Forging press state monitoring system based on intelligent sensor

Abstract

The invention discloses a forging press state monitoring system based on an intelligent sensor, relates to the field of forging press state monitoring, and aims at monitoring the state of a forging press by synchronously acquiring a vibration signal representing the mechanical state of equipment and an electromagnetic interference signal reflecting the noise characteristics of a field environment and introducing a cross-modal deep learning model. And deeply mining and establishing a nonlinear mapping relation between the two heterogeneous signals, guiding a model to learn and identify a propagation normal form and pollution characteristics of noise in a vibration signal, and further reconstructing a high-fidelity pure fault signal. Based on this, deep time-frequency domain feature enhancement and intelligent analysis are performed on the purified signals, so that accurate diagnosis of early weak faults of the forging press is realized. Therefore, the safety and the stability of equipment operation can be improved, the service life and the operation cycle of the forging press are effectively prolonged, so that the non-planned shutdown loss and the maintenance cost are reduced, and the intelligent level of health management of the forging and pressing equipment is comprehensively improved.

Description

Technical Field

[0001] This application relates to the field of forging press condition monitoring, and more specifically, to a forging press condition monitoring system based on intelligent sensors. Background Technology

[0002] With the continuous upgrading and development of high-end manufacturing, the stability and reliability of forging and pressing machinery, as the cornerstone of modern industry, directly affect the efficiency and safety of the entire production chain. During operation, forging and pressing equipment is subjected to severe impact loads and high-frequency vibrations for extended periods, making its key components highly susceptible to early failures such as fatigue damage and wear. If these minor issues are not detected and addressed in time, they can escalate into major downtime accidents, causing not only huge economic losses but also potentially serious safety problems. Therefore, to move beyond the traditional reliance on manual experience for periodic maintenance, developing an intelligent monitoring system capable of real-time and accurate perception of equipment health status and early warning of potential faults has become crucial for ensuring the long-term safe and stable operation of forging and pressing equipment and advancing towards intelligent production. In recent years, the rapid advancement of intelligent sensing technology has provided strong support for achieving this goal. By deploying sensors at key locations on the equipment, rich dynamic signals can be continuously collected, providing a solid foundation for data-driven equipment health status assessment and predictive maintenance strategies. Its research and application have significant practical and engineering value.

[0003] However, existing technologies still face significant technical bottlenecks in signal acquisition and processing. Forging workshops typically deploy various electrical equipment such as high-power motors, frequency converters, and welding equipment. These devices generate strong electromagnetic interference (EMI) during operation, creating a complex electromagnetic noise environment. This broadband noise can easily couple into the vibration signals acquired by sensors through spatial radiation or circuit conduction, causing severe signal contamination. Simultaneously, the characteristic signals generated when equipment experiences early-stage faults are often very weak, with energy potentially far below the ambient noise level. Existing signal denoising algorithms often face a dilemma when processing such weak signals submerged in strong noise: strong filtering may filter out effective fault characteristic information as noise, leading to signal distortion; insufficient filtering strength may result in a large amount of residual noise interfering with or even completely obscuring fault characteristics, significantly reducing the accuracy of subsequent fault diagnosis and identification.

[0004] Therefore, there is an urgent need for an optimized forging press condition monitoring system based on intelligent sensors. Summary of the Invention

[0005] This application is made in order to solve the above-mentioned technical problems.

[0006] According to one aspect of this application, a forging press condition monitoring system based on intelligent sensors is provided, comprising: The raw signal acquisition module is used to acquire raw vibration signals and raw EMI signals; The preprocessing and segmentation module is used to preprocess and segment the original vibration signal and the original EMI signal to obtain the noisy vibration tensor and the electromagnetic interference reference tensor. The cross-modal denoising module is used to input the noisy vibration tensor and the electromagnetic interference reference tensor into the cross-modal denoising network to obtain the denoised vibration tensor. The feature enhancement and extraction module is used to enhance and extract the time-frequency domain features of the denoised vibration tensor to obtain the fault feature spectrum; The fault diagnosis module is used to perform fault diagnosis based on the fault feature spectrum to obtain the fault type and confidence score.

[0007] Compared with existing technologies, this application provides a forging press condition monitoring system based on intelligent sensors. This system simultaneously collects vibration signals characterizing the mechanical state of the equipment and electromagnetic interference signals reflecting the noise characteristics of the surrounding environment. It introduces a cross-modal deep learning model to deeply mine and establish the nonlinear mapping relationship between the two heterogeneous signals. This guides the model to learn and identify the propagation paradigm and contamination characteristics of noise in the vibration signal, thereby reconstructing a high-fidelity clean fault signal. Based on this, the purified signal undergoes in-depth time-frequency domain feature enhancement and intelligent analysis to achieve accurate diagnosis of early, subtle faults in the forging press. This improves the safety and stability of equipment operation, effectively extends the service life and operating cycle of the forging press, reduces unplanned downtime losses and maintenance costs, and comprehensively enhances the intelligent level of health management for forging equipment. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a block diagram of a forging press condition monitoring system based on smart sensors according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the data flow of a forging press condition monitoring system based on smart sensors according to an embodiment of this application.

[0011] Figure 3This is a block diagram of the preprocessing and segmentation module in the intelligent sensor-based forging press condition monitoring system according to an embodiment of this application.

[0012] Figure 4 This is a block diagram of a cross-modal noise reduction module in a smart sensor-based forging press condition monitoring system according to an embodiment of this application.

