distance global temporal dependencies. The output feature vectors of the three branches are dynamically integrated by an adaptive feature weighting fusion module, which automatically assigns weights to the features of each branch based on the characteristics of the current input data. The fused feature vector is then fed into a fully connected classifier, ultimately outputting the refined diagnostic report. The training dataset of this secondary fusion diagnostic model contains a large number of severe fault cases labeled with precise fault tags, ensuring that it can provide high-confidence accurate diagnosis when triggered.

[0013] Furthermore, after the refined diagnostic report is sent back to the edge computing unit, the edge computing unit will package and encrypt the complete context information of this early warning event, including the initial early warning confidence level, trigger time, hash digest of the original data fragment, and diagnostic report, and then asynchronously upload it to the remote device health management cloud platform for subsequent online model updates and iterative optimization of the fault knowledge base.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: By constructing a two-level collaborative architecture of a lightweight primary early warning model on the edge side and a complex secondary fusion diagnostic model on a dedicated acceleration unit, intelligent load balancing is achieved.

[0015] During most of the equipment's normal operation, only the primary model with extremely low resource consumption performs high-frequency, low-latency screening, effectively ensuring real-time performance. Only when the primary model detects potential anomalies is the secondary model activated as needed for in-depth diagnosis, avoiding the continuous running overhead of complex models.

[0016] Without sacrificing fault diagnosis accuracy, the average computing load of edge devices is reduced by more than 70%, and the end-to-end latency of early warning is stably controlled within 200 milliseconds, fully meeting the stringent requirements of large die forging presses for real-time safety monitoring.

[0017] By confining the processing of highly sensitive raw data to a dedicated local acceleration unit, and uploading only diagnostic results and data summaries, the bandwidth consumption of the factory network is significantly reduced, and data security is improved. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the two-level collaborative early warning and diagnosis mechanism in this invention; Figure 3 This is a logical flow diagram of the multi-source sensor data acquisition and streaming preprocessing stage in this invention; Figure 4 This is a diagram illustrating the triggering and collaborative decision-making logic framework of the lightweight primary early warning model and the secondary fusion diagnostic model in this invention. Figure 5 This is a schematic diagram of the multi-branch heterogeneous feature extraction and adaptive fusion principle framework of the secondary fusion diagnostic model in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the edge computing unit, the dedicated accelerated inference unit, and the remote cloud platform in this invention. Detailed Implementation

[0019] Please refer to the attached document. Figure 1 To be continued Figure 6This invention provides a fault early warning method for large forging presses based on a fusion model. Its core lies in constructing a two-level collaborative architecture between an edge computing unit and a dedicated accelerated inference unit to achieve high-precision, low-latency fault early warning in resource-constrained industrial environments. The method collects equipment operating status data in real time through a multi-source sensor array. After streaming preprocessing, a lightweight primary early warning model deployed on the edge computing unit performs rapid anomaly screening. Only when the screening results exceed a preset threshold is a secondary deep diagnostic process triggered, transmitting the original data fragments to the dedicated accelerated inference unit, where a complex secondary fusion diagnostic model performs refined fault identification. The entire process strictly follows the S-step sequence to ensure the feasibility and logical consistency of the technical solution.

[0020] The method includes the following steps: S1, through a multi-source sensor array deployed on key components of a large die forging press, synchronously collects multi-modal real-time status data during equipment operation; S2, perform a preset streaming data preprocessing operation on the multimodal real-time state data to generate a standardized time-series feature sequence; S3, the standardized time-series feature sequence is input into a lightweight primary early warning model that is pre-built and deployed on the edge computing unit. The model performs the first round of rapid fault screening and outputs a primary early warning confidence level that characterizes the probability of abnormal operation of the current device. S4, determine whether the confidence level of the primary warning exceeds a preset first threshold; If S5 is not exceeded, the equipment is deemed to be operating normally, and the current warning cycle ends. S6, if exceeded, triggers a secondary deep diagnostic process, which transmits the cached original multimodal real-time status data fragments within the same time period to a dedicated accelerated inference unit that is physically isolated from but logically coordinated with the edge computing unit through a securely isolated data channel. S7. On the dedicated accelerated inference unit, a more complex secondary fusion diagnostic model is loaded and run. This model performs fine-grained cross-modal feature extraction and spatiotemporal correlation analysis on the input original multimodal real-time state data fragments, and outputs a refined diagnostic report containing specific fault types, fault locations and fault severity. S8, the refined diagnostic report is sent back to the edge computing unit, and the edge computing unit executes the corresponding local alarm, data recording or sends an emergency shutdown command to the host computer system according to the content of the report.

