Field bus DP network intelligent diagnosis method and system

Abstract

The embodiment of the application discloses a field bus DP network intelligent diagnosis method and system, relates to the technical field of industrial automation control, and the system comprises: a signal enhancement unit which is provided with a differential amplification circuit, the signal enhancement unit comprises an adaptive gain control module and a noise suppression module, and is used for adjusting the intensity of a transmission signal of a physical layer to meet a target threshold value; a data acquisition unit which comprises a high-speed analog-digital conversion module and a field programmable gate array processing module, and is used for collecting target signals of the physical layer and data frames of a data link layer online; a data processing unit which comprises a multi-core processor, is internally provided with a fault mode library and a feature extraction algorithm module, is used for performing real-time analysis, fault detection and diagnosis processing on the collected target signals and data frames, and obtains diagnosis processing results; and a communication interface unit which supports various industrial communication interface protocols and is used for forwarding the diagnosis processing results.

Description

Technical Field

[0001] This application relates to the technical field of industrial automation control, and in particular to a fieldbus DP network intelligent diagnostic method and system. Background Technology

[0002] With the rapid development of industrial automation technology, fieldbus technology has been widely used in industrial control systems. Among them, PROFIBUS-DP, as a high-speed and reliable fieldbus standard, plays a crucial role in industrial fields such as power, chemical, and metallurgy. Especially in the control systems of large thermal power units, the DP bus network undertakes the critical task of connecting various intelligent instruments, actuators, and controllers, and its operational stability and reliability are directly related to the safe operation of the entire generator unit.

[0003] However, in actual operation, DP bus segments frequently encounter various complex technical problems. The most common are brief, sudden equipment failures and offline phenomena, characterized by short duration, high frequency, and complex causes. Traditional diagnostic methods mainly rely on handheld diagnostic devices or simple online monitoring tools, which have significant limitations. First, traditional diagnostic devices often only detect the signal quality at the physical layer, failing to deeply analyze the data frame content at the data link layer, leading to incomplete diagnostic results. Second, for transient faults lasting only milliseconds, traditional diagnostic methods struggle to accurately capture and locate them, often missing the optimal diagnostic window. Furthermore, when intermittent DP bus faults occur, maintenance personnel often spend considerable time troubleshooting, impacting production efficiency and, more importantly, posing potential risks to the safe operation of generator sets. Existing diagnostic technologies mostly remain at the passive response stage, lacking proactive prevention and intelligent analysis capabilities. Simultaneously, with the increasing number of devices in industrial fields and the improvement of communication speeds, higher demands are placed on the real-time performance, accuracy, and intelligence level of diagnostic equipment.

[0004] Therefore, there is an urgent need for a technical solution that can achieve comprehensive, real-time, and intelligent diagnosis of DP bus networks, taking into account both physical layer signal quality and data link layer data frame analysis. In particular, it should provide accurate and rapid diagnosis and location capabilities for transient and sudden faults that are difficult to detect using traditional methods, thereby significantly improving maintenance efficiency and ensuring the safe and stable operation of industrial sites. Summary of the Invention

[0005] This application proposes an intelligent diagnostic method and system for fieldbus (DP) networks to address the shortcomings of the aforementioned prior art.

[0006] According to a first aspect of the embodiments of this application, a fieldbus (DP) network intelligent diagnostic system is provided, connected in the fieldbus communication link of an industrial control system, including a signal enhancement unit, a data acquisition unit, a data processing unit, and a communication interface unit, wherein... The signal enhancement unit is configured with a differential amplifier circuit, and the signal enhancement unit includes an adaptive gain control module and a noise suppression module, which are used to adjust the strength of the transmitted signal of the physical layer to meet the target threshold. The data acquisition unit includes a high-speed analog-to-digital conversion module and a field-programmable gate array (FPGA) processing module, which are used to acquire the target signal of the physical layer and the data frame of the data link layer online. The data processing unit includes a multi-core processor and has a built-in fault mode library and feature extraction algorithm module, which is used to perform real-time analysis, fault detection and diagnosis on the collected target signal and data frame, and obtain the diagnosis results. The communication interface unit supports multiple industrial communication interface protocols and is used to forward the diagnostic processing results.

[0007] In some implementations, the diagnostic processing results include signal strength statistics, wherein, The adaptive gain control module implements multi-level gain adjustment based on a digital potentiometer, dynamically generates control logic based on the signal strength statistics, and controls the output signal amplitude to remain stable within a preset range through a closed-loop feedback mechanism.

[0008] In some implementations, the high-speed analog-to-digital conversion module employs a multi-channel interleaved sampling mode; The field-programmable gate array (FPGA) processing module has a built-in first-in-first-out (FIFO) memory and adds a frame header containing timestamps, trigger types, and verification information to the acquired data before transmission.

[0009] In some implementations, the fault mode library includes physical layer fault modes, data link layer fault modes, and application layer fault modes. Each fault mode is represented as a multidimensional feature vector after feature extraction by the feature extraction algorithm module.

[0010] In some implementations, the feature extraction algorithm module employs a cascaded process of wavelet transform and Fourier transform to extract the time-domain, frequency-domain, and time-frequency-domain features of the fault signal.

[0011] According to a second aspect of this application, a fieldbus DP network intelligent diagnostic method is provided, implemented based on the above-mentioned fieldbus DP network intelligent diagnostic system, including: A sliding window algorithm is used to continuously monitor the data flow of the fieldbus DP network. The sliding window has a preset window size and sliding step size, which is used to synchronously record the data frame transmission status of the data link layer and the signal quality parameters of the physical layer. Based on the abnormal data identified during continuous monitoring, feature extraction and anomaly classification are performed. Based on a preset fault mode library, fault matching is performed on the abnormal data after feature extraction and anomaly classification. The fault mode library includes multi-layer fault modes of the physical layer, data link layer and application layer. A fuzzy matching algorithm is used to calculate the similarity between the current features of the abnormal data and the patterns in the fault mode library, and the matching result between the current features of the abnormal data and the fault mode library is calculated. If the similarity exceeds a preset threshold, it is determined that the match is successful and the corresponding fault location information is output. If the similarity does not exceed the preset threshold, it is determined that there is no matching fault mode in the current monitoring window. Based on the matching results, a structured fault diagnosis report is generated and maintenance recommendations are provided.

[0012] In some implementations, the feature extraction and anomaly classification processing based on the abnormal data identified during continuous monitoring includes: Based on the cascaded processing of wavelet transform and Fourier transform, the abnormal data is analyzed and its features are extracted. Based on a recognition algorithm that uses a preset threshold and a preset anomaly pattern, the abnormal data that has undergone feature extraction processing is matched and classified with the preset anomaly pattern. Specifically, the wavelet transform performs multi-resolution analysis on the abnormal data based on a preset wavelet basis and denoises the abnormal data based on an improved threshold function for primary feature pre-extraction; the Fourier transform performs adaptive window function selection based on overlapping segmentation processing to extract the time-domain features, frequency-domain features, and time-frequency-domain features of the abnormal data for deep frequency-domain feature analysis.

[0013] In some implementations, the fault diagnosis report includes at least a fault timestamp, fault type, fault location, severity, and confidence index; the maintenance recommendations are generated based on statistical analysis of historical fault data and a rule engine, and cover equipment adjustment, parameter optimization, network maintenance, and preventive maintenance measures.

[0014] In some implementations, generating a structured fault diagnosis report includes: The structured fault diagnosis report is generated based on the hierarchical early warning mechanism. The structured fault diagnosis report is divided into multiple levels, and different levels correspond to different processing priorities.

[0015] In some embodiments, the method further includes: Intermittent faults in the matching results are identified, and the recurrence probability of the intermittent faults is assessed by fusing a time-series prediction model of historical fault event sequences with a physical model of equipment degradation, and a preventive maintenance plan is generated.