[0013] Figure 5 This is a block diagram of the feature enhancement and extraction module in the intelligent sensor-based forging press condition monitoring system according to an embodiment of this application. Detailed Implementation

[0014] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0015] To address the problems mentioned above in the background technology, this application proposes a forging press condition monitoring system based on intelligent sensors. Figure 1 This is a block diagram of a forging press condition monitoring system based on smart sensors according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow of a forging press condition monitoring system based on smart sensors according to an embodiment of this application. Figure 1 and Figure 2 As shown, the forging press condition monitoring system 100 based on intelligent sensors includes: a raw signal acquisition module 110 for acquiring raw vibration signals and raw EMI signals; a preprocessing and segmentation module 120 for preprocessing and segmenting the raw vibration signals and raw EMI signals to obtain a noisy vibration tensor and an electromagnetic interference reference tensor; a cross-modal denoising module 130 for inputting the noisy vibration tensor and electromagnetic interference reference tensor into a cross-modal denoising network to obtain a denoised vibration tensor; a feature enhancement and extraction module 140 for performing time-frequency domain feature enhancement and extraction on the denoised vibration tensor to obtain a fault feature spectrum; and a fault diagnosis module 150 for performing fault diagnosis based on the fault feature spectrum to obtain the fault type and confidence score.

[0016] In the aforementioned intelligent sensor-based forging press condition monitoring system, the raw signal acquisition module 110 is used to acquire raw vibration signals and raw EMI signals. It should be understood that since faults in key components of the forging press during operation are characterized by vibration signals, and electromagnetic interference generated by high-power electrical equipment in the workshop contaminates the vibration signals, relying solely on a single vibration signal is insufficient to distinguish fault characteristics from noise. Therefore, this application uses dual sensors to simultaneously acquire mechanical vibration signals from the equipment and environmental electromagnetic interference signals, thereby comprehensively capturing equipment status information and noise interference characteristics. This provides raw data support for subsequently establishing a mapping relationship between noise and vibration signals and achieving accurate noise reduction, ensuring that fault characteristics are not masked by noise, and laying a data foundation for early fault diagnosis.

[0017] Specifically, in one possible embodiment, the implementation process of the original signal acquisition module 110 is as follows: First, a high-sensitivity piezoelectric accelerometer is used as the main sensor and installed in key monitoring parts such as the flywheel bearing housing and main cylinder of the forging press to collect mechanical vibration signals. Simultaneously, a ring-shaped electromagnetic probe is used as an auxiliary sensor and deployed near electromagnetic interference sources such as the inverter main cable and motor control cabinet to collect electromagnetic interference signals. Through a multi-channel data acquisition device with hardware synchronization capabilities, the signal acquisition triggering mechanisms of the two sensors are bound together to ensure strict alignment of the two signals in the time dimension. During the acquisition process, the same sampling trigger frequency and data transmission rate are maintained, achieving synchronous acquisition of the original vibration signal and the original EMI signal.

[0018] In the aforementioned intelligent sensor-based forging press condition monitoring system, the preprocessing and segmentation module 120 is used to preprocess and segment the original vibration signal and original EMI signal to obtain the noisy vibration tensor and electromagnetic interference reference tensor. It should be understood that the original acquired vibration signal and EMI signal contain environmental clutter, signal attenuation, and other interference, and there are issues with dimensional differences and the inability to directly input continuous data streams into deep learning models. Therefore, this application further performs a series of processing steps on the original vibration signal and original EMI signal, including conditioning, segmentation, standardization, and tensor quantization, to optimize signal quality, eliminate data differences, and adapt to the model input format. This improves the usability and consistency of the signal, ensuring that the processed data retains fault characteristics and noise correlation information while meeting the input requirements of cross-modal denoising networks, providing a qualified data carrier for subsequent accurate establishment of signal mapping relationships and efficient completion of denoising tasks.

[0019] In particular, in one specific embodiment, Figure 3 This is a block diagram of the preprocessing and segmentation module in a smart sensor-based forging press condition monitoring system according to an embodiment of this application. Figure 3As shown, the preprocessing and segmentation module 120 includes: a signal conditioning and digitization unit 121, used to perform analog signal conditioning and synchronous digitization on the original vibration signal and the original EMI signal to obtain a digital vibration signal stream and a digital EMI signal stream; a data stream synchronous segmentation unit 122, used to perform data stream synchronous segmentation on the digital vibration signal stream and the digital EMI signal stream to obtain a set of vibration segments and a set of EMI segments; and a normalization and tensor quantization unit 123, used to normalize and tensor quantize the set of vibration segments and the set of EMI segments to obtain a noisy vibration tensor and an electromagnetic interference reference tensor.

[0020] Specifically, the signal conditioning and digitization unit 121 is used to perform analog signal conditioning and synchronous digitization on the original vibration signal and the original EMI signal to obtain digital vibration signal stream and digital EMI signal stream. It should be understood that since the original vibration signal and EMI signal exist in analog form, they are susceptible to signal distortion due to transmission line loss and environmental interference, and subsequent data processing and feature analysis need to be based on calculable digital signals. Therefore, this application further conditions and optimizes the two analog signals and performs synchronous digitization conversion to correct signal distortion, filter out invalid noise, and convert them into a calculable digital form. This improves the stability and integrity of the signal, ensuring that the digitized signal accurately reflects the original physical quantity characteristics. Simultaneously, synchronous processing ensures the time alignment of the two signals, providing a high-quality digital signal foundation for subsequent segmentation and feature correlation analysis, and avoiding the impact of signal distortion or asynchrony on subsequent processing results.

