A forging hydraulic press key element fault diagnosis method and system

Abstract

The application belongs to the technical field of fault diagnosis, and particularly relates to a forging hydraulic press key element fault diagnosis method and system, which comprises the following steps: collecting multi-source heterogeneous time sequence data reflecting element operation state in real time; preprocessing the multi-source heterogeneous time sequence data; extracting high-recognition fault features under different sensors, different elements and different dimensions from the preprocessed multi-source data; integrating the multi-source high-recognition fault features, analyzing and calculating characteristic parameters reflecting coupling correlation between key elements, and forming a fusion feature vector representing single element fault and multi-element coupling fault; constructing an adaptive deep learning model combined with a fuzzy neural network; inputting the fusion feature vector into the adaptive deep learning model to obtain a fault diagnosis result, and pushing the fault diagnosis result to an interactive interface in real time. The application can improve the accuracy and efficiency of fault diagnosis, and guarantee the stability, operation efficiency and reliability of equipment production.

Description

Technical Field

[0001] This application belongs to the field of fault diagnosis technology, and in particular relates to a fault diagnosis method and system for key components of a forging hydraulic press. Background Technology

[0002] Forging hydraulic presses are core equipment in the heavy manufacturing industry, widely used in the forming of large components in aerospace, shipbuilding, and other industries. Key components in their hydraulic systems, such as hydraulic pumps, relief valves, and hydraulic cylinders, directly determine the equipment's operational stability and machining accuracy. These components are prone to wear, seal failure, and jamming under long-term high-load conditions. Failure to diagnose these issues promptly can lead to increased forging scrap rates, equipment downtime, and even safety accidents. Therefore, a technical solution is urgently needed to improve the accuracy and efficiency of fault diagnosis, ensuring the stability, efficiency, and reliability of equipment production. Summary of the Invention

[0003] This application provides a method and system for fault diagnosis of key components of a forging hydraulic press, which can realize fault diagnosis of the forging hydraulic press, improve the accuracy and efficiency of fault diagnosis, and ensure the stability, operating efficiency and reliability of equipment production.

[0004] In a first aspect, this application provides a method for diagnosing faults in key components of a forging hydraulic press, including: By arranging multiple types of condition monitoring sensors on the key functional components of the forging hydraulic press, multi-source heterogeneous time-series data reflecting the operating status of the components are collected in real time; the key functional components are the core components in the forging hydraulic press system structure whose operating status directly affects the overall stability of the forging hydraulic press, the stability of the hydraulic system, and the processing accuracy; the key functional components include at least one of the following: hydraulic pump, relief valve, and hydraulic cylinder; Preprocessing of multi-source heterogeneous time-series data includes data cleaning, noise reduction, and adaptive normalization based on operating conditions. High-identity fault features from different sensors, components, and dimensions are extracted from preprocessed multi-source data. These multi-source high-identity fault features are integrated, and feature parameters reflecting the coupling relationship between key components are analyzed and calculated to form a fusion feature vector characterizing single-component faults and multi-component coupled faults. An adaptive deep learning model combining a fuzzy neural network is constructed; the fused feature vector is input into the adaptive deep learning model to obtain fault diagnosis results, the fault diagnosis results including at least fault status judgment, fault component location, fault type identification and fault severity assessment, and the fault diagnosis results are pushed to the interactive interface in real time to guide the user to perform equipment maintenance operations or confirm the trigger command of the automatic maintenance process of the equipment.

[0005] Secondly, embodiments of this application provide a fault diagnosis system for key components of a forging hydraulic press, applied to the fault diagnosis method for key components of a forging hydraulic press described in the first aspect. The system includes the following: The acquisition unit is configured to collect multi-source heterogeneous time-series data reflecting the operating status of key functional components of the forging hydraulic press in real time by arranging multiple types of status monitoring sensors on the key functional components; the key functional components are the core components in the forging hydraulic press system structure whose operating status directly affects the overall stability of the forging hydraulic press, the stability of the hydraulic system, and the processing accuracy; the key functional components include at least one of the following: hydraulic pump, relief valve, and hydraulic cylinder. The preprocessing unit is configured to preprocess multi-source heterogeneous time-series data, and the preprocessing includes data cleaning, noise reduction, and adaptive normalization based on operating conditions. The extraction unit is configured to extract high-identity fault features from different sensors, components, and dimensions from preprocessed multi-source data; integrate the multi-source high-identity fault features, analyze and calculate the feature parameters reflecting the coupling relationship between key components, and form a fusion feature vector characterizing single-component faults and multi-component coupled faults. The diagnostic unit is configured to construct an adaptive deep learning model that incorporates a fuzzy neural network; input the fused feature vector into the adaptive deep learning model to obtain fault diagnosis results, the fault diagnosis results including at least fault state judgment, fault component location, fault type identification and fault severity assessment, and push the fault diagnosis results to the interactive interface in real time to guide the user to perform equipment maintenance operations or confirm the trigger command of the automatic maintenance process of the equipment.

[0006] Thirdly, embodiments of this application provide an electronic device, the electronic device comprising: At least one processor, memory, and input / output unit; The memory is used to store computer programs, and the processor is used to call the computer programs stored in the memory to execute the fault diagnosis method for key components of the forging hydraulic press in the first aspect.

[0007] Fourthly, a computer-readable storage medium is provided, comprising instructions that, when executed on a computer, cause the computer to perform the fault diagnosis method for key components of a forging hydraulic press according to the first aspect.

[0008] The technical solution provided in this application involves arranging multiple types of condition monitoring sensors on key functional components of a forging hydraulic press to collect multi-source heterogeneous time-series data reflecting the operating status of these components in real time. These key functional components are core components in the forging hydraulic press system structure whose operating status directly affects the overall stability of the forging hydraulic press, the stability of the hydraulic system, and the machining accuracy. These key functional components include at least one of the following: a hydraulic pump, a relief valve, and a hydraulic cylinder. The multi-source heterogeneous time-series data is then preprocessed, including data cleaning, noise reduction, and adaptive normalization based on operating conditions. Furthermore, high-identifiability fault features from different sensors, components, and dimensions are extracted from the preprocessed multi-source data. These high-identifiability fault features are then integrated, and feature parameters reflecting the coupling correlation between key components are analyzed and calculated to form a fused feature vector characterizing single-component faults and multi-component coupled faults. Finally, an adaptive deep learning model combining fuzzy neural networks is constructed. The fused feature vectors are input into the adaptive deep learning model to obtain fault diagnosis results. The fault diagnosis results include at least fault state judgment, fault component location, fault type identification, and fault severity assessment. The fault diagnosis results are pushed to the interactive interface in real time to guide the user to perform equipment maintenance operations or confirm the trigger command of the equipment's automatic maintenance process. This technical solution, by combining an adaptive deep learning model with fuzzy neural networks, can achieve fault diagnosis of hydraulic presses, thereby improving the accuracy and efficiency of fault diagnosis and ensuring the stability, operating efficiency, and reliability of equipment production. Attached Figure Description

[0009] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the process system of a fault diagnosis method for key components of a forging hydraulic press according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a fault diagnosis system for key components of a forging hydraulic press according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, 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, 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.

[0011] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0012] To address at least one technical problem, embodiments of this application provide a method and system for diagnosing faults in key components of a forging hydraulic press.

[0013] The fault diagnosis scheme for key components of forging hydraulic presses provided in this application embodiment can also be executed by electronic devices, such as servers, server clusters, or cloud servers. These electronic devices can also be terminal devices such as mobile phones, computers, tablets, wearable devices, or dedicated devices (such as dedicated terminal devices with a fault diagnosis system for key components of forging hydraulic presses). These electronic devices can also carry the chips described in the above embodiments. Alternatively, these electronic devices can also install service programs for executing the fault diagnosis scheme for key components of forging hydraulic presses.

[0014] Figure 1 A schematic diagram of a fault diagnosis method for a key component of a forging hydraulic press provided in this application embodiment includes the following steps: Step 101: By arranging multiple types of condition monitoring sensors on the key functional components of the forging hydraulic press, multi-source heterogeneous time-series data reflecting the operating status of the components are collected in real time. Step 102: Preprocess the multi-source heterogeneous time series data. The preprocessing includes data cleaning, noise reduction, and adaptive normalization based on operating conditions. Step 103: Extract highly identifiable fault features from the preprocessed multi-source data under different sensors, different components, and different dimensions; Step 104: Integrate the multi-source high-identity fault features, analyze and calculate the feature parameters that reflect the coupling relationship between key components, and form a fusion feature vector that characterizes single component faults and multi-component coupled faults. Step 105: Construct an adaptive deep learning model that combines fuzzy neural networks; Step 106: Input the fused feature vector into the adaptive deep learning model to obtain the fault diagnosis result, and push the fault diagnosis result to the interactive interface in real time to guide the user to perform equipment maintenance operations or confirm the trigger command of the automatic maintenance process of the equipment.

[0015] In this embodiment, the key functional components are core components in the forging hydraulic press system structure whose operating state directly affects the overall stability of the forging hydraulic press, the stability of the hydraulic system, and the machining accuracy. Further optionally, the key functional components include at least one of the following: a hydraulic pump, a relief valve, and a hydraulic cylinder. By arranging various types of condition monitoring sensors on these components, multi-source heterogeneous time-series data reflecting the operating state of the components can be accurately collected, providing basic data support for subsequent diagnostics.

[0016] Optionally, preprocessing includes data cleaning, noise reduction, and adaptive normalization based on operating conditions. Data cleaning removes outliers and missing values ​​from the collected data, while noise reduction suppresses noise introduced by mechanical vibration and electromagnetic interference at the forging site. Adaptive normalization based on operating conditions dynamically adjusts parameters according to actual operating conditions, mapping sensor data to a standard range that matches the current operating conditions. This effectively eliminates the interference of operating condition changes on subsequent feature extraction, ensuring high reliability of the preprocessed multi-source data.

[0017] Optionally, the fault diagnosis results should include at least fault status assessment, fault component location, fault type identification, and fault severity evaluation. These results comprehensively reflect the operational health status of key components of the forging hydraulic press and are then pushed to the interactive interface in real time. This not only clearly guides users to perform corresponding equipment maintenance operations based on the severity of the fault but also allows users to confirm the trigger commands for the automatic maintenance process, achieving efficient integration of diagnostic results and maintenance actions.