[0016] The beneficial effects of the fieldbus (DP) network intelligent diagnostic method and system according to the embodiments of this application include at least the following: This application's embodiment of a fieldbus DP network intelligent diagnostic system ensures the original quality of the analyzed signal through physical layer signal enhancement; it employs high-speed, high-precision data acquisition to capture nanosecond-level instantaneous anomalies; and combined with a multi-level fault mode library and cascaded feature extraction algorithms, it achieves comprehensive and in-depth analysis from signal quality to protocol integrity, avoiding the one-sidedness of single-dimensional diagnosis. It can achieve a diagnostic accuracy rate approaching or exceeding 95%, effectively identifying faults that are difficult to detect using traditional methods and significantly improving diagnostic accuracy and comprehensiveness. This application embodiment can automatically complete the entire process of monitoring, analysis, matching, and location in real time. Through fuzzy matching algorithms and a fault mode library, it can directly output the topology path coordinates of the faulty device after successful matching and generate a structured diagnostic report containing specific location, type, and severity. This achieves accurate and rapid fault location, guiding maintenance personnel directly to the problem point, reducing fault location time by an average of over 80%, greatly reducing the blind troubleshooting work of maintenance personnel, and effectively improving the efficiency of fault location and troubleshooting. This application embodiment utilizes a high-speed data acquisition unit and a sliding window monitoring algorithm to achieve seamless, high-density sampling and analysis of data streams. It solves the problem of fault omission caused by slow response speeds or low sampling rates in traditional methods. It can effectively capture, record, and analyze transient faults lasting only nanoseconds to milliseconds, as well as randomly occurring intermittent faults, achieving a breakthrough in the capture and analysis capabilities for transient and intermittent faults. This application embodiment introduces a recurrence probability assessment model to predict the future occurrence probability of intermittent faults. Based on the probability exceeding a threshold, a preventative maintenance plan is automatically generated. Combined with a confidence accumulation decision mechanism, false alarms are effectively filtered, improving the reliability of alarms. This transforms the maintenance strategy from post-fault repair to pre-fault prevention, achieving an intelligent upgrade from passive response to proactive prevention. It not only diagnoses existing faults but also predicts potential risks, guiding preventative maintenance. The communication interface unit of the diagnostic module in this application embodiment supports multiple mainstream industrial protocols, facilitating integration with existing systems. Modular hardware and software design makes functional units relatively independent, which facilitates upgrades or customization according to needs, enhances the system's life cycle and applicability, can flexibly adapt to different industrial sites, reduces deployment and modification difficulty, and makes the system highly versatile and easy to integrate and expand. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of a fieldbus DP network intelligent diagnostic system according to an embodiment of this application; Figure 2 This is a flowchart illustrating an intelligent diagnostic method for a fieldbus DP network according to an embodiment of this application. Figure 3 This is a software architecture diagram of a fieldbus DP network intelligent diagnostic system according to an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the various embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this disclosure to facilitate a better understanding of the disclosure. However, the technical solutions claimed in this disclosure can be implemented even without these technical details and with various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of this disclosure. The various embodiments can be combined with and referenced by each other without contradiction.

[0019] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed embodiments of the present application, but merely to illustrate selected embodiments of the present application. Other embodiments obtained by those skilled in the art based on the embodiments of the present application without inventive effort are all within the scope of protection of the embodiments of the present application.

[0020] This application discloses an intelligent diagnostic method and system for PROFIBUS-DP networks. The method, based on an intelligent diagnostic system, is primarily applied to fault diagnosis and maintenance of PROFIBUS-DP bus networks in industrial control systems. It is suitable for automation control systems in industries such as power, chemical, and metallurgy, particularly distributed control systems for large thermal power units. The purpose of this application is to achieve comprehensive, real-time, and intelligent diagnosis of DP bus networks from the physical layer to the data link layer through innovative hardware design and intelligent diagnostic algorithms. This significantly improves the accuracy of fault diagnosis and maintenance efficiency, solving the technical challenge of accurately measuring and locating transient, sudden equipment failures and offline problems that occur during DP bus network segment operation in existing technologies.

[0021] See attached document Figure 1As shown, this fieldbus DP network intelligent diagnostic system is connected to the fieldbus communication link of an industrial control system. The fieldbus DP network intelligent diagnostic system includes a signal enhancement unit, a data acquisition unit, a data processing unit, and a communication interface unit.

[0022] In some implementations, the signal enhancement unit is configured with a differential amplifier circuit, which includes an adaptive gain control module and a noise suppression module for adjusting the strength of the transmitted signal at the physical layer to meet a target threshold.

[0023] Traditional DP bus diagnostic devices often directly sample and analyze the raw signal, which can easily lead to the omission of important fault information when the signal quality is poor. The signal enhancement unit in this application is the core innovation; it employs an advanced differential amplifier circuit design and incorporates an adaptive gain control module and a noise suppression module to enhance the transmission signal of the DP bus physical layer.

[0024] For example, the adaptive gain control module can automatically adjust the amplification factor according to the real-time strength of the DP bus signal. Preferably, the gain adjustment range of the adaptive gain control module is 0dB to 40dB, the adjustment accuracy is 1dB, and the response time is less than 10μs. Based on this, the best signal quality can be obtained under various signal strength conditions.

[0025] For example, the noise suppression module uses digital filtering technology, which can effectively filter out electromagnetic interference and radio frequency interference in industrial environments, improving the signal-to-noise ratio by more than 20dB.

[0026] In one specific embodiment of this application, the signal enhancement unit is implemented using a high-performance differential amplifier circuit. The circuit uses the AD8132 differential amplifier as its core component, which has a bandwidth of 500MHz and a slew rate of 1000V / μs, meeting the amplification requirements of the high-speed DP bus signal. The input of the differential amplifier circuit employs a precision resistor matching network with a matching accuracy of 0.1%, ensuring signal symmetry and common-mode rejection capability. Gain control is achieved using the AD8403 digital potentiometer, which has 256 adjustment levels. Its control logic is dynamically generated by the ARM processor based on the signal strength statistics output from the A / D conversion module. The adjustment process uses a closed-loop feedback mechanism to ensure that the output signal amplitude of the differential amplifier circuit remains stable within a preset dynamic range. For example, the adjustment accuracy is 1dB, and the response time is less than 10μs. The noise suppression module uses an active filter circuit design, employing the MAX275 active filter chip to form an 8th-order Butterworth low-pass filter with a cutoff frequency set to 20MHz, effectively filtering out high-frequency noise.

[0027] In some implementations, the data acquisition unit includes a high-speed analog-to-digital conversion module (high-speed A / D conversion module) and a field-programmable gate array (FPGA) processing module, used to acquire the target signal of the physical layer and the data frame of the data link layer online.

[0028] For example, this high-speed analog-to-digital converter module employs a multi-channel interleaved sampling mode.

[0029] For example, the field-programmable gate array (FPGA) processing module has an internal first-in-first-out (FIFO) memory and adds a frame header containing timestamps, trigger types, and verification information to the acquired data before transmission.

[0030] The data acquisition unit in this application embodiment is the foundation for achieving high-precision diagnosis. The data acquisition unit in this application embodiment includes a high-speed A / D conversion module and an FPGA processing module. The high-speed A / D conversion module and the FPGA processing module realize DMA data transmission through the PCIe bus protocol.

[0031] For example, the high-speed A / D conversion module has a sampling rate of 100MS / s and a resolution of 12 bits. This high-performance configuration enables the diagnostic module to capture transient signal anomalies with a duration of not less than 10ns, providing a sufficient data foundation for subsequent analysis.

[0032] For example, the FPGA processing module is responsible for data preprocessing and caching. It adopts a parallel processing architecture and can process multiple data channels simultaneously to ensure data real-time performance.

[0033] More importantly, the data acquisition unit supports real-time acquisition of data frames with a maximum transmission rate of 12Mbps, fully meeting the high-speed communication requirements of modern industrial control systems.

[0034] In one specific embodiment of this application, the hardware implementation of the data acquisition unit adopts a combined architecture of a high-speed A / D converter and an FPGA. The A / D converter uses an ADS5400 chip, internally built with a 32KB FIFO memory. The data encapsulation unit adds frame header information containing timestamps, trigger types, and CRC16 checksums before transmitting the data to the ARM memory. This device has 12-bit resolution and a sampling rate of 1GS / s, enabling accurate capture of transient signal anomalies on the DP bus. The FPGA uses a Spartan-6 series XC6SLX45, which has abundant logic resources and high-speed I / O interfaces, enabling parallel data processing and real-time caching. The FPGA internally implements a 32KB, 16-bit FIFO memory, capable of caching sufficient data for subsequent analysis.

[0035] In some implementations, the data processing unit includes a multi-core processor with a built-in fault mode library and feature extraction algorithm module, used to perform real-time analysis, fault detection, and diagnostic processing on the acquired target signal and data frame, and obtain diagnostic processing results.

[0036] The data processing unit is the core intelligence of this application embodiment. This data processing unit adopts a high-performance ARM processor (ARM Cortex-A9 multi-core processor) architecture with a main frequency exceeding 1GHz, and integrates large-capacity memory and various peripheral interfaces. Most importantly, the data processing unit incorporates an innovative fault mode library and feature extraction algorithm. The fault mode library is a significant technological innovation of this application embodiment, comprising three layers: physical layer fault modes, data link layer fault modes, and application layer fault modes. Physical layer fault modes include typical fault modes such as signal attenuation, noise interference, reflection, and crosstalk; data link layer fault modes include protocol layer faults such as frame errors, checksum errors, timeout errors, and retransmission errors; and application layer fault modes include application layer faults such as device offline, parameter configuration errors, and communication protocol mismatch. This multi-layered fault mode library design makes the diagnostic results more comprehensive and accurate.