[0021] Specifically, in one possible embodiment, the signal conditioning and digitization unit 121 is implemented as follows: First, the original vibration signal is conditioned by converting the weak charge signal output by the piezoelectric sensor into a voltage signal using a charge amplifier, and then filtering out high-frequency noise higher than half the sampling frequency using a low-pass anti-aliasing filter. Simultaneously, a signal amplifier is used to boost the amplitude of the original EMI signal, and a dedicated electromagnetic interference filter is used to suppress interference in irrelevant frequency bands. Then, the two conditioned analog signals are connected to a multi-channel analog-to-digital converter, a unified sampling frequency and quantization precision are set, and the conversion process is controlled by a hardware synchronous triggering mechanism to ensure that each data point of the two signals corresponds one-to-one in time, ultimately outputting a synchronized digital vibration signal stream and a digital EMI signal stream.

[0022] Specifically, the data stream synchronization segmentation unit 122 is used to synchronously segment the digital vibration signal stream and the digital EMI signal stream to obtain a set of vibration segments and a set of EMI segments. It should be understood that due to the massive amount of continuous digital signal stream data, direct processing would lead to excessive computational load, and fault characteristics typically exist in the form of signal changes within local time periods, requiring segmentation to extract local features. Therefore, this application further employs a synchronous segmentation strategy to cut the two digital signal streams, thereby extracting signal segments of fixed length and maintaining the temporal correspondence between segments. This transforms continuous signals into discrete analysis units, reducing data processing volume, improving the efficiency of subsequent feature analysis, and ensuring that each vibration segment and its corresponding EMI segment contain information from the same time period. This provides structured samples for establishing cross-modal feature associations and accurately identifying noise pollution patterns, avoiding feature extraction bias due to inappropriate data formats.

[0023] Specifically, in one possible embodiment, the data stream synchronization segmentation unit 122 is implemented as follows: First, a fixed segment length and sliding step size are set. The segment length is determined based on the forging press operating cycle and the duration of the fault characteristic. The sliding step size is set according to a certain proportion of the segment length to ensure characteristic continuity. Then, the synchronization segmentation trigger mechanism is activated, starting from the beginning position of the two digital signal streams, simultaneously extracting data of a set length to form the first set of vibration segments and EMI segments. Next, according to the set step size, the extraction window is synchronously moved on the two signal streams, and the extraction operation is repeated until the end of the signal stream. Finally, all extracted vibration segments are sequentially collected to form a vibration segment set, and the corresponding EMI segments are collected to form an EMI segment set, ensuring that the segments in the two sets correspond one-to-one.

[0024] Specifically, the standardization and tensorization unit 123 is used to standardize and tensorize the vibration segment set and the EMI segment set to obtain the noisy vibration tensor and the electromagnetic interference reference tensor. It should be understood that since the vibration segments and EMI segments come from different types of sensors, there are problems with inconsistent dimensions and large differences in amplitude range. Therefore, this application further performs standardization processing and tensor construction on the two sets of segments to eliminate differences in data dimensions and amplitudes, and integrate the discrete segments into a unified data format. This enables the processed data to have a unified numerical range and structural standard, improving the stability and convergence speed of model training, and ensuring that the cross-modal denoising network can efficiently read and mine the deep correlation between the two data streams, providing highly adaptable input data for accurate denoising.

[0025] Specifically, in one possible embodiment, the standardization and tensor unit 123 is implemented as follows: First, standardization is performed on each segment in the vibration segment set and the EMI segment set: For a single vibration segment, the mean and standard deviation of all its data points are calculated, and each data point is converted into a standardized value through linear transformation; for a single EMI segment, the mean and standard deviation are calculated using the same method and standardization is completed, so that the values ​​of all segments are distributed within a reasonable range. Subsequently, the standardized vibration segments are stacked according to the dimensions of sample number, signal channel number, and time step, where the sample number is the total number of vibration segments, the signal channel number corresponds to the measurement dimension of the vibration signal, and the time step corresponds to the number of data points in each segment. Then, the EMI segments are stacked according to the same dimensions to ensure that each dimension of the vibration tensor and the EMI tensor is completely matched, ultimately forming a noisy vibration tensor and an electromagnetic interference reference tensor, which can be directly used for subsequent cross-modal data correlation processing.

[0026] In the aforementioned forging press condition monitoring system based on intelligent sensors, the cross-modal denoising module 130 is used to input the noisy vibration tensor and the electromagnetic interference reference tensor into the cross-modal denoising network to obtain the denoised vibration tensor. It should be understood that since the noisy vibration tensor still contains noise components related to the electromagnetic interference reference tensor, and noise and fault features are highly intertwined in the time and frequency domains, processing the noisy vibration tensor alone cannot accurately separate the two. Therefore, this application further inputs the noisy vibration tensor and the electromagnetic interference reference tensor into the cross-modal denoising network to explore the nonlinear mapping relationship between electromagnetic interference and vibration noise, achieving targeted noise suppression. In this way, while removing electromagnetic interference noise, early fault features in the vibration signal are preserved to the maximum extent, providing a high-fidelity signal foundation for subsequent feature extraction and fault diagnosis, and avoiding the loss of fault information or noise residue affecting diagnostic accuracy.