[0018] As an optional embodiment, in step 102, a wavelet transform-based denoising algorithm and an adaptive filter are used to jointly process multi-source heterogeneous time-series data to obtain preprocessed data. This suppresses noise introduced by mechanical vibration and electromagnetic interference at the forging site while preserving the integrity of fault characteristics. Furthermore, the normalization parameters of the preprocessed data are dynamically adjusted according to the actual operating conditions, mapping the sensor data in the preprocessed data to a standard range matching the current operating conditions, thus obtaining preprocessed multi-source data to eliminate data interference caused by changes in operating conditions.

[0019] Specifically, to address the interference noise from the forging site mixed in multi-source heterogeneous time-series data, a wavelet transform-based denoising algorithm and an adaptive filter are used in conjunction. The wavelet transform denoising algorithm, based on the distribution differences of signal and noise at different scales in the wavelet domain, decomposes the multi-source heterogeneous time-series data to obtain wavelet coefficients at different frequency scales. Then, it suppresses the noisy wavelet coefficients according to a preset threshold rule, retaining effective coefficients related to fault characteristics and reconstructing the signal. The adaptive filter dynamically adjusts its filtering parameters by analyzing the characteristics of interference components in the data in real time, allowing the filtering process to adapt to changes in interference. The combined effect of these two methods can specifically suppress vibration interference caused by the operation of mechanical components and electromagnetic interference generated by electrical equipment at the forging site, while avoiding excessive denoising that could lead to the loss of fault characteristics. Ultimately, preprocessed data retaining complete fault characteristics is obtained.

[0020] After noise reduction, the normalization parameters are dynamically adjusted based on the actual operating conditions of the forging hydraulic press. The actual operating conditions mainly refer to key operating parameters such as the hardness of the forging material being processed and the set forging force. The normal fluctuation range of sensor data varies under different operating conditions. For example, the normal range of the hydraulic system pressure signal is higher when processing high-hardness forgings than when processing low-hardness forgings. In this case, the mean and standard deviation parameters in the normalization process need to be adjusted based on the historical normal data statistics of the corresponding operating conditions. Various sensor data such as vibration signals, pressure signals, flow signals, and temperature signals in the preprocessed data are uniformly mapped to a standard range that matches the current operating conditions. This ensures that the sensor data under different operating conditions have a consistent comparison benchmark, effectively eliminating the interference of data fluctuations caused by changes in operating conditions on subsequent feature extraction and fault diagnosis. Finally, preprocessed multi-source data is obtained. This data not only removes on-site interference noise but also eliminates the influence of differences in operating conditions, providing a high-quality data foundation for high-identity fault feature extraction in subsequent steps.

[0021] As an optional embodiment, step 103 involves extracting highly identifiable fault features from the preprocessed multi-source data under different sensors, components, and dimensions. This includes extracting multi-dimensional basic features corresponding to a single signal source from the vibration, pressure, flow, and temperature signals included in the preprocessed multi-source data. Furthermore, based on the operational correlation between the hydraulic pump, relief valve, and hydraulic cylinder under real-time operating conditions, interactive features between key components are extracted from the preprocessed multi-source data. Specifically, this includes extracting the linkage change features between the hydraulic pump output flow and the relief valve inlet pressure, the matching features between the relief valve pressure regulation value and the hydraulic cylinder movement speed, and the correspondence features between the hydraulic cylinder inlet / outlet pressure difference and the hydraulic pump working load. Finally, the multi-dimensional basic features of a single signal source and the interactive features between key components are screened, eliminating redundant and interfering features, and retaining features that are sensitive to faults and meet preset identification conditions to obtain highly identifiable fault features.

[0022] Specifically, firstly, for the vibration, pressure, flow, and temperature signals contained in the preprocessed multi-source data, multi-dimensional basic features corresponding to each signal source are extracted. The extraction process is based on the physical characteristics and fault characterization patterns of different signal types. For vibration signals, the focus is on capturing the signal's statistical characteristics in the time domain, its spectral distribution in the frequency domain, and its energy changes in the time and frequency domains. These features can reflect vibration anomalies caused by wear or jamming of key components. For pressure signals, the focus is on the stability duration of signal fluctuations and response delays. These features can reflect pressure changes caused by sealing failures or abnormal pressure regulation. For flow and temperature signals, the flow pulsation frequency deviation and the stable range of temperature change rates are extracted respectively. These parameters can correspond to early signs of abnormal component loads or heat dissipation failures. Multi-dimensional extraction ensures that fault information of each signal source is comprehensively captured.

[0023] Furthermore, based on the operational correlation between the hydraulic pump, relief valve, and hydraulic cylinder under real-time operating conditions, the interaction characteristics between key components are extracted. This process is based on the working principle of the hydraulic system. The output flow of the hydraulic pump directly affects the inlet pressure of the relief valve. When wear occurs in the hydraulic pump, causing flow fluctuations, the inlet pressure of the relief valve will change accordingly. By analyzing the changing trends of the two within the same time window, the linkage change characteristics can be extracted. The pressure regulation value of the relief valve determines the movement speed of the hydraulic cylinder. If the pressure regulation of the relief valve is abnormal, the speed of the hydraulic cylinder will deviate from the normal range. By comparing the matching degree between the pressure regulation value and the measured speed value, the matching characteristics can be extracted. There is a corresponding relationship between the inlet and outlet pressure difference of the hydraulic cylinder and the working load of the hydraulic pump. An increase in pump load will lead to a change in the cylinder pressure difference. By establishing the correlation curve between the two, corresponding characteristics can be extracted. These interaction characteristics can effectively reflect the fault propagation and coupling effects between multiple components.

[0024] Finally, the multi-dimensional basic features of a single signal source and the interaction features between key components are screened. The screening process is based on the sensitivity and identifiability of the features to faults. By analyzing the differences between features under normal and fault conditions, redundant features with minor differences are eliminated, and interference features caused by minor fluctuations in the field environment are also excluded. Only features that respond significantly to fault changes and can distinguish different fault types are retained. For example, the kurtosis value in vibration signals is sensitive to wear faults, and the response delay of pressure signals has high identifiability for sealing failures. These retained features together form highly identifiable fault features under different sensors, components, and dimensions, providing an accurate and efficient feature foundation for feature integration and feature vector construction in subsequent steps, effectively improving the accuracy and efficiency of subsequent fault diagnosis.

[0025] Optionally, in step 103, multi-dimensional basic features corresponding to a single signal source are extracted from the vibration signal, pressure signal, flow signal, and temperature signal included in the preprocessed multi-source data. These features include: for vibration signals, time-domain statistical features, frequency-domain spectral features, and time-frequency-domain energy features. The time-domain statistical features include peak value, root mean square (RMS), kurtosis, and skewness. The frequency-domain spectral features include peak value, center frequency, and frequency band energy percentage. The time-frequency-domain energy features include the energy values ​​of each frequency band after wavelet packet decomposition. For pressure signals, pressure fluctuation amplitude, pressure stabilization duration, and pressure response delay time are extracted. For flow signals, flow pulsation frequency, flow deviation value, and instantaneous peak flow are extracted. For temperature signals, temperature change rate, temperature stabilization range, and over-temperature duration are extracted.

[0026] Specifically, for vibration signals, time-domain statistical features, frequency-domain spectral features, and time-frequency-domain energy features are extracted based on the differences in fault characterization across different dimensions. Time-domain statistical features focus on the overall fluctuation characteristics of the signal over time. By capturing the signal peak value, the instantaneous vibration intensity of the component is reflected; the root mean square reflects the average level of vibration energy; and kurtosis and skewness can identify abnormal deviations in signal distribution. For example, when a hydraulic pump bearing is worn, the kurtosis of the vibration signal will significantly increase, and the skewness will also change due to asymmetric vibration caused by wear particle impact. Frequency-domain spectral features are extracted by analyzing the frequency distribution pattern of the signal. The spectral peak value corresponds to the characteristic frequency of the component fault, the center frequency reflects the concentrated frequency band of the main vibration energy, and the frequency band energy proportion can distinguish the energy distribution differences between normal vibration and fault vibration. For example, when the overflow valve core is stuck, the characteristic frequency of its vibration signal will show a new spectral peak value, and the corresponding frequency band energy proportion will also increase significantly. Time-frequency-domain energy features utilize wavelet packet decomposition technology to decompose the vibration signal into different frequency bands and extract the energy values ​​of each band. This method can effectively capture the dynamic changes of vibration energy in the time-frequency domain during fault development, and is particularly suitable for the feature identification of early micro-wear faults.

[0027] For pressure signals, features are extracted based on their pressure variation patterns under component failures. Pressure fluctuation amplitude reflects the range of pressure signal fluctuation under stable operating conditions; when a micro-leak occurs in the hydraulic cylinder seals, the pressure fluctuation amplitude will be significantly larger than under normal operating conditions. Pressure stabilization time records the time it takes for the pressure to reach a stable value from a fluctuating state; wear on the pressure regulating components of the relief valve will prolong the pressure stabilization time. Pressure response delay time is the interval from the issuance of the control command to the actual pressure change; hydraulic pump flow regulation failure will increase the pressure response delay time. These features are directly related to critical states such as the sealing performance and adjustment accuracy of pressure-related components.

[0028] For flow signals, features are extracted by combining their correlation with component load and operating status. Flow pulsation frequency is the frequency of periodic fluctuations in the flow signal. Wear of hydraulic pump blades can cause abnormal deviations in flow pulsation frequency. Flow deviation is the difference between the actual flow and the rated flow. Abnormal overflow of the relief valve can cause the flow deviation to exceed the normal range. Instantaneous flow peak is the maximum flow in a short period of time. Sudden changes in the speed of hydraulic cylinder movement or pipeline blockage can cause an abnormal increase in instantaneous flow peak. These features can accurately reflect the working status of flow-related components.