[0037] For example, the diagnostic processing results include signal strength statistics.

[0038] The adaptive gain control module uses a digital potentiometer to achieve multi-level gain adjustment, dynamically generates control logic based on the signal strength statistics, and controls the output signal amplitude to remain stable within a preset range through a closed-loop feedback mechanism.

[0039] For example, the fault mode library includes physical layer fault modes, data link layer fault modes and application layer fault modes. Each fault mode is represented as a multi-dimensional feature vector after feature extraction by the feature extraction algorithm module.

[0040] For example, the feature extraction algorithm module uses a cascaded process of wavelet transform and Fourier transform to extract the time-domain features, frequency-domain features and time-frequency-domain features of the fault signal.

[0041] Feature extraction algorithms are a key technology for achieving intelligent diagnosis. This application employs an innovative algorithm combining wavelet transform and Fourier transform, capable of simultaneously extracting the time-domain, frequency-domain, and time-frequency-domain features of fault signals. Wavelet transform is suitable for analyzing non-stationary signals and can effectively capture the abrupt changes in instantaneous faults; Fourier transform is suitable for analyzing stationary signals and can accurately identify the spectral characteristics of periodic faults. After initial feature pre-extraction is performed on the FPGA, deep feature analysis is conducted on the ARM. The combination of these two algorithms achieves a feature extraction accuracy of over 99.5%, providing a reliable foundation for subsequent fault matching and localization.

[0042] In one specific embodiment of this application, the hardware implementation of the data processing unit adopts an ARM processor architecture. The processor selected is an i.MX6Q, which has four ARM Cortex-A9 cores, a clock speed of 1GHz, and built-in 2GB DDR3 memory and 8GB eMMC storage. The processor runs a Linux operating system, providing abundant software resources and development tools. The data processing unit also integrates various peripheral interfaces, including a USB 2.0 interface, an SD card interface, and an Ethernet interface, facilitating system expansion and maintenance.

[0043] In some implementations, the communication interface unit supports multiple industrial communication interface protocols for forwarding the diagnostic processing results.

[0044] The communication interface unit ensures the diagnostic module's versatility and scalability. This unit supports multiple communication protocols, including PROFIBUS-DP, Industrial Ethernet, and RS485. This multi-interface design allows the diagnostic module to be easily integrated into existing industrial control systems without requiring large-scale modifications. Furthermore, physical layer compatibility is achieved through a protocol conversion chip, enabling remote diagnostics. Maintenance personnel can remotely access diagnostic results via a network, significantly improving maintenance efficiency.

[0045] In one specific embodiment of this application, the hardware implementation of the communication interface unit supports multiple communication protocols. The PROFIBUS-DP interface is implemented using the SPM1812 chip, which conforms to the PROFIBUS-DP standard and supports a maximum communication rate of 12Mbps. The industrial Ethernet interface is implemented using the KSZ8863 chip, which supports 10 / 100Mbps adaptive Ethernet communication. The RS485 interface is implemented using the MAX485 chip, which has ±15kV ESD protection and is suitable for harsh electromagnetic environments in industrial settings.

[0046] This application embodiment of the intelligent diagnostic system for a fieldbus DP network includes four core modules: a signal enhancement unit, a data acquisition unit, a data processing unit, and a communication interface unit. Exemplarily, this application embodiment also includes auxiliary modules such as a power management module and a clock management module.

[0047] This application's embodiment of a fieldbus DP network intelligent diagnostic system ensures the original quality of the analyzed signal through physical layer signal enhancement; it employs high-speed, high-precision data acquisition to capture nanosecond-level instantaneous anomalies; and combined with a multi-level fault mode library and cascaded feature extraction algorithms, it achieves comprehensive and in-depth analysis from signal quality to protocol integrity, avoiding the one-sidedness of single-dimensional diagnosis. It can achieve a diagnostic accuracy rate approaching or exceeding 95%, effectively identifying faults that are difficult to detect using traditional methods and significantly improving diagnostic accuracy and comprehensiveness. This application embodiment can automatically complete the entire process of monitoring, analysis, matching, and location in real time. Through fuzzy matching algorithms and a fault mode library, it can directly output the topology path coordinates of the faulty device after successful matching and generate a structured diagnostic report containing specific location, type, and severity. This achieves accurate and rapid fault location, guiding maintenance personnel directly to the problem point, reducing fault location time by an average of over 80%, greatly reducing the blind troubleshooting work of maintenance personnel, and effectively improving the efficiency of fault location and troubleshooting. This application embodiment utilizes a high-speed data acquisition unit and a sliding window monitoring algorithm to achieve seamless, high-density sampling and analysis of data streams. It solves the problem of fault omission caused by slow response speeds or low sampling rates in traditional methods. It can effectively capture, record, and analyze transient faults lasting only nanoseconds to milliseconds, as well as randomly occurring intermittent faults, achieving a breakthrough in the capture and analysis capabilities for transient and intermittent faults. This application embodiment introduces a recurrence probability assessment model to predict the future occurrence probability of intermittent faults. Based on the probability exceeding a threshold, a preventative maintenance plan is automatically generated. Combined with a confidence accumulation decision mechanism, false alarms are effectively filtered, improving the reliability of alarms. This transforms the maintenance strategy from post-fault repair to pre-fault prevention, achieving an intelligent upgrade from passive response to proactive prevention. It not only diagnoses existing faults but also predicts potential risks, guiding preventative maintenance. The communication interface unit of the diagnostic module in this application embodiment supports multiple mainstream industrial protocols, facilitating integration with existing systems. Modular hardware and software design makes functional units relatively independent, which facilitates upgrades or customization according to needs, enhances the system's life cycle and applicability, can flexibly adapt to different industrial sites, reduces deployment and modification difficulty, and makes the system highly versatile and easy to integrate and expand.

[0048] See attached document Figure 2 As shown, the intelligent diagnostic method for a fieldbus DP network is implemented based on the above-mentioned intelligent diagnostic system for a fieldbus DP network, and includes the following steps 110-150.

[0049] Step 110: The sliding window algorithm is used to continuously monitor the data flow of the fieldbus DP network.

[0050] Step 110 can be understood as the real-time monitoring stage.

[0051] Among them, the data frame transmission status and the signal quality parameters of the physical layer.

[0052] In the real-time monitoring phase, this application employs an innovative sliding window algorithm to continuously monitor the data stream. Preferably, the sliding window algorithm has a window size of 100ms and a sliding step size of 10ms. This parameter setting ensures both real-time monitoring and sufficient analytical accuracy, enabling the capture of transient faults lasting no less than 10ms.

[0053] Step 120: Based on the abnormal data identified during continuous monitoring, perform feature extraction and anomaly classification.

[0054] Step 120 can be understood as the feature extraction stage and the anomaly classification stage.

[0055] In the feature extraction stage, this application utilizes a preset threshold and a pattern recognition algorithm to identify abnormal patterns. For example, the preset threshold is set based on statistical analysis of historical fault data and employs a dynamic adjustment mechanism to adapt to different working environments and equipment states. For example, the pattern recognition algorithm uses machine learning technology to establish an accurate fault identification model by training a large number of fault samples.

[0056] The feature extraction algorithm in this application is implemented collaboratively on FPGA and ARM processor, balancing real-time performance and flexibility.

[0057] For example, the feature extraction algorithm of this application embodiment first constructs a four-stage pipeline processing unit in the FPGA, namely: the first stage is a signal conditioning unit, which implements configurable gain amplification and anti-aliasing filtering. The gain control register is dynamically configured by the ARM through the AXI-Lite interface. The filter adopts a half-band filter cascade structure. The first stage has a decimation factor of 2, passband ripple <0.01dB, and stopband attenuation >60dB, effectively reducing the clock frequency of subsequent processing; the second stage is a trigger detection unit, which implements two trigger modes: edge triggering and window function energy triggering. Edge triggering can be configured with a hysteresis comparator to prevent noise-induced false triggering. The hysteresis width is set to 1.5 times the noise amplitude. Window function energy triggering calculates the signal energy within the sliding window. The window length is configured to 256 sampling points. When the energy change rate exceeds a preset threshold (default 30%), a trigger mark is generated. The trigger position is accurately recorded to a 32-bit timestamp counter with a resolution of 10ns; the third stage is a feature pre-extraction unit, which completes the primary feature calculation before the data stream enters the DDR3 cache, and uses Xilinx FIR. The IP core performs mean and root mean square calculations; the CORDIC algorithm IP core performs amplitude-phase demodulation, and the pre-extracted feature vector (8-dimensional) is stored in memory along with the original data for in-depth analysis on the ARM side, reducing data transmission volume by 60%; the fourth level is a data encapsulation unit, which encapsulates the original sampled data and pre-extracted features according to a custom frame format. The frame header includes a synchronization word (0x5A5A5A5A), frame length, trigger type, timestamp, and CRC16 checksum to ensure data integrity. The encapsulated data is transmitted to the ARM memory via AXI-DMA, with the transfer block size set to 4KB to match the Linux memory page size.