[0027] In particular, in one specific embodiment, Figure 4 This is a block diagram of a cross-modal noise reduction module in a smart sensor-based forging press condition monitoring system according to an embodiment of this application. Figure 4 As shown, the cross-modal denoising module 130 includes: a feature encoding unit 131, used to perform feature encoding on the noisy vibration tensor and the electromagnetic interference reference tensor to obtain the original vibration time-domain feature vector and the electromagnetic interference time-domain feature vector; a feature interaction unit 132, used to input the original vibration time-domain feature vector and the electromagnetic interference time-domain feature vector into the cross-modal attention interaction module to obtain the context vector; and a feature fusion and decoding unit 133, used to perform feature fusion and decoding on the context vector and the original vibration time-domain feature vector to obtain the denoised vibration tensor.

[0028] Specifically, the feature encoding unit 131 is used to perform feature encoding on the noisy vibration tensor and the electromagnetic interference reference tensor to obtain the original vibration time-domain feature vector and the electromagnetic interference time-domain feature vector. It should be understood that since both the noisy vibration tensor and the electromagnetic interference reference tensor are three-dimensional time-domain data, with high dimensionality and containing redundant information, direct cross-modal interaction would lead to a large computational load and low correlation accuracy. Therefore, this application further performs feature encoding on the two tensors separately to extract key features from the time-domain data, achieving data dimensionality reduction and information condensation. In this way, high-dimensional tensors can be transformed into compact time-domain feature vectors, reducing the computational load of subsequent interaction modules and highlighting the core features of electromagnetic interference and vibration signals, providing a clear feature foundation for accurately establishing the correlation between the two, and avoiding redundant information interfering with the feature correlation effect.

[0029] Specifically, in one possible embodiment, the feature encoding unit 131 is implemented as follows: For the noisy vibration tensor, a one-dimensional convolutional neural network is used as the encoding branch, and multiple sets of convolutional kernels are set to capture vibration features at different time scales. The data dimension is gradually compressed through convolution and pooling operations, and finally the original vibration time-domain feature vector is output. For the electromagnetic interference reference tensor, a structure-matched one-dimensional convolutional neural network is used as the encoding branch. Similarly, noise feature patterns of EMI signals are extracted through convolution and pooling, and the electromagnetic interference time-domain feature vector is output. The parameters of the two encoding branches are trained independently to ensure that the features extracted by each branch can accurately reflect the core information of the corresponding data.

[0030] Specifically, the feature interaction unit 132 is used to input the original vibration time-domain feature vector and the electromagnetic interference time-domain feature vector into the cross-modal attention interaction module to obtain a context vector. It should be understood that since the original vibration time-domain feature vector and the electromagnetic interference time-domain feature vector are still in an independent state, the correspondence between the electromagnetic interference features and the vibration noise features cannot be clearly defined, and direct fusion will lead to a deviation in the noise suppression direction. Therefore, this application further inputs the two feature vectors into the cross-modal attention interaction module to dynamically calculate the correlation weights of the two features, focusing on the electromagnetic interference features associated with vibration noise. In this way, a context vector containing accurate noise correlation information can be generated, enabling the subsequent fusion process to directionally identify and suppress noise components in the vibration features, avoiding the introduction of new interference due to irrelevant electromagnetic interference features, and improving the accuracy of noise separation.

[0031] Specifically, in one possible embodiment, the feature interaction unit 132 is implemented as follows: First, the original vibration time-domain feature vector is used as the query feature vector to locate the area affected by interference. Simultaneously, the electromagnetic interference time-domain feature vector is used as the associated feature vector and the information feature vector, respectively. The associated feature vector is used to calculate the correlation with the query feature vector, and the information feature vector provides detailed information about the interference features. The module first calculates the correlation weight between the query feature vector and the associated feature vector through matrix operations. The weight value directly reflects the degree of correlation between each part of the vibration feature and the interference. Then, this weight is applied to the information feature vector, and a context vector is generated through weighted summation. This vector accurately depicts the distribution pattern of electromagnetic interference in the vibration features.

[0032] Specifically, the feature fusion and decoding unit 133 is used to fuse and decode the context vector and the original vibration time-domain feature vector to obtain the denoised vibration tensor. It should be understood that since the context vector contains the correlation information between electromagnetic interference and vibration noise, while the original vibration time-domain feature vector contains fault features and residual noise, relying solely on a single vector cannot achieve signal reconstruction and noise separation. Therefore, this application further fuses and decodes the two vectors to utilize the context vector to directionally separate the noise components in the original vibration features and reconstruct the time-domain signal. In this way, the fused clean features can be transformed into a denoised vibration tensor with the same dimensions as the original noisy vibration tensor, preserving both the early fault features in the vibration signal and completely removing electromagnetic interference noise, providing a high-quality signal source for subsequent time-frequency domain feature extraction.

[0033] Specifically, in one possible embodiment, the feature fusion and decoding unit 133 is implemented as follows: First, feature fusion preparation is performed by inputting the context vector into a one-dimensional linear transformation module. Through parameter mapping, it is transformed into an interference feature vector with the same dimension as the original vibration time-domain feature vector, ensuring that the two can be directly fused. Then, a subtraction fusion method is used to perform element-wise difference operations on the original vibration time-domain feature vector and the transformed interference feature vector. This operation removes interference components from the original vibration features, obtaining refined features containing only the core fault information. Finally, signal decoding is performed by using a one-dimensional transposed convolutional network as the decoding module. Multiple sets of transposed convolutional kernels are used to gradually expand the dimension of the refined features, restoring the temporal resolution and complete structure of the signal. The final output is a denoised vibration tensor that is completely identical to the input noisy vibration tensor in terms of sample number and time step, ensuring compatibility with subsequent processing.