[0029] For temperature signals, features are extracted based on their relationship with the heating and cooling states of components. The rate of temperature change is the amount of temperature change per unit time; insufficient lubrication of the hydraulic pump bearings will accelerate the rate of temperature change. The stable temperature range is the time period during which the temperature remains within the normal range; blockage of the cooling channels in the hydraulic cylinder will shorten the stable temperature range. The duration of overheating is the cumulative time the temperature exceeds the normal threshold; prolonged overload of the relief valve will increase the duration of overheating. These features can promptly detect temperature anomalies caused by component friction overload, providing support for early fault warning. By extracting the above multi-dimensional basic features from the four types of signals, the characteristic manifestations of different fault types of key components can be comprehensively covered, laying a rich and accurate foundation for subsequent interactive feature extraction and feature screening, further improving the accuracy of overall fault diagnosis and early identification capabilities.

[0030] As an optional embodiment, in step 104, the multi-source high-identity fault features are integrated, and the feature parameters reflecting the coupling correlation between key components are analyzed and calculated to form a fused feature vector characterizing single-component faults and multi-component coupled faults. Specifically, in the first step, the high-identity fault features extracted from different sensors, different components, and different dimensions are classified and labeled, and divided into hydraulic pump feature group, overflow valve feature group, and hydraulic cylinder feature group according to their respective key components. Each group of features includes the time-domain statistical features, frequency-domain spectral features, and time-frequency-domain energy features of the vibration signal of the corresponding component, as well as the feature parameters corresponding to the pressure signal, flow signal, and temperature signal. In the second step, the correlation weight of the features within each key component is calculated, and the linear correlation degree between different features within the same component is analyzed based on the Pearson correlation coefficient, such as the correlation coefficient between the kurtosis value of the hydraulic pump vibration signal and the flow pulsation frequency of the flow signal. Features with a correlation coefficient greater than a preset correlation threshold are classified into the core feature subset of the component to reduce feature redundancy within a single component. In the third step, the coupling correlation feature parameters between key components are analyzed and calculated, and a component coupling relationship model is constructed based on the working principle of the hydraulic system. The dynamic matching degree between the hydraulic pump output flow characteristics and the relief valve inlet pressure characteristics is calculated. The dynamic matching degree is the ratio of the rate of change of the hydraulic pump output flow to the rate of change of the relief valve inlet pressure within a certain time window. If the ratio exceeds the matching range under normal operating conditions, it is marked as an abnormal coupling characteristic. The response synchronicity between the relief valve pressure regulation value characteristics and the hydraulic cylinder movement speed characteristics is calculated. Response synchronicity is the time difference between the relief valve pressure regulation command and the change in hydraulic cylinder speed. If the time difference is greater than a preset response threshold, it is marked as a coupling delay characteristic. The correlation strength between the hydraulic cylinder inlet and outlet pressure difference characteristics and the hydraulic pump working load characteristics is calculated. The correlation strength is the linear regression coefficient between the hydraulic cylinder inlet and outlet pressure difference and the hydraulic pump working load. If the coefficient deviates from the standard coefficient range, it is marked as a coupling imbalance characteristic. In the fourth step, the feature dimensions of the core feature subset of the components and the coupling correlation feature parameters between components are unified. Principal component analysis is used to map features of different dimensions to the same feature space, retaining principal components whose variance contribution in the feature space is greater than a preset variance threshold. The fifth step involves concatenating the core features of the components and the coupled features after unifying the dimensions according to preset weights. The weights of the core features are determined based on the degree of influence of the components on the overall stability of the forging hydraulic press. The weights of the hydraulic pump features are higher than those of the relief valve features, and the weights of the relief valve features are higher than those of the hydraulic cylinder features. The weights of the coupled features are adjusted according to the current forging force conditions. The greater the forging force, the higher the weights of the coupled features. Finally, a fused feature vector is formed that can simultaneously characterize single component failures (reflected by the core features of the components) and multi-component coupled failures (reflected by the coupled features).

[0031] Specifically, the process of integrating multi-source high-identity fault features, analyzing and calculating feature parameters reflecting the coupling correlation between key components, and forming a fused feature vector in step 104 is as follows: First, the highly identifiable fault features extracted from different sensors, components, and dimensions were classified and labeled. The classification was based on the key component to which the feature belonged, thus dividing the system into hydraulic pump feature groups, relief valve feature groups, and hydraulic cylinder feature groups. Each feature group fully encompasses all types of signal features of the corresponding component. The hydraulic pump feature group covers the time-domain statistical features, frequency-domain spectral features, and time-frequency domain energy features of its vibration signal, as well as the feature parameters corresponding to pressure, flow, and temperature signals. The relief valve and hydraulic cylinder feature groups also integrate the full-dimensional features of their corresponding components according to the same logic. Through classification and labeling, the scattered features are formed into a structured feature set, laying the foundation for subsequent targeted processing.

[0032] Next, the correlation weights of the internal features of each key component are calculated. This process is based on the Pearson correlation coefficient to analyze the degree of linear correlation between different features within the same component. For example, the correlation between the kurtosis value of the vibration signal and the flow pulsation frequency of the flow signal in the hydraulic pump feature group is analyzed. If the two show a significant synchronous change trend due to hydraulic pump wear failure, their correlation coefficient will be at a high level. Features with a correlation coefficient greater than the preset correlation threshold are classified into the core feature subset of the component, and redundant features with low correlation are eliminated. For example, the temperature stability range feature in the hydraulic pump temperature signal that changes slowly and has weak correlation with other features can be excluded. In this way, feature redundancy within a single component is greatly reduced, ensuring that the features input to the subsequent model are both representative and concise.

[0033] Then, based on the working principle of the hydraulic system, a component coupling relationship model is constructed, and the coupling correlation characteristic parameters between key components are analyzed and calculated. For the hydraulic pump and relief valve, the dynamic matching degree between the two is calculated. This parameter is obtained by the ratio of the rate of change of the hydraulic pump output flow to the rate of change of the relief valve inlet pressure within a certain time window. When the hydraulic pump output flow fluctuates due to blade wear, the relief valve inlet pressure will change accordingly. If the ratio of the two rates of change exceeds the stable range under normal operating conditions, it is marked as an abnormal coupling characteristic. For the relief valve and hydraulic cylinder, the response synchronization between the two is calculated. This parameter is the time difference between the relief valve receiving the pressure regulation command and the hydraulic cylinder's movement speed changing accordingly. If the relief valve experiences a pressure regulation delay due to valve core jamming, this time difference will be greater than the preset response threshold, which is marked as a coupling delay characteristic. For hydraulic cylinders and hydraulic pumps, the correlation strength between the two is calculated. This parameter is the linear regression coefficient of the pressure difference between the inlet and outlet of the hydraulic cylinder and the working load of the hydraulic pump. If the abnormal adjustment of the hydraulic pump load causes the pressure difference of the cylinder to deviate from the normal correlation law, the coefficient will exceed the standard coefficient range, which is marked as a coupling imbalance feature. These coupling correlation feature parameters accurately capture the fault propagation and mutual influence law between multiple components.

[0034] Subsequently, the feature dimensions of the core feature subset of components and the coupling correlation feature parameters between components are unified using principal component analysis (PCA). Both types of features are input into the PCA model, which uses linear transformation to map high-dimensional features to the same low-dimensional feature space. Simultaneously, principal components with a variance contribution greater than a preset variance threshold are retained. Principal components with a high variance contribution indicate richer fault information; for example, principal components that simultaneously reflect the core features of a hydraulic pump and the pump-valve coupling features are preferentially retained. Dimensional unification eliminates dimensional differences and redundancy between different types of features, providing a basis for direct feature fusion.

[0035] Finally, the unified-dimensional core features of the components and the coupled features are concatenated according to preset weights. The weights are dynamically adjusted based on the importance of the two types of features. The weights of the core features are determined based on the component's impact on the overall stability of the forging hydraulic press. As the power source of the hydraulic system, the hydraulic pump has a higher feature weight than the relief valve responsible for pressure regulation, and the relief valve has a higher feature weight than the hydraulic cylinder that performs the action. The weights of the coupled features are adjusted according to the current forging force conditions. The greater the forging force, the stronger the load transfer and interaction between components, and the higher the risk of coupling failure. Therefore, the weights of the coupled features are increased accordingly. For example, the weights of the coupled features are significantly higher when machining high-hardness forgings than when machining low-hardness forgings. Through weighted concatenation, the final fused feature vector can accurately characterize single-component failures through the core features of the components and effectively reflect multi-component coupling failures through the coupled features, providing comprehensive and condition-adaptive feature inputs for accurate diagnosis by the subsequent adaptive deep learning model.

[0036] As an optional embodiment, in step 105, constructing an adaptive deep learning model combined with a fuzzy neural network includes: pre-training an adaptive deep learning model constructed using an adaptive neurofuzzy inference system with multi-scenario fault sample data; the multi-scenario fault sample data covers fault data under different models of forging hydraulic presses, different forging materials, and different forging force conditions, and the sample data includes single-component fault samples and multi-component coupled fault samples. During the pre-training process, the fault diagnosis accuracy, fault location accuracy, and multi-component coupled fault recognition rate are used as objective functions to optimize the fuzzy membership function parameters of the adaptive neurofuzzy inference system layer and the initial values ​​of the multi-head attention weights of the Transformer layer. The adaptive neural fuzzy inference system layer is based on the failure mechanism of key components in forging hydraulic presses. It automatically generates a fuzzy rule base for fault diagnosis through data-driven processing. This fuzzy rule base contains the association logic between multi-dimensional features in different fault types and multi-scenario fault sample data. The different fault types are wear, seal failure, and jamming. The multi-dimensional features include time-domain statistical features, frequency-domain spectral features, and component interaction features. Furthermore, the rule generation process incorporates feature association patterns under different forging materials and forging force conditions. The contribution weight of each fault feature to the fault category is calculated using a fuzzy membership function, obtaining a fuzzy representation of the fault features as a fuzzy feature matrix. The adaptive deep learning model adopts a Transformer network architecture. The input is the fuzzy feature matrix. A multi-head attention mechanism associates fuzzy feature data corresponding to different sensors and key components. The fuzzy feature data includes hydraulic pump flow characteristics, relief valve pressure characteristics, and hydraulic cylinder temperature characteristics. Different sensors and key components correspond to different fuzzy feature data. A global feature association matrix is ​​constructed based on the fuzzy feature data to capture the coupling relationships between multiple components. Then, an encoder layer extracts deep fault information from the global feature association matrix to complete the preliminary diagnosis of the fault state. The adaptive deep learning model's parameters are fine-tuned using a preset number of target application scenario samples. This preset number is determined based on the differences in device models across target application scenarios; if the difference between the device models in the target application scenario and those in the pre-training samples exceeds a preset threshold, the number of target application scenario samples is increased. During fine-tuning, the fuzzy rule base of the adaptive neural fuzzy inference system layer and the attention head weights of the Transformer layer are optimized. Optimizing the fuzzy rule base of the adaptive neural fuzzy inference system layer involves supplementing the fault feature association rules specific to the target scenario, while optimizing the attention head weights of the Transformer layer involves adjusting the attention priority of different features in the target scenario, allowing the model to initially adapt to the target scenario. Transfer learning and online optimization mechanisms are employed, combining the equipment model of the forging hydraulic press with real-time collected operating condition data to optimize the model parameters of the adaptive deep learning model. The equipment model parameters of the forging hydraulic press include rated pressure, displacement, and motor power, while the real-time operating condition data includes pressure fluctuation amplitude and flow pulsation frequency.