[0058] The feature extraction algorithm in this application embodiment further includes deep feature extraction on the ARM side, based on a Linux user-space feature extraction daemon running on the ARM side, with a priority set to real-time SCHED_FIFO 96. For wavelet transform, when performing multi-resolution analysis, the wavelet basis selection needs to balance time-frequency localization and computational complexity. This application embodiment uses Design of Experiments (DOE) to determine the optimal wavelet basis: 18 wavelet bases (db2-db10, sym4-sym8, coif3-coif5) are tested on 10 typical fault waveforms, with evaluation metrics being reconstruction signal-to-noise ratio and feature separability (inter-class distance / intra-class distance). Experimental results show that the db4 wavelet has the best overall score for 90% of fault types, and is therefore selected as the default wavelet basis. The number of decomposition layers is not fixed at 5, but dynamically adjusted according to the effective bandwidth of the signal. The DP bus signal energy is mainly concentrated below 20MHz. According to the Nyquist sampling theorem, when the sampling rate is 100MS / s, the maximum effective number of layers is... Therefore, the actual number of decomposition layers is limited to 2-3 layers to avoid unnecessary computation. Wavelet thresholding denoising uses an improved VisuShrink threshold. Where σ is the noise standard deviation, N is the signal length, and α is the adaptive coefficient. α is dynamically adjusted based on the local signal-to-noise ratio (SNR): when SNR < 10 dB, α = 0.8, enhancing denoising; when SNR > 20 dB, α = 1.2, preserving more details. The noise standard deviation σ is estimated using the absolute median (MAD) of the first-order wavelet coefficients. The Fourier transform employs overlapping segmentation to overcome the inter-frame feature loss problem of FFT. Each frame has a data length of 1024 points, with a frame overlap rate of 50%, meaning 512 points are slid each time. The selection of the window function needs to consider the trade-off between spectral leakage and main lobe width. In this application, the window function is adaptively selected for different fault types: for narrowband interference requiring high frequency resolution, the Blackman-Harris window (sidelobe attenuation > 92dB) is used; for transient pulses requiring high time resolution, the Harris window (β=5.0, narrow main lobe) is used. The window function coefficients are pre-calculated and stored in a lookup table to avoid real-time computation overhead. Frequency domain feature extraction includes not only the conventional power spectrum but also introduces bispectral features to detect nonlinear phase coupling phenomena, which is particularly effective in identifying nonlinear distortion caused by connector oxidation. The bispectrum is estimated indirectly, with a segment length of 256 points, a segment overlap rate of 75%, and a Parzen window used for smoothing.

[0059] The third feature extraction algorithm in this embodiment includes feature fusion and dimensionality reduction. Initially, the fused feature dimension can reach over 50 dimensions, which directly inputs into the classifier would lead to the curse of dimensionality. This embodiment employs a two-stage dimensionality reduction strategy: the first stage uses mutual information for feature selection, retaining features with mutual information > 0.3 with the fault category, typically reducing the dimension to 20-25 dimensions; the second stage uses Linear Discriminant Analysis (LDA), projecting to a low-dimensional subspace (c1 dimension, where c is the number of fault categories) with the highest class discriminative power, ultimately controlling the feature vector dimension to 12-15 dimensions.

[0060] Furthermore, to overcome the limitations of univariate mutual information filtering, this application proposes a physical prior-guided Bayesian network feature selection method. This method first constructs a feature-dependent prior model based on the physical topology and signal transmission mechanism of the DP bus network: physical parameters such as bus terminal impedance, cable length, and signal reflection coefficient are used as prior knowledge and transformed into structural constraints of the Bayesian network through expert experience. Second, the K2 algorithm is used in conjunction with prior knowledge to learn the network structure, where nodes represent feature variables and edges represent conditional dependencies between features. Then, the Markov blanket of each feature is calculated based on the network structure, retaining the subset of features directly related to the target variable (fault type). Finally, a feature co-importance metric is introduced, defined as... Here, H represents conditional entropy, used to evaluate the collaborative discriminative ability among features. In specific implementation, the system first loads the device physical model parameters to generate an initial Bayesian network structure; then, it uses historical fault data for parameter learning and employs maximum a posteriori estimation (MAP) to optimize the conditional probability table; in the feature selection stage, features with high Markov blanket weights and high collaborative importance are prioritized, forming 12-15 dimensional optimized feature vectors. For real-time performance assurance, the feature extraction algorithm is vectorized on the ARM NEON coprocessor. For wavelet transform convolution operations, the NEON intrinsics instruction set is used to compute four sets of 32-bit floating-point data in parallel, improving single-cycle throughput by 3.8 times. The FFT algorithm uses the arm_cfft_f32 function optimized by the ARM Performance Library, with a 1024-point complex FFT taking only 58μs. Here, NEON is the SIMD (Single Instruction Multiple Data) extended instruction set module built into the ARM Cortex-A series processors, and Intrinsics is a set of built-in C / C++ function interfaces. ARM PerformanceLibrary is the official high-performance mathematical function library released by ARM, containing optimized functions such as FFT, matrix operations, and filtering. The `arm_cfft_f32` function is a dedicated interface within this library for "single-precision floating-point complex FFT (Fast Fourier Transform)". "c" represents complex, "fft" refers to Fast Fourier Transform, and "f32" indicates 32-bit floating-point data.

[0061] For example, the K2 algorithm, which combines prior knowledge with network structure learning, includes the following steps: The prior knowledge encoding steps include: transforming expert knowledge into node order constraints: if an expert believes that A may lead to B, then ensure that A is ranked before B in the ordering; and constructing a prior graph. It includes edges that must exist (such as "power failure → communication interruption") and edges that are prohibited from existing.

[0062] The improved K2 algorithm steps include: initialization, which is to determine the node order π based on prior knowledge, so that the parent node precedes the child node; for each node... a) Initialize the parent node set b) Evaluate candidate parent node Y (satisfying that Y is located in π). Previously and ( (a) Forbidden edge set), c) Calculate the score increment: Using BDeu rating: ,in, This represents the i-th feature node (or random variable) currently being evaluated, corresponding to the fault feature parameters in the patent text; Let X{i} represent the set of parent nodes of node X{i}, that is, the set of parent variables connected by directed edges that directly point to X{i} in the Bayesian network; Y represents candidate parent nodes, as those that may be added. The candidate nodes are scored and tested; π represents the node topology order (node ​​order constraint) determined by prior knowledge to ensure that the causal relationship direction conforms to the physical prior; α represents the equivalent sample size parameter in BDeu scoring, which is set to 10 here to smooth the probability estimation. This represents the number of possible states (the number of values ​​a discrete variable can take) of the j-th node (variable) X{i}. Represents the set of parent nodes The total number of all parent node state combinations, i.e., the dimension of the parent node joint state space; This represents the number of samples in the dataset that satisfy the combination of the i-th node taking the k-th state and its parent node taking the j-th state (configuration of the j-th parent node); α=10 is the equivalent sample size. for Number of states Given the number of parent node combinations, d) select Y that maximizes Δscore to add. Repeat until Δscore < threshold (set to 0.01); verify whether the learning results meet the requirements. Edges that violate prior knowledge will be manually reviewed.

[0063] The parameter learning steps include: using maximum a posteriori estimation (MAP). The prior distribution uses the Dirichlet distribution, with parameters... Posterior distribution parameters: .

[0064] For example, the Markov Blanket is the smallest set of features in a Bayesian network that is directly related to the target variable, and its calculation formula is as follows: Markov blanket composition ; Where Spouses(X) represents The specific calculation steps for the other parent nodes of X's child nodes include: determining the set of parent nodes of X. ;Sure child node set For each child node C∈ Determine its parent node set Parents(C); calculate the spouse node set: The result of merging these elements is a Markov blanket: .