[0021] In step S1, the multi-source sensor array is precisely installed at key mechanical and thermodynamic nodes of the large forging press to comprehensively cover the core operating status of the equipment. Specifically, the vibration acceleration sensor is rigidly fixed to the outside of the bearing housing of the main drive shaft to capture high-frequency mechanical vibration signals caused by gear meshing errors, shaft imbalance, or bearing wear; the pressure and flow sensors are integrated into the oil inlet and return lines of the main hydraulic cylinder to monitor the system working pressure and the instantaneous flow rate of hydraulic oil, respectively, to reflect the servo valve response characteristics, seal leakage, or pump efficiency decay; the displacement sensor adopts the non-contact eddy current principle and is installed on both sides of the slider guide rail to measure the vertical stroke deviation and sway angle of the slider in real time, to detect guide rail wear, increased guide clearance, or off-center load conditions; the temperature sensor is a platinum resistance temperature measuring element embedded in the stator winding slot of the motor to directly sense the winding temperature rise and avoid insulation aging caused by poor heat dissipation or overload; the oil quality sensor is connected in series in the main lubrication circuit and continuously monitors the contamination degree, moisture content, and oxidation degree of the lubricating oil through the composite sensing principle of dielectric constant and viscosity.

[0022] All sensors are equipped with independent signal conditioning circuits and establish a hard real-time communication connection with the edge computing unit via an industrial Ethernet bus. To ensure the time consistency of multi-source data, the system incorporates a hardware-level timestamp synchronization mechanism. This mechanism, based on the IEEE 1588 precision time protocol, uses a master clock source to broadcast synchronization pulses to each sensor node, strictly controlling the time alignment error of each data source to within 1 microsecond. The sampling frequency of the vibration acceleration sensor is set to 10 kHz to fully capture the characteristic frequencies of bearing faults and their harmonic components; the sampling frequency of the pressure and flow sensors is set to 1000 Hz, sufficient to analyze the transient processes of hydraulic shock and pressure fluctuations; the sampling frequency of the remaining sensors is uniformly set to 100 Hz to meet the monitoring requirements of slowly changing thermal and displacement parameters.

[0023] In step S2, the streaming data preprocessing operation is executed in a pipelined manner within the real-time operating system kernel of the edge computing unit, ensuring a processing latency of less than 10 milliseconds. This operation first performs a sliding window truncation on the raw data streams from each sensor. The window length is fixed at 5 seconds, and the sliding step size is 500 milliseconds, forming continuously overlapping data blocks to balance the sufficiency of feature extraction with the timeliness of temporal response. Subsequently, zero-mean normalization is performed on the data within each window. The mean and standard deviation parameters are not dynamically calculated but are statistically derived from historical data collected during the device's factory no-load testing phase and stored in the edge computing unit's read-only memory to eliminate the computational overhead and drift risk introduced by online statistics. The normalization formula is expressed as:

[0024] in, These are the original sampled values. and These represent the mean and standard deviation of the corresponding sensor channels during the factory no-load testing phase. After normalization, the system downsamples the data from all channels, using a linear interpolation algorithm to resample the vibration signal from 10 kHz, the hydraulic signal from 1000 Hz, and the other signals from 100 Hz to a reference frequency of 200 Hz. This reference frequency is chosen based on the Nyquist sampling theorem and the input dimension constraints of the primary early warning model, significantly reducing the computational load of subsequent models while preserving key fault frequency band information. Finally, the resampled multi-channel data is aligned by time steps and concatenated into a two-dimensional matrix. The row dimension represents the number of time steps (1000 time steps for a 5-second window), and the column dimension represents the total number of sensor channels (assuming 12 effective channels). This matrix is ​​the standardized time-series feature sequence, serving as the sole input to the primary early warning model.

[0025] In step S3, the lightweight primary early warning model is designed as a deep separable convolutional neural network. Its overall architecture includes 7 convolutional layers and a fully connected output layer, with the total number of parameters strictly controlled to within 500,000. The network first extracts the local temporal patterns of each channel along the time dimension using one-dimensional deep convolutional layers. Each deep convolutional kernel has a size of 7 and a stride of 2. Subsequently, it achieves information interaction between channels through pointwise convolutional layers, with a kernel size of 1, and the number of output channels is gradually compressed. Batch normalization and ReLU activation functions are introduced in the intermediate layers to enhance nonlinear expressive power and accelerate convergence.