[0037] Specifically, in step 105, the implementation process of constructing an adaptive deep learning model combined with a fuzzy neural network is as follows: First, the adaptive deep learning model constructed using an adaptive neurofuzzy inference system is pre-trained with multi-scenario fault sample data. The multi-scenario fault sample data covers fault data under different forging hydraulic presses of different models, different forging materials, and different forging force conditions. It includes single-component fault samples such as hydraulic pump wear and relief valve sealing failure, as well as multi-component coupled fault samples such as hydraulic pump failure causing abnormal relief valve pressure. In the pre-training stage, the fault diagnosis accuracy, fault location accuracy, and multi-component coupled fault recognition rate are used as objective functions. The fuzzy membership function parameters of the adaptive neurofuzzy inference system layer and the initial values ​​of the multi-head attention weights of the Transformer layer are continuously iteratively optimized. For example, for sealing failure faults under different forging materials, the contribution weight of the pressure response delay feature in the fuzzy membership function to the fault category is adjusted. At the same time, the attention weight of the Transformer layer to the correlation features of pressure signal and temperature signal is optimized to ensure that the model has a preliminary fault diagnosis capability in various scenarios.

[0038] The adaptive neuro-fuzzy inference system is built upon the fault mechanism of key components in forging hydraulic presses. It automatically generates a fuzzy rule base for fault diagnosis through data-driven processing. This fuzzy rule base stores the association logic between different fault types and multi-dimensional features. Different fault types include wear, seal failure, and jamming. Multi-dimensional features encompass time-domain statistical features, frequency-domain spectral features, and component interaction features. The rule generation process incorporates feature association patterns under different forging materials and forging force conditions. For example, under high-hardness forging processing conditions, the correlation strength between increased vibration signal kurtosis and hydraulic pump wear is increased. The contribution weight of each fault feature to the fault category is calculated using a fuzzy membership function, transforming the dispersed fault features into a fuzzy feature matrix, thus achieving fuzzy representation of fault features. This approach effectively handles the uncertainty of fault features and avoids the omission of weak fault signals by traditional crisp features.

[0039] The adaptive deep learning model adopts the Transformer network architecture. The model input is the aforementioned fuzzy feature matrix. Through a multi-head attention mechanism, it associates the fuzzy feature data corresponding to different sensors and key components. Different sensors and key components correspond to exclusive fuzzy feature data. For example, the hydraulic pump flow feature is converted from the data collected by the flow sensor, the relief valve pressure feature is generated from the data collected by the pressure sensor, and the hydraulic cylinder temperature feature comes from the fuzzification processing result of the temperature sensor. The multi-head attention mechanism can simultaneously capture the correlation between multiple sets of features. Based on these fuzzy feature data, a global feature correlation matrix is ​​constructed, such as capturing the dynamic correlation between the fuzzy features of the hydraulic pump flow and the fuzzy features of the relief valve pressure, so as to accurately capture the coupling correlation between multiple components. The encoder layer performs deep processing on the global feature correlation matrix to extract the deep fault information contained in the matrix. For example, it can identify the abnormal relief valve pressure caused by hydraulic pump wear from the pump-valve coupling features, and complete the preliminary diagnosis of the fault state.

[0040] Next, the parameters of the adaptive deep learning model are fine-tuned using a preset number of target application scenario samples. The preset number is determined based on the difference in equipment models between the target application scenarios and the pre-training samples. If the equipment models in the target application scenarios differ significantly from those in the pre-training samples, for example, if the equipment type in the target application scenarios is a large-tonnage forging hydraulic press while the pre-training samples are mainly small and medium-tonnage equipment, the number of target application scenario samples is increased to ensure the fine-tuning effect. During the fine-tuning process, the focus is on optimizing the fuzzy rule base of the adaptive neural fuzzy inference system layer and the attention head weights of the Transformer layer. When optimizing the fuzzy rule base, fault feature association rules specific to the target scenario are added. For example, based on the structural characteristics of the hydraulic pump of the target equipment, association rules between flow pulsation frequency and wear faults are added. When optimizing the attention head weights, the attention priority of different features in the target scenario is adjusted. For example, if the heat dissipation performance of the hydraulic cylinder of the target equipment is weak, the attention weight of the temperature change rate feature in fault diagnosis is increased, so that the model can initially adapt to the equipment characteristics and working conditions of the target application scenario.

[0041] Optionally, when determining the required number and type of target application scenario samples based on equipment type differences, a multi-dimensional evaluation mechanism for equipment type differences must first be constructed. This mechanism revolves around the core dimensions affecting fault characteristic performance, including but not limited to: differences in equipment rated parameters (such as rated pressure, displacement, and motor power), differences in core component structure (such as the number of blades in a hydraulic pump, the adjustment method of the relief valve, and the sealing structure of the hydraulic cylinder), and differences in operating condition adaptation range (such as maximum forging force and the type of compatible forging material). This mechanism quantifies the degree of difference between the target equipment and the equipment in the pre-training samples. For example, the analytic hierarchy process (AHP) can be used to assign weights to each evaluation dimension. Rated pressure and core component structure have a greater impact on fault characteristics, so their weights are set at a higher level, followed by the operating condition adaptation range. Then, by calculating the deviation rate of the target equipment and the pre-training sample equipment in each dimension parameter, and combining the weights, a comprehensive difference value is obtained, which serves as the core basis for sample increment.

[0042] For example, if the overall difference value is in the mild range (e.g., difference ≤ 30%), it indicates that the fault characteristics of the target equipment and the pre-training sample equipment are highly similar. Only 20%-30% of the samples need to be added to the original preset sample size to supplement fault samples corresponding to the slight structural differences unique to the target equipment. If the difference value is in the moderate range (30% < difference ≤ 60%), it means that there are significant differences in rated parameters or core component structures, and the fault characteristics will show obvious differences. The sample size needs to be increased by 50%-80%, focusing on supplementing samples of fault characteristic changes caused by parameter differences. If the difference value is in the severe range (difference > 60%), such as when the target equipment is a large-tonnage forging hydraulic press (rated pressure 31.5MPa, displacement 200mL / r) while the pre-training samples are small-to-medium-tonnage equipment (rated pressure 16MPa, displacement 80mL / r), the two differ greatly in hydraulic system load bearing and fault propagation paths. The sample size needs to be increased by 100%-150% to ensure coverage of fault characteristics under all operating conditions of the target equipment.

[0043] For example, to address differences in rated parameters, if the target equipment has a higher rated pressure, it is necessary to supplement fault samples specific to high-pressure operating conditions, such as high-pressure leakage fault samples caused by hydraulic pump shaft seal wear under high-pressure conditions, and overflow valve jamming fault samples during high-pressure regulation. To address differences in core component structures, if the target equipment's hydraulic cylinder uses a combined sealing structure (different from the single-lip seals in the pre-training samples), it is necessary to supplement samples of sealing failure faults specific to this sealing structure, such as samples of slow pressure drop caused by localized wear of the combined seals. To address differences in operating condition adaptability, if the target equipment mainly processes special steel (the pre-training samples mainly use ordinary carbon steel), it is necessary to supplement samples of multi-component coupling faults under high forging forces, such as coupling fault samples where hydraulic pump wear under high load causes overflow valve pressure fluctuations, leading to hydraulic cylinder movement speed deviations. Simultaneously, the sample types must consider both single-component faults and multi-component coupling faults, and cover common fault types of the target equipment (wear, seal failure, jamming), ensuring that the supplemented samples can learn the unique fault characteristic correlation patterns of the target equipment, ultimately achieving model adaptation to the target application scenario.

[0044] Finally, a transfer learning and online optimization mechanism was adopted to continuously optimize the model parameters by combining the equipment model parameters of the forging hydraulic press with the real-time collected operating condition data. The equipment model parameters include rated pressure, displacement, and motor power, while the real-time operating condition data covers pressure fluctuation amplitude, flow pulsation frequency, and more. During the transfer learning phase, general modules in the pre-trained model that are not related to the equipment model, such as the fuzzy membership function basic structure and the Transformer encoder architecture, were transferred to the target model. Only parameters strongly related to the equipment model, such as the attention weights of the Transformer layer on flow characteristics for a specific displacement hydraulic pump, were retrained to reduce the sample requirements for adapting the model to new equipment. During the online optimization phase, the deviation between the model's diagnostic results and the actual fault conditions was monitored in real time. If it was found that the model's accuracy in diagnosing jamming faults under a certain forging force condition decreased, the association rules between flow deviation features and jamming faults in the fuzzy rule base were dynamically updated based on the real-time collected operating condition data such as flow pulsation frequency and pressure fluctuation amplitude. At the same time, the attention weights of the Transformer layer on the coupling features of flow and pressure signals were adjusted to ensure that the model can continuously adapt to changes in operating conditions and maintain high diagnostic accuracy during long-term operation of the equipment.