[0065] The Markov blanket weight calculation includes: calculating the Markov blanket weight for each feature Y in MB(X). Where I represents conditional mutual information, the calculation formula is: The larger the conditional mutual information value, the greater the contribution of the feature to the discrimination of the target variable. Feature selection includes: retaining features with MBW(Y)>0.15; and sorting the retained features in descending order of MBW to form the final feature subset. The feature selection process in this embodiment adopts a three-level progressive selection strategy, specifically including the following first to third levels. The first level is a preliminary selection based on mutual information, that is, calculating the mutual information between each feature and the fault category: The first level retains features with mutual information values ​​I(X;Y) > 0.3, reducing the feature dimension from 50+ to 20-25. The second level is a physical prior screening based on Markov blankets, that is, calculating the Markov blanket weight MBW(Y) for each feature, retaining features with MBW(Y) > 0.15, and further reducing the dimension to 15-18. This level of screening incorporates prior knowledge of the physical characteristics of the DP bus to ensure that the screening results conform to engineering practice. The third level is a final screening based on co-importance, that is, calculating the co-importance of features: In this context, X represents the feature variable (such as fault features extracted from signal amplitude, spectral energy, etc.); Y represents the target variable, i.e., the fault category label (such as short-circuit fault, open-circuit fault, interference fault, etc.); I(X;Y) represents the mutual information between feature X and fault category Y, quantifying the degree of statistical dependence between the two; T represents the target variable (fault type), with the same meaning as Y, and is the standard notation in conditional entropy calculation; S represents the selected feature subset, i.e., the set of features retained after the previous stage screening; MBW(Y) represents the Markov blanket weight, measuring the contribution of feature Y to the conditional mutual information of the target variable in the Markov blanket; CI(Y) represents the feature co-importance, defined as the reduction in uncertainty of the target variable after adding feature Y; H(Y|T) represents the conditional entropy, i.e., the residual uncertainty of Y under the known variable T; H(Y|Z,T) represents the conditional entropy of Y under the known selected feature subset S and the target T; H represents conditional entropy, T is the target variable (fault type), and S is the selected feature subset. Features are selected in descending order of CI(Y) until the cumulative CI reaches 90% of the total CI, ultimately forming an optimized feature vector of 12-15 dimensions. The three-level selection process is progressive rather than replacement-based; each level refines the optimization based on the results of the previous level. The final selection result retains features highly correlated with the fault category, considers the collaborative discriminative effect between features, and conforms to the physical characteristics of the DP bus, ensuring the discriminative power and engineering practicality of the feature vector.

[0066] Step 130: Based on the preset fault mode library, perform fault matching on the abnormal data after feature extraction and anomaly classification.

[0067] Steps 130-140 can be understood as the fault matching stage.

[0068] The fault mode library includes multi-layered fault modes at the physical layer, data link layer, and application layer.

[0069] The fault mode library in this application is constructed in a layered, incremental, and verifiable manner to ensure the systematic nature and timeliness of the knowledge.

[0070] In some implementations, building a fault mode library first includes the acquisition and preprocessing of multi-source heterogeneous data.

[0071] In some implementations, the sources for multi-source heterogeneous data acquisition are divided into three categories: historical work order data, online monitoring data, and expert experience data. For historical work order data, fault records from the past three years are extracted from the CMMS system using ETL tools, with key fields such as equipment number, fault description, handling measures, and repair time extracted. Preprocessing employs natural language processing technology, integrating knowledge graph and causal reasoning techniques to perform entity recognition on unstructured text. In practical implementation, the system first constructs a professional ontology library for industrial control, containing core concepts such as equipment components, fault phenomena, and causal relationships. Approximately 500 entity categories and 1200 relationship types are extracted through expert interviews and standard documents. Second, a BERT-BiLSTM-CRF model is used for basic entity recognition, but a counterfactual reasoning mechanism is introduced: for identified potential fault entities, the system generates a counterfactual hypothesis, "How would the system behavior change if the fault did not exist?" This hypothesis is verified by querying historical data or simulation models, thus distinguishing between real faults and occasional anomalies. Third, a context-aware entity disambiguation algorithm based on knowledge graphs is used, employing a graph attention network (GAT) to calculate the semantic correlation between entities, resolving the problem of synonymy, such as distinguishing whether "communication interruption" refers to a network-wide interruption or a single node interruption. Finally, a causal discovery algorithm is introduced, using the PC algorithm and LiNGAM model to mine implicit causal relationships from fault description text and construct fault propagation paths.

[0072] For example, the specific implementation of calculating the semantic association degree between entities using Graph Attention Network (GAT) in this application embodiment includes the following sub-steps 111-114.

[0073] Sub-step 111, constructing a graph structure, includes: using entities in the fault description text as graph nodes, and co-occurrence relationships between entities as edges; entity feature vectors are generated through the BERT model with a dimension of 768.

[0074] Sub-step 112, attention mechanism calculation, includes: for node i and its neighbor node j, calculating the attention coefficient:

[0075] in, The original feature vector of entity i; W represents the shared weight matrix (128×768) used for feature linear transformation, a is the attention mechanism parameter vector (256×1), || represents the vector concatenation operation, LeakyReLU is the activation function with a slope of 0.

[0076] Sub-step 113, correlation calculation, includes: the update of node i is represented as: The semantic association between entities i and j is defined as follows: ; in, Let i represent the representation vector of node i after being updated by the graph attention network; This represents the attention weight coefficient of node i to its neighbor node j.

[0077] Sub-step 114, threshold judgment: when sim(i,j)>0.75, entities i and j are determined to have a strong semantic relationship, where sim(i,j) represents the cosine similarity between entities i and j; when 0.5 ≤ sim(i,j) ≤ 0.75, a moderate semantic relationship is determined; when sim(i,j)<0.5, the relationship is determined to be weak and is not considered.

[0078] For example, the PC algorithm is indeed a standard algorithm in the field of causal discovery (proposed by Peter Spirtes and Clark Glymour). This application embodiment combines it with the LiNGAM model to mine causal relationships. Specifically, it includes the following sub-steps 121-123.

[0079] Sub-step 121, PC algorithm application, includes: constructing a fully connected graph, with nodes representing fault characteristic variables; gradually removing irrelevant edges through conditional independence testing (using the G² statistic, where G represents the causal graph structure; significance level set to 0.05); determining the direction of undirected edges, applying the V-structure rule (if X and Y are not directly connected but are both connected to Z, and Z is not in the conditional independence set of X and Y, then X→Z←Y); propagating directionality to complete the entire causal graph construction.

[0080] Sub-step 122, LiNGAM model application, includes: establishing a linear non-Gaussian model based on the causal structure obtained by the PC algorithm: x = Bx + e; where B is a strict lower triangular connection matrix, representing the strength of the linear causal relationship between variables; x is the vector of observed variables; e is the vector of non-Gaussian noise; the B matrix is ​​estimated using the ICA (Independent Component Analysis) algorithm, and the causal direction is verified by the non-Gaussianity test (using negative entropy measurement).

[0081] Sub-step 123, causal relationship confirmation, includes: retaining only the causal relationships confirmed by both the PC algorithm and the LiNGAM model; Calculate causal strength: ; in, Let represent the element in the i-th row and j-th column of matrix B, and let represent the causal effect strength of variable j on variable i. This represents the p-value for statistical significance testing. A valid causal relationship is confirmed when causal_strength > 0.6.

[0082] For example, the processing and analysis flow after constructing the fault propagation path includes the following sub-steps 131-134: Sub-step 131, identification of key path nodes, includes: calculating the betweenness centrality of each node: , where BC(v) represents the betweenness centrality of node v, which measures the criticality of the node in the shortest path; This represents the total number of shortest paths from node s to node t; This represents the number of shortest paths passing through v. Nodes in the top 20% of those with betweenness centrality are considered critical nodes.

[0083] Sub-step 132, Fault Impact Scope Assessment, includes: performing a breadth-first search (BFS) starting from the fault source; calculating the impact coefficient representation. ,in, Let be the distance from node i to the fault source. is the node importance weight; IC represents the impact coefficient, which quantifies the impact range of the fault source. When IC>0.8, it is judged as a major impact fault.

[0084] Sub-step 133, root cause localization, employs Bayesian backward reasoning, including: ; Where P(rootcause|symptoms) represents the posterior probability of the root cause in Bayesian backward inference; P(symptoms|rootcause) represents the likelihood probability, the probability of the current symptom occurring given the root cause; k represents the safety factor, with a value of 1.5, used to calculate the preventive maintenance cycle. Nodes with a wide impact and low probability of occurrence are given priority as root causes.

[0085] Sub-step 134, maintenance strategy generation, includes: implementing preventive maintenance on critical nodes, with a maintenance cycle of T = k / λ, where λ is the failure rate and k is the safety factor (set to 1.5); deploying signal enhancement equipment or adding redundant channels for propagation paths with high impact factors; highlighting critical nodes and the scope of impact in the diagnostic report to guide maintenance personnel to prioritize their handling.