[0026] The network ends with a global average pooling layer, compressing the temporal dimension into a single point. A fully connected layer with a single neuron, combined with a sigmoid activation function, directly outputs a preliminary warning confidence level between 0 and 1. Before deployment, the model was trained end-to-end using an offline historical dataset covering more than 10 early fault modes, including normal equipment operation, early minor bearing wear, hydraulic system micro-leakage, and minor motor overheating. To maintain a high anomaly detection rate with a minimal network structure, knowledge distillation was employed during training to transfer key discriminative features from a superior but parameter-intensive teacher model (such as ResNet-1D). The soft-label output of the teacher model, together with the true label, constitutes the loss function, guiding the student model to learn a smoother probability distribution boundary. After training, the model was quantized into an eight-bit integer format and compiled into an inference engine optimized for the ARM architecture of edge computing units, ensuring that a single inference time does not exceed 15 milliseconds.

[0027] In step S4, the first threshold is set to 0.7. This value is determined by plotting the receiver operation characteristic curve on the validation set and selecting the point with the maximum Youden index, aiming to balance the false alarm rate and the false negative rate. The threshold judgment operation is performed in the application layer logic of the edge computing unit. If the primary warning confidence is less than or equal to 0.7, the process proceeds to step S5, determining that the device is operating normally, the current warning cycle ends, and the system immediately begins processing the data of the next sliding window. If the primary warning confidence is greater than 0.7, the process proceeds to step S6, triggering the secondary deep diagnostic process.

[0028] In step S6, the edge computing unit immediately activates the pre-allocated circular buffer, which continuously caches all raw multimodal real-time state data from the past 10 seconds, maintaining the original frequency of each sensor at the sampling rate. Once the trigger condition is met, the system extracts a 5-second raw data segment from the buffer, perfectly aligned with the window used for the primary warning. This segment contains high-fidelity raw signals that have not been downsampled or normalized. Subsequently, this data segment is encapsulated into an encrypted data packet and transmitted to the dedicated accelerated inference unit via a dedicated PCIe Gen 3 data channel physically isolated from the factory control network. This data channel has hardware-level error checking and retransmission mechanisms to ensure data integrity. The dedicated accelerated inference unit is physically a separate embedded graphics processor module, integrating a dedicated tensor computing core and 4GB of high-bandwidth memory. It is normally in a deep sleep state, consuming less than 1 watt; only upon receiving a valid data packet is the edge computing unit send a wake-up signal to initiate the model loading and inference process.

[0029] In step S7, the secondary fusion diagnostic model is loaded into the GPU memory of the dedicated accelerated inference unit and begins inference. This model is a multi-branch heterogeneous fusion network containing three parallel feature extraction branches. The first branch is a one-dimensional convolutional neural network, whose input is the original multi-channel temporal signal. It extracts local temporal patterns under different receptive fields through multi-layer dilated convolutional kernels (dilation rates of 1, 2, and 4), paying particular attention to impact pulses and periodic modulation features. The second branch is a graph attention network. This network first abstracts the key components of the large forging press into a graph structure: the main motor, gearbox, crankshaft, connecting rod, slider, worktable, main hydraulic cylinder, and servo valve—a total of eight components—are treated as nodes in the graph. The physical connections between components (e.g., shaft-bearing), force transmission paths (e.g., connecting rod-slider), and hydraulic coupling relationships (e.g., servo valve-main cylinder) are treated as edges in the graph, constructing a static adjacency matrix. Sensor data is mapped to the initial feature vectors of the corresponding nodes. The graph attention layer dynamically aggregates neighbor information by calculating the importance weights between nodes, learning the propagation path and impact range of faults in the component network. The third branch is a bidirectional gated cyclic unit network, whose forward and backward GRU units capture long-distance temporal dependencies from the past to the future and from the future to the past, respectively, effectively modeling the evolution trend of device state.

[0030] The output feature vectors from the three branches are concatenated and then input into an adaptive feature weighting fusion module. This module consists of two fully connected layers and a Softmax function. It dynamically generates the weight coefficients of the three branches based on the global statistical characteristics of the current input data (such as energy entropy and kurtosis coefficient), achieving on-demand feature fusion. The fused high-dimensional feature vector is finally fed into a three-layer fully connected classifier. This classifier outputs a probability distribution vector whose dimension corresponds to a predefined set of fault categories (such as "main bearing outer ring wear", "servo valve jamming", "slider guide rail wear", etc.), and attaches a regression head to predict the fault severity level (levels one to five). The entire inference process takes no more than 150 milliseconds on a dedicated accelerated inference unit, and the output result is the refined diagnostic report.