[0045] As an optional embodiment, in step 106, the fused feature vector is input into the adaptive deep learning model to obtain the fault diagnosis result, including: splitting the fused feature vector into a component-independent feature subset and a component-coupled feature subset. The component-independent feature subset contains single-component features corresponding to each key component, including time-domain statistical features of vibration signals, fluctuation amplitude of pressure signals, etc. The component-coupled feature subset contains cross-component correlation features, including the linkage feature between hydraulic pump output flow and relief valve inlet pressure, and the matching feature between relief valve pressure regulation value and hydraulic cylinder movement speed. The component-independent feature subset is input into the ANFIS layer of the adaptive deep learning model. The ANFIS layer calculates the fuzzy membership degree of the component-independent feature subset based on the pre-trained fuzzy rule base, and outputs the preliminary confidence level of any fault among wear, sealing failure, and jamming for each component. The component-coupled feature subset is input into the Transformer layer of the model. The Transformer layer uses a multi-head attention mechanism to... A global correlation analysis is performed on a subset of component coupling features to select the coupling features with the highest correlation to the fault, outputting the range of associated components for multi-component coupling faults and the corresponding fault type confidence. The confidence of single-component faults output by the ANFIS layer and the confidence of coupled faults output by the Transformer layer are weighted and fused. The weight coefficient of the coupled fault confidence is dynamically adjusted according to the current forging force condition and forging material. The greater the forging force and the higher the hardness of the forging material, the higher the weight coefficient of the coupled fault confidence. The fault status is determined based on the confidence result after weighted fusion. When the confidence of any fault type is greater than a preset confidence threshold, a fault is determined to exist. The fault type with the highest confidence is extracted as the final fault type. Combined with the range of associated components output by the Transformer layer, the fault component location result is determined. The built-in fault severity assessment module of the model is called to generate corresponding assessment indicators, and the results are integrated to form a fault diagnosis result that includes fault status judgment, fault component location, fault type identification, and fault severity assessment.

[0046] Specifically, in step 106, the fused feature vector formed in step 104 is first split into a hierarchical subset to obtain a component-independent feature subset and a component-coupled feature subset. The component-independent feature subset covers the single-component features corresponding to each key component, such as the time-domain statistical features of the vibration signal of the hydraulic pump, the pressure signal fluctuation amplitude of the relief valve, and the temperature change rate of the hydraulic cylinder. These features only reflect the operating status of a single component. The component-coupled feature subset contains cross-component correlation features, such as the linkage feature between the hydraulic pump output flow and the relief valve inlet pressure, and the matching feature between the relief valve pressure regulation value and the hydraulic cylinder movement speed. These features reflect the interaction relationship between multiple components. By splitting, different types of features can be adapted to the processing needs of different levels of the model.

[0047] Next, the component-independent feature subset is input into the ANFIS layer of the adaptive deep learning model. The ANFIS layer calls the fuzzy rule library generated in the pre-training stage. This rule library has stored the association logic between different fault types and single component features, such as the association rule between increased vibration signal kurtosis and hydraulic pump wear, and the association rule between abnormal pressure fluctuation amplitude and relief valve seal failure. Based on these rules, the ANFIS layer calculates the fuzzy membership degree of each feature in the component-independent feature subset, determines the probability that each feature belongs to the three types of faults: wear, seal failure, and jamming, and finally outputs the preliminary confidence level of each key component for each type of fault, such as the confidence level of hydraulic pump wear fault and relief valve seal failure fault, thus completing the preliminary judgment of single component faults.

[0048] Simultaneously, a subset of component coupling features is input into the Transformer layer of the model. The Transformer layer performs global correlation analysis on the coupling features through a multi-head attention mechanism. This mechanism can simultaneously focus on the correlation strength between multiple sets of coupling features. For example, it focuses on analyzing the correlation between the linkage features of hydraulic pump flow and relief valve pressure and the matching features between relief valve pressure and hydraulic cylinder speed, and filters out the coupling features with the highest correlation to the fault. For example, when fluctuations in hydraulic pump flow cause abnormal relief valve pressure, which in turn leads to deviations in hydraulic cylinder speed, the Transformer layer will prioritize capturing this set of coupling features, and then output the range of associated components for multi-component coupling faults, clarifying the component combinations involved in the fault, and outputting the confidence level of the corresponding fault type for the coupling fault. For example, it outputs the confidence level of coupling faults involving hydraulic pump and relief valve that belong to the category of abnormal pressure caused by wear.

[0049] Then, the single-component fault confidence scores output by the ANFIS layer and the coupled fault confidence scores output by the Transformer layer are weighted and fused. The weight coefficients are dynamically adjusted according to the current forging force conditions and forging material. The principle is that the greater the forging force, the higher the hardness of the forging material, the greater the load on the component, and the higher the probability of coupled faults caused by load transmission between multiple components. Therefore, it is necessary to increase the weight coefficient of the coupled fault confidence score. For example, when processing high-hardness special steel and the forging force is large, the weight of the coupled fault confidence score will be significantly higher than that of the single-component fault confidence score. However, when processing low-hardness ordinary steel and the forging force is small, the weight of the coupled fault confidence score will be appropriately reduced. Dynamic weighting ensures that the fusion result can adapt to the fault occurrence pattern under the current working conditions.

[0050] Finally, the fault status is determined based on the weighted fusion confidence score. When the confidence score of any fault type exceeds a preset confidence threshold, a fault is identified in a key component of the forging hydraulic press. The type with the highest confidence score is extracted from all fault types as the final fault type. For example, if the confidence score of hydraulic pump wear fault is the highest, the final fault type is determined to be hydraulic pump wear. Combining the range of associated components output by the Transformer layer, the components involved in the fault are identified, completing the fault component localization. Simultaneously, the built-in fault severity assessment module is invoked. This module generates assessment indicators based on key parameters in the fusion features (such as the root mean square vibration signal corresponding to wear fault and the flow deviation value corresponding to jamming fault). Finally, the fault status judgment results, fault component localization results, fault type identification results, and fault severity assessment indicators are integrated to form a complete fault diagnosis result, providing a comprehensive basis for subsequent maintenance operations.

[0051] Optionally, in the above steps, the various features in the component-independent feature subset are first preprocessed and adapted to ensure that the feature format is consistent with the input requirements of the ANFIS layer. The component-independent feature subset includes vibration signal time-domain statistical features (such as peak value and root mean square), pressure signal fluctuation amplitude, and temperature change rate. Due to differences in acquisition dimensions and units, these features need to be normalized and mapped to the preset feature input range of the ANFIS layer. This range is consistent with the value range of features in the fuzzy rule base during the pre-training stage; for example, the root mean square value of the vibration signal is mapped to the 0-1 range to avoid affecting the accuracy of fuzzy membership calculation due to differences in units. Next, the features are matched against the fuzzy rule base. The ANFIS layer will call the corresponding component's exclusive fuzzy rule subset from the rule base based on the component type (hydraulic pump, relief valve, hydraulic cylinder) to which the features in the component-independent feature subset belong. If the input is the kurtosis value of the vibration signal and the deviation value of the flow signal of a hydraulic pump, the ANFIS layer will automatically locate the "hydraulic pump fault diagnosis rule subset" in the rule base. This subset contains pre-trained association rules such as "high vibration signal kurtosis value → hydraulic pump wear" and "large flow signal deviation value → hydraulic pump blade failure". At the same time, based on the feature type (such as vibration feature, flow feature), it further filters out the rules in the rule base that are directly related to the current input feature and eliminates interference from irrelevant rules. For example, when the input is the fluctuation amplitude of the pressure signal, only the fault rules related to the pressure feature are matched. Then, the fuzzy membership degree calculation process is started. For each independent input feature, the ANFIS layer calls the corresponding fuzzy membership degree function (such as triangular membership degree function, Gaussian membership degree function) to calculate the membership degree value of the feature belonging to different fault types (wear, seal failure, jamming). Taking the root mean square (RMS) feature of the vibration signal of a hydraulic pump as an example, if the feature value is located at the boundary between the feature intervals corresponding to "light wear" and "moderate wear," the membership function will calculate its membership degree for "light wear" (e.g., 0.3) and "moderate wear" (e.g., 0.7) respectively. Simultaneously, the membership value is adjusted by combining the association weights between this feature and the fault type in the rule base. For example, if the pre-training stage determines that the RMS of the vibration signal contributes more to wear faults, its corresponding membership value for wear faults will be appropriately increased. Finally, the membership results of multiple independent features of the same component are fused. The ANFIS layer will perform a weighted summation of the membership values ​​of different features of the same component based on the feature importance weights determined in the pre-training stage.Taking the relief valve as an example, if the input independent features include pressure signal fluctuation amplitude (membership: sealing failure 0.8, wear 0.1) and temperature change rate (membership: sealing failure 0.6, wear 0.2), and the importance weight of pressure fluctuation amplitude is 1.2 times that of temperature change rate, then after fusion, the membership degree of the relief valve belonging to the sealing failure fault is (0.8×1.2+0.6×1)÷(1.2+1)≈0.71, and the membership degree belonging to the wear fault is (0.1×1.2+0.2×1)÷(1.2+1)≈0.15. Finally, the preliminary confidence of the component for each type of fault is output, completing the input and processing flow of the component independent feature subset in the ANFIS layer.