[0086] For example, the online monitoring data comes from the measured waveforms of this diagnostic module. During implementation, monitoring nodes are deployed at 10 typical industrial sites, covering environments such as thermal power plants, chemical plants, and steel mills. Each node collects at least three months of continuous data, totaling approximately 5TB. The raw waveforms are stored in segments, with one HDF5 file stored every 10 minutes. The filename includes a timestamp and device ID. The preprocessing stage performs three steps: deduplication (based on waveform similarity hashing), noise reduction (wavelet soft thresholding), and annotation (field engineers mark the start and end times of the fault).

[0087] Expert experience data was obtained through semi-structured interviews. A standardized questionnaire was designed, containing 30 fault scenario descriptions, and 15 senior engineers were invited to quantitatively rate the fault characteristics (using a 1-5 Likert scale). The Delphi method was used for two rounds of consultation, with a convergence consistency threshold set at 0.8 to ensure consistency in expert perception.

[0088] In some implementations, building a failure mode library further includes hierarchical modeling of failure modes.

[0089] The hierarchical fault mode modeling employs an ontology-based approach to construct a fault mode library. The top-level ontology defines five core concepts: fault phenomenon, fault cause, fault location, fault characteristics, and handling measures. Each concept contains multiple attribute slots; for example, the "fault characteristics" concept includes time-domain characteristic slots, frequency-domain characteristic slots, and time-frequency domain characteristic slots. In physical layer fault mode modeling, signal attenuation is subdivided into gradual attenuation (cable aging) and abrupt attenuation (connector oxidation). The former is characterized by an amplitude drop rate <0.5dB / cycle, while the latter is characterized by an amplitude step drop >3dB. Noise interference modes are categorized by spectral shape into broadband noise (spectral flatness <3dB) and narrowband interference (peak factor >10dB). Reflection modes are differentiated by delay time into near-end reflection (<50ns) and far-end reflection (>100ns), corresponding to impedance mismatches at different locations. The data link layer fault mode introduces the concept of the "Error Burst Index" (EBI). ,in Number of error frames Total number of frames For the length of the error burst, The error rate component reflects the degree of communication quality degradation. An EBI > 0.1 indicates a persistent fault, while an EBI < 0.01 indicates a non-persistent fault. >5 is classified as an intermittent, sudden failure and requires differentiated handling. Application layer failure modes primarily focus on modeling parameter configuration errors. Configuration errors are categorized into static errors (GSD file version mismatch) and dynamic errors (runtime parameters being illegally modified). Static errors are identified through MD5 checksum verification of the GSD file, while dynamic errors are traced by monitoring the timestamps of parameter write operations and the source IP address of the operation.

[0090] In some implementations, the third step in building a fault mode library includes feature parameter standardization and vectorization.

[0091] For example, all feature parameters are normalized to the [0,1] interval. For physical layer amplitude features, max-min normalization is used. , where x represents the original feature parameter value (the unnormalized measured value of the physical layer amplitude); x' represents the normalized feature value, mapped to the interval [0,1]; This indicates the upper limit of the signal level (or the maximum observation value of the dataset) as specified by the PROFIBUS standard. t represents the lower limit of the signal level as specified by the PROFIBUS standard (or the minimum observation value of the dataset); t represents the original time parameter (such as fault duration, signal rise time, etc.); t' represents the log-normalized time characteristic value. The reference time constant (e.g., 10ns or 1ms, depending on the signal type) is used for normalizing the time scale. and Signal level extremes are set according to the PROFIBUS standard. For time-dependent characteristics, logarithmic normalization is used. ,in The baseline time is used to accommodate temporal features ranging from nanoseconds to milliseconds. Vectorization employs sparse coding techniques. Each fault mode is represented as a 128-dimensional sparse vector, with non-zero dimensions corresponding to salient features. Sparsity is controlled between 15% and 25%, preserving discriminability while reducing computational load. Vector calibration utilizes semi-supervised learning, co-training with 10% labeled samples and 90% unlabeled samples, iterating until the vector space cosine similarity stabilizes (change between adjacent iterations < 0.001).

[0092] Step 140: Use a fuzzy matching algorithm to calculate the similarity between the current features of the abnormal data and the patterns in the fault mode library, and calculate the matching result between the current features of the abnormal data and the fault mode library.

[0093] Specifically, if the similarity exceeds a preset threshold, it is determined to be a successful match, and the corresponding fault location information is output; if the similarity does not exceed the preset threshold, it is determined that there is no matching fault mode in the current monitoring window.

[0094] In the fault matching stage, this application employs an innovative fuzzy matching algorithm. This algorithm uses a membership function to calculate fault similarity, with a similarity threshold set at 0.85. When the similarity exceeds the threshold, the system determines that a match is successful. This fuzzy matching method can effectively handle the variability and uncertainty of fault characteristics, improving the accuracy and robustness of the matching.

[0095] Step 150: Based on the matching result, generate a structured fault diagnosis report and provide maintenance recommendations.

[0096] Step 150 can also be understood as the report generation stage.

[0097] For example, the fault diagnosis report may include at least the fault timestamp, fault type, fault location, severity, and confidence index.

[0098] For example, the maintenance recommendations are generated based on statistical analysis of historical fault data and a rules engine, and cover equipment adjustment, parameter optimization, network maintenance, and preventative maintenance measures.

[0099] In some implementations, generating a structured fault diagnosis report includes: generating the structured fault diagnosis report based on a hierarchical early warning mechanism, wherein the structured fault diagnosis report is divided into multiple levels, and different levels correspond to different processing priorities.

[0100] In some implementations, the method further includes: identifying intermittent faults in the matching results, assessing the recurrence probability of the intermittent faults by fusing a time-series prediction model of historical fault event sequences with a physical model of equipment degradation, and generating a preventative maintenance plan.

[0101] In the report generation stage of this application, a detailed fault diagnosis report is generated and maintenance recommendations are provided. Preferably, the generated fault diagnosis report adopts a graded early warning mechanism, classifying the severity of the fault into four levels: Level 1 fault (minor), Level 2 fault (moderate), Level 3 fault (serious), and Level 4 fault (urgent). Different levels correspond to different processing priorities and maintenance strategies, helping maintenance personnel to rationally arrange maintenance work. The maintenance recommendations are generated based on historical fault data statistical analysis and an expert knowledge base, including equipment replacement recommendations, parameter adjustment recommendations, network optimization recommendations, and preventive maintenance plans.

[0102] The fault matching algorithm in this application seeks a balance between accuracy and real-time performance, and adopts a hierarchical matching and confidence accumulation mechanism.