[0034] In particular, in the complex industrial environment of strong electromagnetic interference in forging presses, the vibration signal of the sensor and the electromagnetic interference noise (EMI) are not simply linearly superimposed. The electromagnetic field may produce a nonlinear modulation effect on the sensor's sensitive element or signal transmission link, causing the noise shape and amplitude to change with the instantaneous state of the actual vibration signal, exhibiting a dynamic coupling characteristic that is strongly correlated with the signal state.

[0035] In the technical mechanism of the above embodiments, after obtaining the vibration features and electromagnetic interference context vector, feature fusion is performed using a simple subtraction or concatenation method, which constitutes its core technical defect. Subtraction fusion implicitly assumes that noise and signal are linearly superimposed, failing to remove the aforementioned nonlinear, state-dependent coupled noise. Concatenation fusion, on the other hand, blurs the signal-noise decoupling task with a clear physical background into a general black-box fitting task. This not only places higher demands on the performance and training data volume of the subsequent decoder network, but more importantly, it fails to fully utilize a crucial prior physical knowledge: the core role of the electromagnetic interference context vector should be to apply a refined dynamic correction to the vibration features, rather than acting as a parallel independent information source. Therefore, both methods ignore the necessary dynamic gating relationship between the electromagnetic interference context vector and the vibration features, failing to reflect the channel-specific, time-varying nonlinear modulation effect applied to different feature components in the electromagnetic interference context vector and vibration features, resulting in incomplete denoising and easy damage to the useful signal.

[0036] To address the aforementioned technical shortcomings, this solution proposes a Dynamic Noise Decoupling Gated Network (DNDG) mechanism. It abandons the simple feature fusion paradigm and introduces a gating unit driven by the electromagnetic interference context vector to dynamically and nonlinearly modulate and correct the original vibration features, thereby achieving deep decoupling of signal and noise. Specifically, the feature fusion and decoding unit 133 includes: an adaptive gating generation subunit, used to input the context vector into an EMI-driven adaptive gating module to obtain information retention gating and noise suppression gating; a feature dynamic decoupling subunit, used to perform gated modulation-based dynamic decoupling of the original vibration time-domain feature vector based on information retention gating, noise suppression gating, and the context vector to obtain refined vibration features; and a signal reconstruction subunit, used to reconstruct the signal from the refined vibration features to obtain a denoised vibration tensor.

[0037] Specifically, firstly, the context vector is input into an EMI-driven adaptive gating module to obtain information retention gating and noise suppression gating. It should be understood that in this preferred mechanism, a refined signal for controlling subsequent feature retention and correction needs to be dynamically generated based on the real-time perceived electromagnetic interference noise pattern, replacing the fixed fusion strategy in the original mechanism. During execution, the context vector, generated by the cross-modal attention mechanism and containing the most relevant noise pattern at the current moment, is input in parallel to two structurally independent but functionally different fully connected neural network layers. Information retention gating and noise suppression gating are generated respectively through nonlinear transformations of their respective network layers and the application of a Sigmoid activation function.

[0038] This process simulates an intelligent decision-making process: the information retention gating, by learning context vectors, can dynamically judge and quantify the credibility of each feature dimension in vibration features. If a feature dimension is less contaminated by noise, its gating value approaches 1, and vice versa. In this way, the noise information contained in the context vector can be transformed into control weights for the retention and suppression of vibration feature channels, providing a dynamic and adaptive basis for subsequent refined feature correction.

[0039] Its calculation process is defined by the following formula:

[0040] in, Representative information retention gating; Represents noise suppression gating; Use the Sigmoid activation function; and These are the weight matrices for generating the two gated neural network layers, respectively. and This is the corresponding bias vector; This is the input EMI context vector.

[0041] Building upon this, based on information retention gating, noise suppression gating, and context vectors, the original vibration time-domain feature vectors are dynamically decoupled under gated modulation to obtain refined vibration features. In other words, using the dynamically gated signal generated in the previous step, a nonlinear separation operation with clear physical meaning is performed on the vibration features to achieve effective decoupling of noise and signal. This is the core innovation of this scheme.

[0042] During execution, on the one hand, a noise correction component is generated based on the context vector. Specifically, the context vector is passed through an independent nonlinear transformation network layer to generate a noise correction component directly related to the noise morphology. This nonlinear transformation network layer is a fully connected feedforward type, with the number of neurons in the input layer matching the dimension of the context vector, transmitting only signals without information loss. The hidden layer is set to a single layer, with twice the number of neurons as the input layer. The weight matrix is ​​initialized using Xavier, and the bias is initially set to 0. Nonlinear mapping is achieved using the tanh activation function. The number of neurons in the output layer matches the dimension of the original vibration time-domain feature vector, and the noise correction component is output after a linear transformation. On the other hand, based on information retention gating and noise suppression gating, gated modulation and feature refinement are performed on the noise correction component and the original vibration time-domain feature vector to obtain refined vibration features. Specifically, element-wise gated modulation and difference are performed on the vibration features and the newly generated noise correction component using information retention gating and noise suppression gating. This process no longer involves simple signal subtraction but performs a composite operation: through element-wise multiplication of the vibration features with the information retention gating, adaptive preservation of the original vibration features is achieved, allowing only high-confidence feature components to pass through. Simultaneously, by multiplying the noise suppression gate and the noise correction component element-wise, dynamic weighting of the noise correction intensity is achieved. The final subtraction operation completes a refined, state-dependent denoising process. In this way, a refined vibration characteristic after dynamic decoupling and fine correction can be obtained, which eliminates the influence of nonlinear coupling noise to the greatest extent while ensuring that the true signal characteristics related to the fault are preserved.