[0052] As an optional embodiment, in step 106, the fault severity assessment module built into the model is invoked to generate corresponding assessment indicators, including: the fault severity assessment module extracts feature parameters strongly correlated with the final fault type from the fused feature vector; if the final fault type is wear, the kurtosis value and root mean square value in the time-domain statistical features of the vibration signal are extracted; if it is sealing failure, the pressure fluctuation amplitude and pressure response delay time of the pressure signal are extracted; if it is jamming, the flow pulsation frequency and flow deviation value of the flow signal are extracted; a mapping model between feature parameters and fault severity is constructed based on multi-scenario fault sample data. The multi-scenario fault sample data covers the full-cycle feature parameter change curves of different models of forging hydraulic presses under different forging force conditions and different forging materials, from the initial stage to the severe stage. The mapping model sets three levels of feature parameter thresholds: the feature parameter values ​​in the initial stage are between the upper limit of the normal range and the slightly abnormal threshold; the values ​​in the moderate stage are between the slightly abnormal threshold and the severely abnormal threshold; and the values ​​in the severe stage exceed the severely abnormal threshold; the currently extracted feature parameter values ​​are substituted into the mapping model to match and obtain the initial level of fault severity; historical operating data of the faulty component is retrieved and combined with the component's operating data. The initial severity level is dynamically corrected based on historical fault records. If the operating time of a faulty component exceeds the rated service life specified in the parameters of the corresponding equipment model and the initial severity level is mild, the initial severity level is adjusted to moderate. If the faulty component has a historical fault record with the same fault type as the current fault and the initial severity level is moderate, the initial severity level is adjusted to severe. An evaluation index is generated based on the corrected fault severity level. The evaluation index is divided into three levels: mild, moderate, and severe faults. Each level corresponds to different maintenance guidance information. Mild faults correspond to low maintenance priority and prompt regular monitoring of characteristic parameter changes. The monitoring cycle is determined based on the current forging task frequency. Moderate faults correspond to medium maintenance priority and suggest maintenance during the next production break. A list of related components to be checked is output. Severe faults correspond to high maintenance priority and simultaneously output fault risk warnings and recommended maintenance cycles. Fault risk warnings include consequences such as hydraulic system pressure imbalance or forging machining accuracy deviations that may result from continued operation of the fault. Recommended maintenance cycles are determined based on the current forging task volume and the estimated remaining life of the component. The estimated remaining life of the component refers to the historical replacement cycle of the component model and the wear rate under the current operating conditions.

[0053] Specifically, in step 106, the evaluation module first accurately extracts feature parameters strongly correlated with the final fault type from the fused feature vector formed in step 104. The extraction logic is based on the physical relationship between the fault type and the feature parameters. If the final fault type is wear, wear will cause changes in the vibration characteristics of the component, so the kurtosis value and root mean square value in the time-domain statistical features of the vibration signal are extracted. If it is a sealing failure, sealing problems will cause pressure instability and response delay, so the pressure fluctuation amplitude and pressure response delay time of the pressure signal are extracted. If it is jamming, jamming will cause abnormal flow transmission, so the flow pulsation frequency and flow deviation value of the flow signal are extracted. By selectively extracting key parameters, it is ensured that the evaluation basis can directly reflect the severity of the fault.

[0054] Next, the evaluation module calls a pre-built mapping model between feature parameters and fault severity. This model is trained based on fault sample data from multiple scenarios. The sample data covers fault data of different models of forging hydraulic presses under different forging force conditions and different forging materials, and includes the full-cycle feature parameter change curves from the initial stage to the severe stage, adaptable to evaluation needs under diverse working conditions. The mapping model presets three levels of feature parameter thresholds to divide the parameter ranges into the initial, moderate, and severe stages. In the initial stage, the feature parameter values ​​are between the upper limit of the normal range and the slightly abnormal threshold, indicating that the fault has just begun and has not yet affected normal operation. In the moderate stage, the values ​​are between the slightly abnormal threshold and the severely abnormal threshold, indicating that the fault has had some impact but can still operate for a short period of time. In the severe stage, the values ​​exceed the severely abnormal threshold, indicating that the fault has seriously threatened equipment safety. The threshold division achieves the initial quantification of fault severity.

[0055] The evaluation module then substitutes the extracted feature parameter values ​​into the mapping model and matches the initial fault severity level based on the threshold range of the parameter values. For example, in a hydraulic pump wear fault, if the vibration signal kurtosis value is between the mild and severe abnormal thresholds, the initial level is matched as moderate wear. To improve the accuracy of the evaluation, the module further retrieves historical operating data of the faulty component and dynamically corrects the initial level by combining the component's operating time with historical fault records. If the operating time of the faulty component has exceeded the rated service life specified in the parameters of the corresponding equipment model, long-term operation will reduce the component's fault resistance, and even if the initial level is mild, it needs to be adjusted to moderate. If the component has historical fault records consistent with the current fault type, repeated faults may indicate that the component's performance has continued to deteriorate, and even if the initial level is moderate, it needs to be adjusted to severe. Dynamic correction eliminates the limitations of single-parameter evaluation, making the level determination more consistent with the actual state of the component.

[0056] Finally, the evaluation module generates corresponding evaluation indicators and maintenance guidelines based on the revised fault severity level. The evaluation indicators are divided into three levels: minor, moderate, and severe faults. Maintenance guidelines for each level are designed based on equipment operating requirements. Minor faults correspond to low maintenance priority, prompting regular monitoring of characteristic parameter changes. The monitoring cycle is determined based on the current forging task frequency to avoid over-maintenance impacting production. Moderate faults correspond to medium maintenance priority, recommending maintenance during the next production break. A list of related components to be checked is also output to prevent the fault from spreading to other components. Severe faults correspond to high maintenance priority, simultaneously outputting fault risk warnings and recommended maintenance cycles. The fault risk warning clearly states the potential consequences of continued operation of the fault, such as hydraulic system pressure imbalance affecting overall operation or forging machining accuracy exceeding tolerances leading to product scrap. The recommended maintenance cycle is determined based on the current forging workload and the estimated remaining lifespan of the components. The estimated remaining lifespan of the components references the historical replacement cycle of the same model and the wear rate under current operating conditions, ensuring that maintenance is timely enough to repair the fault while minimizing disruption to the production schedule. This results in an evaluation result that is both accurate and practical, providing clear guidance for subsequent maintenance operations.

[0057] As an optional embodiment, in step 106, the fault diagnosis results are pushed to the interactive interface in real time to guide the user to perform equipment maintenance operations or confirm the trigger command of the automatic maintenance process of the equipment. This includes: splitting the fault diagnosis results into basic diagnostic information and maintenance guidance information according to information type. The basic diagnostic information includes fault status judgment results, fault component location results, fault type identification results, and fault severity assessment level. The maintenance guidance information includes maintenance priority recommendations, maintenance cycle suggestions, related component checklists, and fault risk warnings corresponding to the severity level. The basic diagnostic information and maintenance guidance information are presented in different functional areas of the interactive interface. The basic diagnostic information area uses visual icons to distinguish fault components and fault types, and the maintenance guidance information area displays corresponding operation suggestions in order of maintenance priority. Different operation entry points are provided through the interactive interface for different fault severity assessment levels. For example, a minor fault corresponds to the monitoring plan setting entry, through which the monitoring frequency and warning threshold of characteristic parameters can be adjusted. The monitoring frequency is matched with the current forging task frequency by default. A moderate fault corresponds to the maintenance task creation entry, through which a maintenance work order containing steps for replacing faulty components and related component inspection items is generated. The maintenance work order automatically associates the component specification information in the equipment model parameters. A severe fault corresponds to the simultaneous display of an emergency maintenance confirmation entry and an automatic maintenance process trigger entry. The emergency maintenance confirmation entry is used to manually confirm maintenance instructions and assign maintenance personnel, while the automatic maintenance process trigger entry is used by the user to authorize the system to start the preset equipment shutdown protection and component emergency isolation procedures.

[0058] Specifically, the fault diagnosis results are first broken down by information function type, into basic diagnostic information and maintenance guidance information. For example, basic diagnostic information focuses on the core fault status, including fault status judgment, fault component location, fault type identification, and fault severity assessment level. For instance, it might determine that the equipment has a fault, locate the faulty component as the hydraulic pump, identify the fault type as wear, and assess the level as moderate. Maintenance guidance information provides support for subsequent operations, including maintenance priority recommendations corresponding to the severity level, maintenance cycle suggestions, a checklist of related components, and fault risk warnings. For example, a moderate wear fault corresponds to a medium maintenance priority recommendation, suggesting checking related components such as the overflow valve during the next production interval, and warning that continued operation may lead to abnormal flow. This categorization clearly separates information for different purposes, allowing users to quickly obtain their core needs.

[0059] The two categories of information are then presented in different functional areas of the interface. The interface design follows the principle of prioritizing core information. The basic diagnostic information area is located in the upper half of the interface, using visual icons to intuitively distinguish faulty components and fault types. For example, a red pump-shaped icon marks a faulty hydraulic press, while a sawtooth graphic marks wear faults. Users can quickly identify key fault information without reading text. The maintenance guidance information area is located in the lower half of the interface, displaying operation suggestions sorted from high to low maintenance priority. For example, maintenance suggestions for moderate faults will prioritize the recommended maintenance cycle, followed by a checklist of related components, avoiding cluttered information that could delay user decision-making. This partitioned presentation method not only conforms to user information acquisition habits but also improves information browsing efficiency.

[0060] Then, differentiated operation entry points are designed based on the severity of different faults through an interactive interface, ensuring that the operation entry points accurately match the fault handling requirements. For minor faults, the interface displays a monitoring plan setting entry point. The design principle of this entry point is that minor faults do not require immediate maintenance shutdown. Instead, the monitoring strategy should be dynamically adjusted to prevent the fault from worsening. Through this entry point, users can modify the monitoring frequency and warning threshold of characteristic parameters. The system defaults to associating the monitoring frequency with the current forging task frequency. For example, when forging tasks are frequent, the monitoring frequency of vibration signal kurtosis value is automatically increased to avoid missing signs of fault development due to excessively long monitoring intervals. For moderate faults, the interface displays a maintenance task creation entry point. Moderate faults need to be handled during production breaks. This entry point allows users to generate standardized maintenance work orders. The work order automatically integrates the replacement steps of the faulty component and the associated component inspection items, and automatically matches the component specification information according to the equipment model parameters. For example, for a fault in the overflow valve seal of a certain model of hydraulic press, the work order will automatically fill in the seal specifications and disassembly and installation steps suitable for that model, reducing the amount of manual data entry for users. For severe malfunctions, the interface simultaneously displays an emergency maintenance confirmation entry and an automatic maintenance process trigger entry. Severe malfunctions require rapid handling to prevent equipment damage. The emergency maintenance confirmation entry allows users to manually confirm maintenance commands and assign maintenance personnel, clearly defining responsibilities. The automatic maintenance process trigger entry allows users to authorize the system to initiate preset programs, quickly executing equipment shutdown protection to prevent the fault from escalating. Simultaneously, it initiates emergency isolation of components to prevent the fault from affecting other parts. For example, after a hydraulic cylinder jamming fault is triggered, the system will automatically cut off the hydraulic pump's oil supply to the hydraulic cylinder, isolating the faulty component.