[0103] This application embodiment is based on a complete closed-loop processing mechanism to ensure that each link is closely connected and the logic is clear. For example, the algorithm executes the following steps in the actual diagnostic process: First, during system initialization, a pre-trained fault mode library and membership function parameters are loaded, and the sliding window size is set to 100ms with a step size of 10ms. When data flow is detected, the feature extraction module is triggered for real-time processing, outputting a 24-dimensional feature vector. Then, a coarse matching stage is entered, where the KD-Tree is used to quickly retrieve the 5-10 candidate fault modes with the highest similarity in the 12-dimensional space after LDA dimensionality reduction. Next, a fine matching stage is performed, calculating the weighted Manhattan distance for the candidate modes and introducing a feature correlation penalty mechanism to obtain an accurate similarity value. Then, the system will... The matching results of the previous window are fused with historical window data: if the pattern is detected for the first time, the confidence accumulation counter is initialized; if the matching results of consecutive windows are consistent, the confidence accumulation value is updated and the standard deviation condition is checked; simultaneously, for intermittent faults, the DS evidence theory fusion mechanism is activated to calculate the basic probability allocation and update the confidence; when the confirmation condition is met (three consecutive window matches and the standard deviation < 0.05), the fault confirmation process is triggered, and a diagnostic report containing five mandatory fields is generated; if a multi-pattern matching situation is detected, the composite fault analysis module is activated to determine whether it is a composite fault through the conflict coefficient, and a diagnostic result of multiple faults coexisting is generated. The following are the implementation details: 1) Membership function parameter calibration, including: Gaussian membership function The parameters (c, σ) are not statically set, but dynamically adjusted through online learning. The initial parameter c is taken as the mean of the feature in the fault mode library, and σ is taken as the standard deviation, where x represents the input feature value (such as measured parameters such as signal amplitude and frequency). represents the Gaussian membership function value, indicating the degree to which the feature value x belongs to a certain fault mode; c represents the center value (mean) of the membership function, corresponding to the standard value of that feature in the fault mode library; The standard deviation (width parameter) of the membership function reflects the fluctuation range of the fault mode characteristics. After the system is running, each successful fault diagnosis incorporates the measured characteristic value into the statistical sample of the mode, updates it using a sliding window (window length 100), and recalculates c and σ. To prevent outlier interference, the Huber loss function is used instead of the mean square error, with the threshold δ set at 1.345σ. Samples exceeding δ have their weights reduced to 1 / δ. For non-Gaussian distribution characteristics (such as skewed return loss), a piecewise Gaussian model or trapezoidal membership function is used. Specifically, the feature distribution is fitted using kernel density estimation (KDE), with a horizontal cutoff at 30% of the peak density. The cutoff interval is approximated by a Gaussian function, and the outer interval is approximated by a linear function to ensure that the membership function shape closely matches the true distribution. 2) Dynamic adjustment of weight coefficients, including: feature weights. The initial values ​​are determined using the Gini impurity of the random forest. , where p i C represents the proportion of samples of class i in the node, and C is the total number of classes. During online execution, an adaptive boosting approach is used to dynamically adjust the weights: if a feature continuously causes misclassification in the first N matches (N defaults to 3), its weight decreases by 20%; if it continuously contributes to correct matches (N defaults to 5), the weight increases by 10%, with an upper limit of twice the initial value. After weight adjustment, renormalization is required to ensure... ;in, The weight coefficient of the i-th feature is determined by the Gini importance of the random forest; Gini represents the Gini impurity, which measures the impurity of a node; p irepresents the proportion (probability) of the i-th type of fault samples in the dataset; C represents the total number of fault categories. Different weighting coefficients can be set for different industrial scenarios: thermal power plant scenarios focus on temperature and electromagnetic compatibility related features, with temperature-related features amplified by 1.5 times in the weighting coefficient matrix; chemical plant scenarios focus on corrosion and grounding features, with corresponding feature weights amplified by 1.8 times; the scenario mode is specified by the user through the configuration file, or it can be inferred by automatically identifying MAC address prefixes and equipment manufacturer information; 3) Multi-scale matching implementation, including: coarse-scale matching is performed in the feature dimensionality reduction space (12 dimensions after LDA), using KD-Tree index to accelerate nearest neighbor search, with a search time complexity of O(logM), where K represents the conflict coefficient in DS evidence theory, measuring the degree of contradiction between different evidence sources; M represents the total number of modes in the fault mode library (approximately 300-500). The similarity threshold is set to 0.75 to filter out candidate mode sets (size controlled at 5-10); fine-scale matching is performed in the original feature space (24 dimensions), calculating precise similarity only for the candidate set. To avoid the curse of dimensionality, fine-scale matching uses weighted Manhattan distance, which is more computationally efficient than Euclidean distance and more robust to outliers. A feature correlation penalty term is introduced during distance calculation, meaning that if two highly correlated features (correlation coefficient > 0.8) deviate from their normal values ​​simultaneously, only the one with the larger deviation is included in the distance calculation, avoiding double penalties. 4) Confidence accumulation and decision fusion include: the confidence of a single match may be affected by noise; the implementation of this application uses time-accumulated decision fusion, meaning that if the same pattern is identified in three consecutive sliding windows (each window is 100ms, with a step size of 10ms, i.e., a time range of 270ms), and the similarity standard deviation is < 0.05, it is determined to be a confirmed fault; otherwise, it is considered a suspicious event, only recorded without alarm; for intermittent faults, cumulative evidence theory is used, with the confidence increased by 0.1 for each successful match, up to a maximum of [missing value]. 0.95; if five consecutive windows fail to match, the confidence level decreases by 0.05, and the fault state is cleared when it drops below 0.3; for cases where multiple modes match simultaneously (such as physical layer noise interference and data link layer frame errors often coexist), a fusion decision based on DS evidence theory is implemented; the identification result of each mode is regarded as an evidence body, and the basic probability allocation (BPA) is obtained by similarity normalization; conflicting evidence is synthesized through Yager rules, and when the conflict coefficient K>0.7, it is judged as a composite fault, and a diagnostic report of multiple faults coexisting is generated; 5) Parallel computing and real-time guarantee, including: on the multi-core ARM platform, the fault matching task is divided into 4 sub-tasks according to the slave address range, and each sub-task is bound to an independent core; the pattern library creates 4 read-only copies in memory and places them on different NUMA nodes to avoid cache consistency competition.After coarse-scale matching, the candidate pattern set is passed to each core via shared memory. Fine-scale matching is performed in parallel, and the results are finally aggregated by the main core. In actual tests on a 4-core Cortex-A9 platform, the matching speed is increased to 3.2 times that of a single core, and the latency is reduced from 8ms to 2.5ms. For computationally intensive bispectral feature matching, GPU acceleration is implemented (if an on-chip GPU is integrated). The bispectral matrix is ​​treated as a two-dimensional image, and template matching is computed in parallel using the OpenCL kernel. The workgroup size is set to 16×16, fully utilizing the GPU SIMD unit. Experiments show that the bispectral matching time is reduced from 45ms to 1.2ms, meeting real-time requirements.

[0104] For example, the specific implementation of the feature correlation penalty mechanism introduced in the weighted Manhattan distance calculation in this application includes: constructing a correlation matrix and calculating the Pearson correlation coefficient matrix R between features. Where cov is the covariance and σ is the standard deviation; when At that time, features i and j are determined to be highly correlated; a penalty mechanism is implemented so that for highly correlated feature pairs (i,j), the distance is not calculated repeatedly: Where d(X,Y) represents the weighted Manhattan distance between samples X and Y; Let be the weight of feature i, reflecting its importance in fault diagnosis; x{i} and y{i} are the values ​​of samples X and Y in the i-th feature dimension, respectively; α is the penalty coefficient (set to 0.3) used to control the penalty intensity of repeated calculation of related features; I(·) is the indicator function, which is 1 when the condition is met and 0 otherwise; for cases where multiple features are highly correlated, only the feature with the largest deviation is retained for distance calculation; dynamic adjustment mechanism: update the correlation coefficient matrix R every 1000 frames of data processed; recalculate the penalty term when the correlation coefficient changes by more than 0.2; for specific scenarios such as thermal power plants, a preset correlation matrix is ​​used to accelerate calculation.

[0105] For example, the analysis and processing flow after fusing the matching results with historical window data includes: ① Calculation of fusion results: the matching results of the current window t The results are then weighted and merged with the results from the previous N windows: ,in β represents the fusion matching result value at the current time t; β is the current window weight (set to 0.6); Historical fusion value; This represents the original matching similarity value for the current monitoring window t. When the matching results of consecutive windows are consistent, confidence accumulation is used: ,in, This represents the cumulative confidence level at time t. γ is the cumulative confidence level of the previous time step; γ is the cumulative coefficient (0.1), and θ is the threshold (0.85). ② Fusion result analysis: a) Stability analysis: Calculate the standard deviation σ of the fusion values ​​of the most recent 3 windows. When σ < 0.05, it is judged as a stable match; b) Trend analysis: Calculate the first difference of the fusion values. ,when And if it continues for 3 consecutive windows, it is determined to be a trend of worsening fault; c) Anomaly detection: when and When the difference exceeds 2σ(F), it is judged as an abnormal fluctuation, triggering a data quality check. ③ Decision output: The fault is confirmed when the following conditions are met: (a) (b) σ < 0.05; (c) For intermittent faults, when When the value is between 0.65 and 0.85 and the fluctuation period is stable, it is marked as a potential fault; when multi-mode matching is detected, the conflict coefficient is calculated. When K>0.7, it is determined to be a composite fault. ④ Adaptive adjustment: The sliding window parameters are dynamically adjusted according to the stability of the fusion result: for stable matching, the window step size is increased to 20ms to improve efficiency; for unstable matching, the window step size is decreased to 5ms to improve accuracy; for confirmed faults, the original data of the previous and next 5 windows are automatically saved for subsequent analysis. In order to verify the effectiveness and reliability of the embodiments of this application, a large number of laboratory tests and field tests were conducted. Among them, the laboratory tests mainly include functional tests, performance tests and reliability tests. Functional tests verified whether the various functions of the system meet the design requirements. The test content includes signal enhancement function, data acquisition function, fault diagnosis function, communication function, etc. The test results show that all functions can work normally and meet the design indicators. Performance tests verified whether the performance indicators of the system meet the design requirements. The test content includes diagnostic accuracy, response time, processing capacity, etc. The test results show that the diagnostic accuracy reaches more than 95%, the response time is less than 1 second, and the processing capacity meets the requirements of 12Mbps communication rate. The field test was conducted on the DP bus network of a thermal power plant. The test lasted for 6 months, covering different operating conditions and seasons. The tests included fault simulation tests, long-term stability tests, and actual fault diagnosis tests. Fault simulation tests artificially created various faults to verify the system's diagnostic capabilities. Test results showed that the system could accurately identify and locate various types of faults. Long-term stability tests verified the system's stability and reliability under long-term operating conditions. Test results showed that the system operated stably with a fault-free operating time exceeding 4000 hours. Actual fault diagnosis tests verified the system's diagnostic effectiveness for actual faults. During the tests, the system successfully diagnosed 15 actual faults with a 100% accuracy rate, and the average fault location time was reduced by more than 80%. Comparative tests compared the embodiments of this application with existing diagnostic equipment. The comparison included diagnostic accuracy, diagnostic efficiency, and ease of use. Test results showed that the embodiments of this application significantly outperformed existing equipment in all indicators. In particular, in terms of diagnostic accuracy, the embodiments of this application improved by 15 percentage points compared to the best existing equipment; and in terms of fault location time, the embodiments of this application reduced the time by 60% compared to the fastest existing equipment.