[0043] Its calculation process is defined by the following formula:

[0044] in, Represents the noise correction component; It is the hyperbolic tangent activation function; and To generate the network layer weights and biases for this component; The final output is the refined vibration characteristic; The input represents the original vibration characteristics; ⊙ represents the Hadamard product, which is the element-wise multiplication of a matrix or vector.

[0045] Finally, the refined vibration features are reconstructed to obtain the denoised vibration tensor. That is, the purified, refined vibration features located in the abstract feature space need to be restored to physically meaningful time-domain vibration signals. The execution process is relatively straightforward: the refined vibration features are fed into a pre-trained decoder network. Since the features input to the decoder are already a pure representation with a very high signal-to-noise ratio and noise components largely removed, the decoder does not need to undertake complex decoupling tasks. This ensures a closed loop in the entire denoising process, transforming the feature operation results at the algorithm level into specific signals that engineers can analyze or use in subsequent diagnostic modules. The purpose and effect of this step is to efficiently and accurately reconstruct the denoised vibration tensor, which not only filters out strong electromagnetic interference macroscopically but also preserves the shape and details of early weak fault impact signals microscopically, significantly improving the overall signal fidelity.

[0046] The fundamental purpose of this technical solution is to achieve high-fidelity denoising of contaminated sensor signals in industrial environments with strong electromagnetic interference, such as forging presses, and to accurately restore the true vibration signals reflecting the health status of the equipment. By employing a dynamic noise decoupling gating network, this solution achieves significant technical results: it successfully overcomes the limitations of traditional denoising methods that rely on the simplified assumption of linear superposition of noise and signal, learning and compensating for complex nonlinear coupling effects through a data-driven approach. Furthermore, it can dynamically and nonlinearly gate and modulate each characteristic channel of the vibration signal based on the real-time sensed electromagnetic environment. This effectively suppresses strong interference while preserving, to the maximum extent possible, the weak impact characteristics related to early faults submerged in noise. This not only provides a solid and reliable data foundation for subsequent advanced diagnostic analysis tasks such as health status assessment and remaining life prediction based on this signal, but also greatly improves the robustness and diagnostic accuracy of the entire condition monitoring system.

[0047] In the aforementioned intelligent sensor-based forging press condition monitoring system, the feature enhancement and extraction module 140 is used to enhance and extract the time-frequency domain features of the denoised vibration tensor to obtain a fault feature spectrum. It should be understood that although the denoised vibration tensor has eliminated electromagnetic interference, it still exists in the form of a time-domain signal. Early fault characteristics of key components of the forging press (such as bearings and gears) manifest as weak periodic impacts, which are easily masked by the background vibration of normal equipment operation in the time-domain waveform, making direct identification difficult. Therefore, this application further uses time-frequency domain conversion and feature enhancement to convert the time-domain vibration signal into a frequency-domain feature spectrum, thereby highlighting the frequency information of periodic fault characteristics. In this way, the fault impact characteristics hidden in the time domain can be presented in the frequency domain as clear frequency peaks, forming a fault feature spectrum containing equipment health status information, providing a clear analytical basis for subsequent comparison with theoretical fault frequencies and achieving accurate fault diagnosis.

[0048] In particular, in one specific embodiment, Figure 5 This is a block diagram of the feature enhancement and extraction module in a smart sensor-based forging press condition monitoring system according to an embodiment of this application. Figure 5 As shown, the feature enhancement and extraction module 140 includes: an analytical envelope extraction unit 141, used to perform analytical signal reconstruction and envelope extraction on the denoised vibration tensor to obtain an envelope signal set; and an envelope spectrum analysis unit 142, used to perform envelope spectrum calculation and spectrum generation on the envelope signal set to obtain a fault feature spectrum.

[0049] Specifically, the analytical envelope extraction unit 141 is used to perform analytical signal reconstruction and envelope extraction on the denoised vibration tensor to obtain an envelope signal set. It should be understood that, since the early fault characteristics of the time-domain signal in the denoised vibration tensor are manifested as weak periodic impacts superimposed on the high-frequency resonance components, it is difficult to distinguish fault impacts from normal vibrations directly from the time-domain waveform. The fault impact modulates the amplitude of the high-frequency resonance components, and this amplitude modulation information is implicit in the signal envelope. Therefore, this application further separates the amplitude modulation information through analytical signal reconstruction and envelope extraction to reveal the periodic fault characteristics hidden in the high-frequency resonance. In this way, the weak fault impact characteristics in the time-domain signal can be transformed into clear periodic fluctuations in the envelope signal, forming an envelope signal set containing fault periodic information, laying the foundation for accurate identification of fault characteristic frequencies in subsequent frequency domain analysis.