[0061] The above process, through information classification and differentiated operation entry design, ensures that users can quickly grasp the core information of the fault and choose the most efficient handling method according to the severity of the fault. Minor faults are prevented and maintained by optimizing monitoring strategies, moderate faults are maintained more efficiently by standardizing work orders, and severe faults are responded to quickly by a combination of manual and automatic entry points. Ultimately, this achieves seamless connection between fault diagnosis results and maintenance operations, reduces user decision-making time, reduces production losses caused by equipment downtime due to faults, and ensures the standardization and safety of maintenance operations.

[0062] As an optional embodiment, after step 106, when the user confirms the maintenance operation command or authorizes the automatic maintenance process through the interactive interface, the command execution status is fed back in real time. The feedback content includes the command issuance progress, maintenance resource allocation, and the current operating status of the equipment. If an abnormality occurs after the automatic maintenance process is triggered, the interactive interface will immediately pop up an abnormality warning message and provide a manual intervention entry. After the maintenance operation is completed, the interactive interface receives the maintenance result data returned by the system, automatically associates and stores the fault diagnosis result, maintenance operation record, and maintenance effect verification data, forming a full lifecycle file for the faulty component. The file data is used for subsequent online optimization of the adaptive deep learning model, providing maintenance effect feedback data under actual working conditions for updating the fuzzy rule base and attention weights of the model.

[0063] Specifically, after step 106, the system first activates the instruction execution tracking system. This system relies on the real-time data acquisition module and the status analysis algorithm to continuously capture key information at each stage of instruction execution and convert this information into content that users can intuitively understand and feed back to the interactive interface. The feedback focuses on the core dimensions required for user decision-making and operation. The instruction issuance progress feedback fully presents the entire process from instruction generation and transmission to receipt by the execution terminal. For example, it informs the user that the maintenance instruction has been successfully issued to the terminal equipment responsible for hydraulic pump maintenance, or that the automatic shutdown instruction has been received by the control system and entered the preprocessing stage, allowing the user to clearly grasp the real-time dynamics of the instruction flow. The maintenance resource allocation feedback links various resources required for maintenance activities. For example, it shows that the personnel responsible for this relief valve maintenance are in place, that the seals compatible with this model of relief valve have been dispatched from the spare parts warehouse, and that the dedicated pressure testing tools have been allocated to the maintenance station, helping the user to determine in advance whether the pre-maintenance preparations are ready. The equipment current operating status feedback is achieved through real-time collected sensor data. For example, it displays whether the current pressure, flow rate, and temperature of key components of the hydraulic press have dropped to a safe operating range, or whether the equipment has successfully entered the shutdown protection mode, ensuring that the user is aware of changes in equipment status in a timely manner and providing a safe basis for subsequent maintenance operations.

[0064] If an abnormal situation occurs after the automatic maintenance process is triggered, such as the hydraulic system pressure failing to drop as expected after the equipment receives a shutdown command, or the emergency isolation valve of the faulty component failing to close properly, the system's anomaly monitoring module will immediately detect these deviations from the normal process. A pre-set early warning mechanism will instantly display an anomaly alert on the interactive interface, while also providing a dedicated manual intervention entry. Users can directly send emergency control commands to the control system through this entry, such as forcibly cutting off the power supply to the hydraulic system or activating the backup mechanical isolation device, quickly curbing the development of the abnormal situation and preventing damage to core equipment components or safety accidents caused by the continued escalation of the anomaly.

[0065] Once all maintenance operations are completed, the interactive interface will automatically receive maintenance result data transmitted back from the system backend. This data includes key records and effectiveness verification information from the maintenance process. Maintenance process records include the disassembly time of the faulty component, details of the replacement steps, and operational notes from the maintenance personnel. Maintenance effectiveness verification data includes measured values ​​of characteristic parameters related to the faulty component after maintenance, such as the vibration signal parameters of the hydraulic pump after replacing a worn bearing and the pressure fluctuation data of the relief valve after repairing a seal failure. The interactive interface will then integrate the complete results obtained from this fault diagnosis, including the fault type and severity assessment level, with the maintenance operation records and maintenance effectiveness verification data. These results are categorized and stored according to the unique identifier of the faulty component, forming a full lifecycle file for that component. Taking the hydraulic cylinder of a forging hydraulic press as an example, its file will fully record the diagnosis results of this jamming fault, the type of seal replaced during maintenance, and the improvement in the flow signal after maintenance. It will also link all past fault records and maintenance history of the hydraulic cylinder, constructing a complete data chain from component installation and use to each fault maintenance.

[0066] These archival data are not merely stored as historical records, but also serve as crucial input for the online optimization of subsequent adaptive deep learning models. The model periodically reads the archival data, analyzes the matching degree between maintenance effect verification data and the initial fault diagnosis results, and determines whether the original diagnostic logic needs adjustment. For example, if archival data shows that a certain type of hydraulic pump wear fault was diagnosed as minor wear according to the original fuzzy rule library, but actual maintenance verification data indicates that the fault has reached a moderate level, it indicates a deviation in the correlation logic between vibration signal features and wear levels in the original fuzzy rules. The model will then update the membership function parameters of the corresponding features in the fuzzy rule library based on this actual feedback data, and simultaneously adjust the attention weight of the Transformer layer for this type of vibration feature, enabling subsequent diagnoses of similar faults to more accurately match the actual fault state. Through these steps, the diagnostic accuracy and scenario adaptability of the model under complex actual working conditions can be continuously improved, ensuring that the fault diagnosis method can stably meet the maintenance needs of forging hydraulic presses in the long term.

[0067] In this embodiment, by combining an adaptive deep learning model with a fuzzy neural network, fault diagnosis of the hydraulic press can be achieved, thereby improving the accuracy and efficiency of fault diagnosis and ensuring the stability, efficiency and reliability of equipment production.

[0068] In another embodiment of this application, a fault diagnosis system for key components of a forging hydraulic press is also provided, see [link to relevant documentation]. Figure 2 The system is applied to the aforementioned method, and the system includes at least the following: The acquisition unit is configured to collect multi-source heterogeneous time-series data reflecting the operating status of components in real time by arranging multiple types of status monitoring sensors on the key functional components of the forging hydraulic press. The key functional components are the core components in the forging hydraulic press system structure whose operating status directly affects the overall stability of the forging hydraulic press, the stability of the hydraulic system, and the processing accuracy. The key functional components include at least one of the following: hydraulic pump, relief valve, and hydraulic cylinder. The preprocessing unit is configured to preprocess multi-source heterogeneous time-series data. The preprocessing includes data cleaning, noise reduction, and adaptive normalization based on operating conditions. The extraction unit is configured to extract high-identity fault features from different sensors, components, and dimensions from preprocessed multi-source data; integrate the multi-source high-identity fault features, analyze and calculate the feature parameters reflecting the coupling relationship between key components, and form a fusion feature vector characterizing single-component faults and multi-component coupled faults. The diagnostic unit is configured to construct an adaptive deep learning model that incorporates a fuzzy neural network; input the fused feature vector into the adaptive deep learning model to obtain fault diagnosis results, which include at least fault status judgment, fault component location, fault type identification, and fault severity assessment; and push the fault diagnosis results to the interactive interface in real time to guide the user to perform equipment maintenance operations or confirm the trigger command of the equipment automatic maintenance process.

[0069] The fault diagnosis system for key components of forging hydraulic presses can implement the steps in the aforementioned method embodiments, which will not be elaborated here.

[0070] In another embodiment of this application, an electronic device is also provided, with reference to Figure 3 As shown, the device includes a processor 1110, a communication interface 1120, a memory 1130, and a communication bus 1140. The processor 1110, communication interface 1120, and memory 1130 communicate with each other via the communication bus. The memory 1130 stores computer programs. When the processor 1110 executes the program stored in the memory 1130, it implements the fault diagnosis method for key components of the forging hydraulic press described in the method embodiment. The communication bus 1140 mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc.

[0071] For ease of representation, Figure 3The bus is represented by a single thick line, but this does not imply that there is only one bus or one type of bus. Communication interface 1120 is used for communication between the aforementioned electronic device and other devices. Memory 1130 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor. The aforementioned processor 1110 may be a general-purpose processor, including artificial intelligence processors, graphics processing units (GPUs), machine learning units (MLUs), central processing units (CPUs), network processors (NPs), etc.; it may also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0072] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed, can implement the steps that can be performed by an electronic device in the above method embodiments.

Claims

1. A method for diagnosing faults in key components of a forging hydraulic press, characterized in that, The method includes: By arranging multiple types of condition monitoring sensors on the key functional components of the forging hydraulic press, multi-source heterogeneous time-series data reflecting the operating status of the components are collected in real time; the key functional components are the core components in the forging hydraulic press system structure whose operating status directly affects the overall stability of the forging hydraulic press, the stability of the hydraulic system, and the processing accuracy; the key functional components include at least one of the following: hydraulic pump, relief valve, and hydraulic cylinder; Preprocessing of multi-source heterogeneous time-series data includes data cleaning, noise reduction, and adaptive normalization based on operating conditions. High-identity fault features from different sensors, components, and dimensions are extracted from preprocessed multi-source data. These multi-source high-identity fault features are integrated, and feature parameters reflecting the coupling relationship between key components are analyzed and calculated to form a fusion feature vector characterizing single-component faults and multi-component coupled faults. Construct an adaptive deep learning model that combines a fuzzy neural network; input the fused feature vector into the adaptive deep learning model to obtain fault diagnosis results, the fault diagnosis results including at least fault status judgment, fault component location, fault type identification and fault severity assessment, and push the fault diagnosis results to the interactive interface in real time to guide the user to perform equipment maintenance operations or confirm the trigger command of the automatic maintenance process of the equipment. The step of inputting the fused feature vector into the adaptive deep learning model to obtain the fault diagnosis result includes: The fused feature vector is split into a component-independent feature subset and a component-coupled feature subset. The component-independent feature subset contains the single-component features corresponding to each key component; the component-coupled feature subset contains cross-component related features. The component-independent feature subset is input into the ANFIS layer of the adaptive deep learning model. Based on the fuzzy rule base generated by pre-training, the ANFIS layer calculates the fuzzy membership degree of the component-independent feature subset and outputs the preliminary confidence of each component for any of the faults among wear, sealing failure, and jamming. The component-coupled feature subset is input into the Transformer layer of the model. The Transformer layer performs global correlation analysis on the component-coupled feature subset through a multi-head attention mechanism, selects the coupling features with the highest correlation with the fault, and outputs the range of associated components for multi-component coupled faults and the corresponding fault type confidence. The single-component fault confidence output from the ANFIS layer and the coupled fault confidence output from the Transformer layer are weighted and fused. The weight coefficient of the coupled fault confidence is dynamically adjusted according to the current forging force condition and forging material. The greater the forging force and the higher the hardness of the forging material, the higher the weight coefficient of the coupled fault confidence. The fault status is determined based on the weighted fusion confidence results. When the confidence of any fault type is greater than the preset confidence threshold, a fault is determined to exist. The fault type with the highest confidence is extracted as the final fault type. The fault component location result is determined by combining the range of associated components output by the Transformer layer. The fault severity assessment module built into the model is called to generate corresponding assessment indicators. The results are integrated to form a fault diagnosis result that includes fault status judgment, fault component location, fault type identification, and fault severity assessment.