[0031] In step S8, the dedicated accelerated inference unit sends the generated refined diagnostic report back to the edge computing unit via the same PCIe channel. Upon receiving the report, the edge computing unit immediately parses its contents: if the report indicates a serious fault (e.g., severity level greater than or equal to level four), it immediately activates the local audible and visual alarm and sends an emergency stop command to the host computer control system via a hard-wired interface; if it indicates a moderate fault, it only records the event log and generates a maintenance work order; if it indicates a minor anomaly, it marks it as pending observation.

[0032] Simultaneously, the edge computing unit packages the complete context information of this early warning event, including the initial warning confidence level, trigger timestamp, SHA-256 hash digest of the original data fragment, and a complete diagnostic report. After encryption using the AES-256 algorithm, it is asynchronously uploaded to the remote equipment health management cloud platform through the factory firewall's unidirectional data egress. This upload operation is executed in a low-priority background thread, without affecting local real-time control tasks. The cloud platform uses the accumulated diagnostic reports and hash digests to periodically update the training datasets of the primary and secondary models, and pushes the optimized model version to the edge via a secure OTA channel, enabling continuous iteration of fault warning capabilities.

[0033] The above method, through the strict sequential execution of S1 to S8, constructs an efficient, secure, and scalable fault early warning system. During the vast majority of normal equipment operation periods, the system consumes only a small amount of computing power from the edge computing unit to perform primary screening, with an average computing load of less than 30%. Only when a potential anomaly is detected is the high-power dedicated acceleration unit activated as needed, thereby improving the overall energy efficiency ratio by more than 3 times. The end-to-end early warning latency (from data acquisition to alarm output) is stably controlled within 200 milliseconds in the worst case (i.e., triggering secondary diagnosis), fully meeting the stringent requirements of real-time safety monitoring for large die forging presses. Simultaneously, the raw, highly sensitive data is always processed within a local closed domain, with only irreversible hash digests and structured diagnostic conclusions uploaded, fundamentally avoiding the risk of data leakage and significantly reducing the bandwidth consumption of the factory network.

Claims

1. A fault early warning method for large-scale forging presses based on a fusion model, characterized in that, include: By deploying a multi-source sensor array on key components of a large die forging press, multi-modal real-time status data during equipment operation is collected synchronously. Perform a preset streaming data preprocessing operation on the multimodal real-time state data to generate a standardized time-series feature sequence; The standardized time-series feature sequence is input into a lightweight primary early warning model that is pre-built and deployed on an edge computing unit. The model performs the first round of rapid fault screening and outputs a primary early warning confidence level that characterizes the probability of abnormal operation of the current device. Determine whether the confidence level of the primary warning exceeds a preset first threshold; If the time limit is not exceeded, the equipment is considered to be operating normally, and the current warning cycle ends. If the limit is exceeded, a secondary deep diagnostic process is triggered, which transmits the cached original multimodal real-time status data fragments within the same time period to a dedicated accelerated inference unit that is physically isolated from but logically coordinated with the edge computing unit through a securely isolated data channel. On the dedicated accelerated inference unit, a more complex secondary fusion diagnostic model is loaded and run. This model performs fine-grained cross-modal feature extraction and spatiotemporal correlation analysis on the input raw multimodal real-time state data fragments, and outputs a refined diagnostic report containing specific fault types, fault locations and fault severity. The refined diagnostic report is sent back to the edge computing unit, which then executes corresponding local alarms, data logging, or sends an emergency shutdown command to the host computer system based on the report's content.

2. The fault early warning method for large forging presses based on a fusion model according to claim 1, characterized in that, The multi-source sensor array includes at least a vibration acceleration sensor mounted on the main drive shaft, a pressure and flow sensor mounted on the hydraulic system, a displacement sensor mounted on the slider guide rail, a temperature sensor mounted on the motor windings, and an oil quality sensor mounted on the lubrication pipeline; the sampling frequency of the vibration acceleration sensor is not less than 10 kHz, the sampling frequency of the pressure and flow sensor is not less than 1000 Hz, and the sampling frequency of the other sensors is not less than 100 Hz; all sensors establish a hard real-time communication connection with the edge computing unit through an industrial Ethernet bus and are configured with a hardware-level timestamp synchronization mechanism to ensure that the time alignment error of each data source is less than 1 microsecond.