[0050] Specifically, in one possible embodiment, the analytical envelope extraction unit 141 is implemented as follows: First, a single time-domain signal segment is extracted from the denoised vibration tensor, and a Hilbert transform is applied to it. The Hilbert transform generates a corresponding orthogonal component for the input real signal, which has a 90-degree phase lag relative to the original signal at all frequencies. Second, the original time-domain signal is used as the real part, and the orthogonal component generated by its Hilbert transform is used as the imaginary part to construct an analytical signal in complex form. This analytical signal is a complex representation containing all amplitude and phase information of the original signal. Finally, the instantaneous modulus of this analytical signal at each time point is calculated, i.e., the sum of the squares of the real and imaginary parts of the complex number is taken, and the square root is obtained. The result of this instantaneous modulus calculation is mathematically the envelope of the original signal. This envelope signal reflects the law of high-frequency resonant energy changing with time, thereby revealing the hidden periodic impact characteristics of the fault. Repeating this process for all time-domain segments in the denoised vibration tensor yields a set of envelope signals containing fault modulation information.

[0051] Specifically, the envelope spectrum analysis unit 142 is used to calculate the envelope spectrum and generate a spectrum diagram of the envelope signal set to obtain a fault feature spectrum. It should be understood that since the envelope signal set still exists in the time domain, the periodic fault characteristics within it manifest as fluctuations in the time domain, making it impossible to directly compare with the theoretical fault characteristic frequencies of key components of the forging press (such as the passing frequency of the inner and outer rings of the bearing, and the gear meshing frequency), thus hindering accurate identification of the fault type. Therefore, this application further performs frequency domain transformation and spectrum diagram generation on the envelope signal set, converting the time-domain fluctuation characteristics into frequency domain peak values ​​to establish a correspondence between the envelope signal and the theoretical fault frequency. This allows for a clear presentation of fault-related characteristic frequency peaks in the frequency domain spectrum, forming a fault feature spectrum. This enables personnel to quickly determine the fault type and location by comparing the characteristic peak values ​​with the theoretical fault frequency, improving the accuracy and efficiency of fault diagnosis.

[0052] Specifically, in one possible embodiment, the envelope spectrum analysis unit 142 is implemented as follows: First, a single envelope signal is extracted from the envelope signal set and subjected to DC component removal processing. Redundant peaks at 0Hz in the frequency domain are eliminated by subtracting the signal's mean. Second, a Hanning window function is applied to the DC-removed envelope signal. The smooth transition characteristics of the window function reduce spectral leakage caused by signal truncation, ensuring concentrated frequency energy. Subsequently, a fast Fourier transform is performed on the windowed envelope signal to convert the time-domain signal into a complex frequency-domain signal. The amplitude of the frequency-domain signal is then calculated to obtain a single-sided amplitude spectrum. Finally, the numerical range of the frequency axis is determined based on the sampling frequency, and the amplitude spectrum is correlated with the frequency axis to generate a spectrum unit containing the frequency-amplitude correspondence. The above operations are repeated for all envelope signals in the envelope signal set, integrating all spectrum units to ultimately form a fault characteristic spectrum.

[0053] In the aforementioned intelligent sensor-based forging press condition monitoring system, the fault diagnosis module 150 is used to perform fault diagnosis based on the fault feature spectrum to obtain the fault type and confidence score. It should be understood that although the fault feature spectrum presents frequency domain peaks related to the equipment status, these peaks are only corresponding data of frequency and amplitude, and cannot directly correlate with the fault modes of specific components of the forging press. Furthermore, there may be false peaks caused by signal fluctuations or residual interference on-site, and misjudgment is likely to occur if only the peak frequency is relied upon. Therefore, this application further compares the fault feature spectrum with a preset theoretical fault frequency database for components and quantitatively evaluates the reliability of the diagnostic results to achieve accurate matching of fault types and determination of diagnostic credibility. In this way, the key components (such as bearings and gears) and specific fault modes (such as outer ring wear and tooth surface cracks) corresponding to the fault can be clearly identified. At the same time, the confidence score distinguishes between real faults and false signals, avoiding false alarms or missed alarms, providing maintenance personnel with clear fault location basis and decision support, ensuring accurate and efficient maintenance measures, and reducing the risk of unplanned downtime.

[0054] Specifically, in one possible embodiment, the fault diagnosis module 150 is implemented as follows: First, a preset theoretical fault frequency database for components is invoked. This database pre-calculates and stores theoretical fault characteristic frequencies and their harmonic frequencies under various fault modes based on the geometric parameters, installation positions, and real-time speeds of key components of the forging press (such as bearings and gears). Second, significant peaks are extracted from the fault characteristic spectrum. To filter out background noise, a dynamic threshold is set, which can be specifically set as the sum of the average value of the global amplitude of the spectrum and three times the standard deviation. Only frequency peaks with amplitudes exceeding this threshold are selected. Next, the extracted peak frequencies are compared one by one with the theoretical fault frequencies in the database. If the deviation rate between a peak frequency and the theoretical frequency is within 2%, it is considered a valid match. Subsequently, a confidence score is calculated for each match. This score comprehensively considers factors such as the normalized amplitude of the matched peak and the integrity of the harmonic components (e.g., whether there are second and third harmonics with amplitudes not less than 30% of the fundamental frequency). Finally, all matching items are sorted from highest to lowest confidence score. If the highest score exceeds the preset confidence threshold (e.g., 0.85), the corresponding fault is defined as the final diagnosed fault type, and the fault type and corresponding confidence score are output synchronously.