2. The fault diagnosis method for key components of a forging hydraulic press according to claim 1, characterized in that, The preprocessing of multi-source heterogeneous time-series data includes: A wavelet transform-based denoising algorithm and an adaptive filter are used to jointly process multi-source heterogeneous time series data to obtain preprocessed data, so as to suppress noise introduced by mechanical vibration and electromagnetic interference at the forging site and retain the integrity of fault characteristics. The normalization parameters of the preprocessed data are dynamically adjusted according to the actual operating conditions, and the sensor data in the preprocessed data is mapped to a standard range that matches the current operating conditions to obtain preprocessed multi-source data.

3. The fault diagnosis method for key components of a forging hydraulic press according to claim 1, characterized in that, The extraction of highly identifiable fault features from preprocessed multi-source data under different sensors, components, and dimensions includes: From the vibration signal, pressure signal, flow signal, and temperature signal contained in the preprocessed multi-source data, multi-dimensional basic features corresponding to a single signal source are extracted respectively. Based on the operational correlation of hydraulic pumps, relief valves, and hydraulic cylinders under real-time working conditions, the interaction features between key components are extracted from preprocessed multi-source data. Among them, the linkage change features between hydraulic pump output flow and relief valve inlet pressure, the matching features between relief valve pressure regulation value and hydraulic cylinder movement speed, and the correspondence features between hydraulic cylinder inlet and outlet pressure difference and hydraulic pump working load are extracted. The highly identifiable fault features are obtained by filtering the multi-dimensional basic features of a single signal source and the interaction features between key components.

4. The fault diagnosis method for key components of a forging hydraulic press according to claim 3, characterized in that, The process involves extracting multi-dimensional basic features corresponding to a single signal source from the vibration, pressure, flow, and temperature signals contained in the preprocessed multi-source data, including: For vibration signals, time-domain statistical features, frequency-domain spectral features, and time-frequency-domain energy features are extracted. The time-domain statistical features include peak value, root mean square value, kurtosis, and skewness. The frequency-domain spectral features include peak value, center frequency, and frequency band energy percentage. The time-frequency-domain energy features include the energy values ​​of each frequency band after wavelet packet decomposition. The pressure fluctuation amplitude, pressure stabilization duration, and pressure response delay time are extracted from the pressure signal. The flow pulsation frequency, flow deviation value, and instantaneous flow peak value are extracted from the flow signal. The temperature change rate, temperature stability range, and duration of overtemperature are extracted from the temperature signal.

5. The fault diagnosis method for key components of a forging hydraulic press according to claim 1, characterized in that, The construction of the adaptive deep learning model combining fuzzy neural networks includes: An adaptive deep learning model is pre-trained using multi-scenario fault sample data and constructed using an adaptive neural fuzzy inference system. The multi-scenario fault sample data covers fault data under different models of forging hydraulic presses, different forging materials, and different forging force conditions. The sample data includes single-component fault samples and multi-component coupled fault samples. The adaptive neural fuzzy inference system layer is based on the failure mechanism of key components of forging hydraulic press. It automatically generates a fuzzy rule base for fault diagnosis through data-driven methods. The fuzzy rule base contains the association logic between different fault types and multi-dimensional features in the multi-scenario fault sample data. The contribution weight of each fault feature to the fault category is calculated through the fuzzy membership function to obtain the fuzzy representation of the fault features as a fuzzy feature matrix. The adaptive deep learning model adopts the Transformer network architecture. The input is a fuzzy feature matrix. The fuzzy feature data corresponding to different sensors and key components are associated through a multi-head attention mechanism. A global feature association matrix is ​​constructed based on the fuzzy feature data. Then, the deep fault information in the global feature association matrix is ​​extracted through the encoder layer to complete the preliminary diagnosis of the fault status. The adaptive deep learning model is fine-tuned by using a preset number of target application scenario samples; the model parameters are optimized by employing transfer learning and online optimization mechanisms, combined with the equipment model of the forging hydraulic press and real-time collected working condition data.

6. The fault diagnosis method for key components of a forging hydraulic press according to claim 1, characterized in that, The built-in fault severity assessment module of the call model generates corresponding assessment indicators, including: The fault severity assessment module extracts feature parameters that are strongly correlated with the final fault type from the fused feature vector. If the final fault type is wear, it extracts the kurtosis value and root mean square value from the time-domain statistical features of the vibration signal. If it is a sealing failure, it extracts the pressure fluctuation amplitude and pressure response delay time from the pressure signal. If it is jamming, it extracts the flow pulsation frequency and flow deviation value from the flow signal. Based on multi-scenario fault sample data, a mapping model between feature parameters and fault severity is constructed. The mapping model sets three levels of feature parameter thresholds. In the initial stage, the feature parameter values ​​are between the upper limit of the normal range and the threshold for mild abnormality. In the moderate stage, they are between the threshold for mild abnormality and the threshold for severe abnormality. In the severe stage, they exceed the threshold for severe abnormality. Substitute the currently extracted feature parameter values ​​into the mapping model to obtain the initial level of fault severity; retrieve the historical operating data of the faulty component, and dynamically correct the initial level by combining the component's operating time with historical fault records to obtain the corrected fault severity level; if the operating time of the faulty component exceeds the rated service life in the parameters of the corresponding equipment model and the initial level is mild, then the initial level is adjusted to moderate; if the faulty component has historical fault records that are consistent with the current fault type and the initial level is moderate, then the initial level is adjusted to severe. An assessment index is generated based on the revised severity level of the fault. The assessment index is divided into the following levels: minor fault, moderate fault, and severe fault. Each level corresponds to different maintenance guidance information.

7. The fault diagnosis method for key components of a forging hydraulic press according to claim 1, characterized in that, The step of pushing the fault diagnosis results to the interactive interface in real time to guide the user to perform equipment maintenance operations or confirm the trigger command for the automatic maintenance process of the equipment includes: The fault diagnosis results are divided into basic diagnostic information and maintenance guidance information according to information type. The basic diagnostic information includes fault status judgment results, fault component location results, fault type identification results, and fault severity assessment level. The maintenance guidance information includes maintenance priority recommendations, maintenance cycle suggestions, related component checklists, and fault risk warnings corresponding to the severity level. Basic diagnostic information and maintenance guidance information are presented in different functional areas of the interactive interface. The basic diagnostic information area uses visual icons to distinguish faulty components and fault types, while the maintenance guidance information area displays corresponding operation suggestions in order of maintenance priority. The interactive interface provides differentiated operation entry points for different fault severity assessment levels.

8. The method for diagnosing faults in key components of a forging hydraulic press according to claim 1, characterized in that, After pushing the fault diagnosis results to the interactive interface in real time to guide the user to perform equipment maintenance operations or confirm the trigger command for the automatic maintenance process of the equipment, the method further includes: After the user confirms the maintenance operation command or authorizes the automatic maintenance process through the interactive interface, the command execution status is fed back in real time. The feedback content includes the command issuance progress, maintenance resource allocation, and the current operating status of the equipment. If an abnormality occurs after the automatic maintenance process is triggered, the interactive interface will immediately pop up an abnormality warning message and provide a manual intervention entry. After the maintenance operation is completed, the interactive interface receives the maintenance result data returned by the system, automatically associates and stores the fault diagnosis result, maintenance operation record, and maintenance effect verification data, forming a full life cycle archive of the faulty component. The archive data is used for subsequent online optimization of the adaptive deep learning model, providing feedback data on the maintenance effect under actual working conditions for updating the fuzzy rule base and attention weights of the model.

9. A fault diagnosis system for key components of a forging hydraulic press, characterized in that, The system, applied to the fault diagnosis method for key components of a forging hydraulic press according to any one of claims 1 to 8, comprises: The acquisition unit is configured to collect multi-source heterogeneous time-series data reflecting the operating status of key functional components of the forging hydraulic press in real time by arranging multiple types of status monitoring sensors on the key functional components; the key functional components are the core components in the forging hydraulic press system structure whose operating status directly affects the overall stability of the forging hydraulic press, the stability of the hydraulic system, and the processing accuracy; the key functional components include at least one of the following: hydraulic pump, relief valve, and hydraulic cylinder. The preprocessing unit is configured to preprocess multi-source heterogeneous time-series data, and the preprocessing includes data cleaning, noise reduction, and adaptive normalization based on operating conditions. The extraction unit is configured to extract high-identity fault features from different sensors, components, and dimensions from preprocessed multi-source data; integrate the multi-source high-identity fault features, analyze and calculate the feature parameters reflecting the coupling relationship between key components, and form a fusion feature vector characterizing single-component faults and multi-component coupled faults. The diagnostic unit is configured to construct an adaptive deep learning model that incorporates a fuzzy neural network; input the fused feature vector into the adaptive deep learning model to obtain fault diagnosis results, the fault diagnosis results including at least fault state judgment, fault component location, fault type identification and fault severity assessment, and push the fault diagnosis results to the interactive interface in real time to guide the user to perform equipment maintenance operations or confirm the trigger command of the automatic maintenance process of the equipment.

Images