3. The fault early warning method for large forging presses based on a fusion model according to claim 2, characterized in that, The streaming data preprocessing operation specifically includes: performing a sliding window truncation on the raw data streams from each sensor, with a window length of 5 seconds and a sliding step size of 500 milliseconds; performing zero-mean normalization on the data within the window, with the mean and standard deviation parameters obtained from historical data collected during the factory no-load testing phase; downsampling the normalized data, uniformly resampling all sensor data to a reference frequency of 200 Hz; and finally, concatenating the resampled multi-channel data into a two-dimensional matrix as the standardized time-series feature sequence.

4. The fault early warning method for large forging presses based on a fusion model according to claim 3, characterized in that, The lightweight primary early warning model is a deep separable convolutional neural network with no more than eight layers and a total number of parameters controlled within 500,000. The input of the model is the standardized temporal feature sequence, and its output layer is a fully connected layer with a single neuron. With the Sigmoid activation function, it directly outputs the primary early warning confidence score. Before deployment, the model is trained end-to-end using an offline historical dataset covering normal equipment operating conditions and various early minor fault modes. Through knowledge distillation technology, key discriminative features are transferred from a higher-performance but more complex teacher model to maintain a high anomaly detection rate under a minimal network structure. The first threshold is set to 0.

7.

5. The fault early warning method for large forging presses based on a fusion model according to claim 4, characterized in that, The dedicated accelerated inference unit is an embedded graphics processor module independent of the main control edge computing unit. It has a dedicated tensor computing core and high-bandwidth memory, and interacts with the edge computing unit at high speed through the PCIe third-generation interface. The accelerated inference unit is physically isolated from the factory control network and is only activated when triggered. It is in a deep sleep state at other times to reduce power consumption.

6. The fault early warning method for large forging presses based on a fusion model according to claim 5, characterized in that, The secondary fusion diagnostic model is a multi-branch heterogeneous fusion network, which includes three parallel feature extraction branches: the first branch is a one-dimensional convolutional neural network, used to extract local temporal patterns within the signals of each sensor; the second branch is a graph attention network, which models the key components of the press as nodes of a graph, and the physical connections and force transmission relationships between components as edges of the graph, using sensor data as node features to learn the dynamic fault propagation correlations between components; the third branch is a bidirectional gated recurrent unit network, used to capture long-distance global temporal dependencies; the output feature vectors of the three branches are dynamically integrated by an adaptive feature weighting fusion module, which automatically assigns weights to the features of each branch according to the characteristics of the current input data; The fused feature vectors are fed into a fully connected classifier, which ultimately outputs the refined diagnostic report.

7. The fault early warning method for large forging presses based on a fusion model according to claim 6, characterized in that, In the graph attention network, key components include the main motor, gearbox, crankshaft, connecting rod, slider, worktable, main hydraulic cylinder, and servo valve, totaling 8 nodes. The physical connections, force transmission paths, and hydraulic coupling relationships between components constitute a static adjacency matrix. Sensor data is mapped to the initial feature vectors of the corresponding nodes. The graph attention layer dynamically aggregates neighbor information by calculating the importance weights between nodes, and learns the propagation path and impact range of faults in the component network.

8. The fault early warning method for large forging presses based on a fusion model according to claim 7, characterized in that, The adaptive feature weighted fusion module consists of two fully connected layers and a Softmax function, which dynamically generates the weight coefficients of three branches based on the global statistical characteristics of the current input data; the global statistical characteristics include energy entropy and kurtosis coefficient.

9. The fault early warning method for large forging presses based on a fusion model according to claim 8, characterized in that, After the refined diagnostic report is sent back to the edge computing unit, the edge computing unit will package and encrypt the complete context information of this early warning event, including the initial early warning confidence level, trigger time, hash digest of the original data fragment, and diagnostic report, and then asynchronously upload it to the remote device health management cloud platform for subsequent online model updates and iterative optimization of the fault knowledge base.

10. The fault early warning method for large forging presses based on a fusion model according to claim 9, characterized in that, The hash digest is generated using the SHA-256 algorithm, the packaging encryption is performed using the AES-256 algorithm, and the upload operation is performed through the one-way data exit of the factory firewall, without affecting the local real-time control task.

Images