[0055] In summary, the intelligent sensor-based forging press condition monitoring system based on the embodiments of this application is explained. It synchronously collects vibration signals characterizing the mechanical state of the equipment and electromagnetic interference signals reflecting the noise characteristics of the surrounding environment. A cross-modal deep learning model is introduced to deeply mine and establish the nonlinear mapping relationship between the two heterogeneous signals. This guides the model to learn and identify the propagation paradigm and contamination characteristics of noise in the vibration signal, thereby reconstructing a high-fidelity clean fault signal. Based on this, the purified signal undergoes in-depth time-frequency domain feature enhancement and intelligent analysis to achieve accurate diagnosis of early, subtle faults in the forging press. This improves the safety and stability of equipment operation, effectively extends the service life and operating cycle of the forging press, thereby reducing unplanned downtime losses and maintenance costs, and comprehensively enhancing the intelligent level of health management for forging equipment.

[0056] As described above, the intelligent sensor-based forging press condition monitoring system according to embodiments of this application can be implemented in various wireless terminals. In one possible implementation, the intelligent sensor-based forging press condition monitoring system according to embodiments of this application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the intelligent sensor-based forging press condition monitoring system can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the intelligent sensor-based forging press condition monitoring system can also be one of many hardware modules of the wireless terminal.

[0057] Alternatively, in another example, the smart sensor-based forging press condition monitoring system and the wireless terminal can also be separate devices, and the smart sensor-based forging press condition monitoring system can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

Claims

1. A forging press condition monitoring system based on intelligent sensors, characterized in that, include: The raw signal acquisition module is used to acquire raw vibration signals and raw EMI signals; The preprocessing and segmentation module is used to preprocess and segment the original vibration signal and the original EMI signal to obtain the noisy vibration tensor and the electromagnetic interference reference tensor. The cross-modal denoising module is used to input the noisy vibration tensor and the electromagnetic interference reference tensor into the cross-modal denoising network to obtain the denoised vibration tensor. The feature enhancement and extraction module is used to enhance and extract the time-frequency domain features of the denoised vibration tensor to obtain the fault feature spectrum; The fault diagnosis module is used to perform fault diagnosis based on the fault feature spectrum to obtain the fault type and confidence score.

2. The forging press condition monitoring system based on intelligent sensors according to claim 1, characterized in that, The preprocessing and segmentation module includes: The signal conditioning and digitization unit is used to perform analog signal conditioning and synchronous digitization on the original vibration signal and the original EMI signal to obtain digital vibration signal stream and digital EMI signal stream; The data stream synchronization segmentation unit is used to perform data stream synchronization segmentation on the digital vibration signal stream and the digital EMI signal stream to obtain the vibration segment set and the EMI segment set. The normalization and tensorization unit is used to normalize and tensorize the vibration fragment set and the EMI fragment set to obtain the noisy vibration tensor and the electromagnetic interference reference tensor.

3. The forging press condition monitoring system based on intelligent sensors according to claim 1, characterized in that, The cross-modal denoising module includes: The feature encoding unit is used to encode the features of the noisy vibration tensor and the electromagnetic interference reference tensor to obtain the original vibration time-domain feature vector and the electromagnetic interference time-domain feature vector. The feature interaction unit is used to input the original vibration time-domain feature vector and the electromagnetic interference time-domain feature vector into the cross-modal attention interaction module to obtain the context vector; The feature fusion and decoding unit is used to fuse and decode the context vector and the original vibration time-domain feature vector to obtain the denoised vibration tensor.

4. The forging press condition monitoring system based on intelligent sensors according to claim 3, characterized in that, The feature fusion and decoding unit includes: An adaptive gating generation subunit is used to input the context vector into the EMI-driven adaptive gating module to obtain information-preserving gating and noise-suppression gating; The feature dynamic decoupling subunit is used to perform feature dynamic decoupling under gated modulation on the original vibration time-domain feature vector based on information retention gate, noise suppression gate and context vector to obtain refined vibration features; The signal reconstruction subunit is used to reconstruct the signal from the refined vibration characteristics to obtain the denoised vibration tensor.

5. The forging press condition monitoring system based on intelligent sensors according to claim 4, characterized in that, The feature dynamic decoupling subunit is used for: Noise correction components are generated based on the context vector; Based on information retention gating and noise suppression gating, the noise correction component and the original vibration time-domain feature vector are gated and modulated and refined to obtain refined vibration features.

6. The forging press condition monitoring system based on intelligent sensors according to claim 5, characterized in that, Based on information retention gating and noise suppression gating, gated modulation and feature refinement are performed on the noise correction component and the original vibration time-domain feature vector to obtain refined vibration features. This includes: gating and refining the noise correction component and the original vibration time-domain feature vector using the following formula: in, Information retention gating; ⊙ represents the Hadamard product, i.e., element-wise multiplication of a matrix or vector; This represents the original vibration time-domain feature vector; For noise suppression gating; For noise correction components; The vibration characteristics are those after refining.

7. The forging press condition monitoring system based on intelligent sensors according to claim 1, characterized in that, The feature enhancement and extraction module includes: The analytical envelope extraction unit is used to perform analytical signal reconstruction and envelope extraction on the denoised vibration tensor to obtain an envelope signal set. The envelope spectrum analysis unit is used to perform envelope spectrum calculation and spectrum generation on the envelope signal set to obtain the fault characteristic spectrum.

Images