[0106] The following two specific application cases further illustrate the practical application effects and value of the embodiments of this application. Case 1: Intermittent fault diagnosis of the DP bus network of a 300MW thermal power plant. The DP bus network of this unit frequently experienced intermittent communication interruptions during operation, each lasting approximately 100ms, with an unpredictable frequency, averaging 2-3 times per week. Traditional diagnostic methods could not capture such transient faults, and maintenance personnel had repeatedly investigated without success. After using the diagnostic module of this application embodiment, the system successfully captured the fault signal, and analysis revealed that it was caused by voltage fluctuations due to the aging of the power supply module of a certain slave device. After replacing the power supply module, the fault was completely eliminated. This case fully demonstrates the unique advantages of the embodiments of this application in capturing and analyzing transient faults. Case 2: Diagnosis of communication quality degradation in the DP bus network of a chemical plant. The DP bus network of this chemical plant recently experienced communication quality degradation, manifested as increased data transmission error rate and prolonged response time. Using traditional diagnostic equipment, all indicators were found to be normal. In-depth analysis using the diagnostic module of this application embodiment revealed that electromagnetic interference generated by a newly added frequency converter on site was affecting the signal quality of the DP bus. The system not only accurately identified the interference source but also provided detailed interference spectrum analysis and suppression suggestions. After implementing shielding measures as suggested, communication quality returned to normal. This case demonstrates the diagnostic capabilities of the embodiments of this application in complex electromagnetic environments.

[0107] This application's embodiment of the intelligent diagnostic method for fieldbus DP networks achieves high-precision detection and location of instantaneous, sudden, and intermittent faults that are difficult to capture by traditional methods by synchronously monitoring physical layer signals and data link layer data, using wavelet and Fourier cascade analysis for deep feature extraction, and relying on a multi-level fault mode library for intelligent matching. At the same time, the confidence accumulation decision and recurrence probability assessment mechanism introduced significantly improves the reliability of diagnostic conclusions and promotes a fundamental shift in maintenance mode from passive response to proactive prediction and intelligent decision support.

[0108] See attached document Figure 3 As shown, the software architecture of this application embodiment adopts a layered modular design, with each layer communicating through clearly defined interfaces to ensure the maintainability and scalability of the system. The following, in conjunction with the appendix... Figure 3 The implementation details of each layer in the driver layer, system service layer (not shown in the figure), application layer, and user interface layer are described in detail.

[0109] The driver layer interacts directly with hardware units, and its stability directly affects the reliability of the entire system. This layer adopts a hybrid architecture combining kernel drivers and user-space libraries.

[0110] In some implementations, for high-speed A / D conversion modules, the driver is developed using the Linux kernel's IIO (Industrial I / O) sub-framework. Specifically, the ADS5400 register mapping must first be configured, the sampling rate set to 100MS / s, and a four-channel interleaved sampling mode enabled to improve data throughput. Internally, the driver implements a circular buffer management system, with a preferred buffer size of 64MB, employing a double-buffering mechanism to ensure continuous data acquisition. The interrupt handler performs only the minimum necessary operations—moving sampled data from the FIFO to the memory buffer—avoiding time-consuming operations in the interrupt context. The user-space library maps the kernel buffer to user space via the mmap() system call, achieving zero-copy data access and keeping latency below 50μs.

[0111] In some implementations, the FPGA processing module driver is based on the PCIe bus protocol. Specifically, the PCIe device driver is registered in the Linux kernel, the BAR0 space is configured for register access, and the BAR2 space is configured for DMA data transfer. DMA transfer uses the Scatter-Gather mode, supporting a maximum of 128 descriptor lists, with each descriptor carrying a 4KB data block. The driver implements an asynchronous notification mechanism; after the FPGA completes preprocessing of a frame of data (such as digital filtering and trigger flagging), it sends a notification to the host via an MSI-X interrupt. The user space can dynamically adjust the cutoff frequency (adjustable from 5MHz to 25MHz) and trigger threshold (adjustable from ±100mV to ±5V) of the FPGA's internal digital filter via the ioctl interface to adapt to different industrial environments.

[0112] In some implementations, the PROFIBUS-DP protocol stack uses the SPC3 protocol chip firmware for secondary development of the communication interface driver. The focus is on implementing differentiated processing for DP-V1 and DP-V2 message parsing: DP-V1 messages use fixed-period polling, while DP-V2 messages support aperiodic data exchange. The driver layer implements hardware offloading for message verification, utilizing the SPC3's CRC check unit to avoid the main CPU participating in redundant calculations. The RS485 interface driver uses GPIO multiplexing to achieve automatic switching of transmit and receive directions, with switching latency controlled within 1μs to prevent bus conflicts.

[0113] The system service layer is built on Linux with a real-time extension patch (PREEMPT_RT) to keep the scheduling latency of critical tasks within 100μs. Process management adopts a priority-based preemptive scheduling strategy, with data acquisition tasks assigned the highest priority (SCHED_FIFO, prio 98), followed by fault diagnosis tasks (prio 95), while non-real-time tasks such as report generation use ordinary time-sharing scheduling.

[0114] In some implementations, memory management employs a memory locking mechanism (mlockall) to prevent critical data from being swapped to disk. For example, 512MB of physical memory is reserved as a shared memory area ( / dev / shm) for the fault mode library and feature templates, accessible only in read-only mode by all diagnostic tasks to avoid duplicate loading. Exemplarily, the file system uses OverlayFS technology, setting the root file system as a read-only layer, while diagnostic logs and configuration files are stored in a writable upper layer to prevent system crashes due to file system corruption.

[0115] In some implementations, the network protocol stack is configured with the SO_PRIORITY socket option to set the EF (Expedited Forwarding) class of DiffServ for diagnostic report uploads and remote control messages, ensuring priority forwarding in industrial Ethernet. Simultaneously, a TCP keepalive mechanism is implemented with a heartbeat interval of 10 seconds. After three timeouts, the connection is disconnected and a local alarm is triggered, preventing remote monitoring failure due to a single point of failure in the network.

[0116] The application layer modules communicate with each other using a publish-subscribe (Pub-Sub) message bus, which is based on the ZeroMQ message queue to reduce module coupling.

[0117] In some implementations, the data acquisition module is configured with multiple asynchronous worker threads, including: a raw data receiving thread responsible for acquiring A / D sampling data from the driver layer; a data parsing thread responsible for parsing the DP frame structure and physical layer features; and a data archiving thread responsible for compressing the raw data (using the LZ4 algorithm, with a compression ratio of approximately 2:1) and storing it on an SSD. Data is transferred between these three threads via a lock-free circular queue with a length of 1024 slots. A backpressure mechanism ensures that memory usage does not exceed a preset limit (default 2GB).

[0118] In some implementations, the signal processing module performs pipelined processing, including: the first stage is DC bias removal and amplitude normalization, using moving average filtering with a window length of 1024 sampling points; the second stage is adaptive wavelet denoising, dynamically selecting a threshold based on the signal-to-noise ratio (using the SURE unbiased estimation method); the third stage is feature extraction, extracting 8-dimensional time-domain features (mean amplitude, variance, peak-to-peak value, rise time, etc.) and 16-dimensional frequency-domain features (power spectral density, harmonic distortion, etc.) from each frame of DP message to form a 24-dimensional original feature vector.

[0119] In terms of the specific implementation of the signal processing module, this application embodiment adds a preprocessing step based on physical mechanisms, which significantly improves the accuracy and robustness of signal feature extraction. The newly added preprocessing stage mainly includes two key steps: baseline drift correction and nonlinear distortion compensation. Baseline drift correction adopts an improved moving least squares (MLS) method, optimized for the low-frequency drift characteristics unique to DP bus signals: First, the original signal is segmented into 20ms segments, and a third-order polynomial is used to fit the trend term within each segment. The fitting weight function is designed as a Gaussian kernel, centered at the midpoint of the segment, with a standard deviation σ=5ms, to ensure a smooth transition of the trend term; Second, an adaptive window mechanism is introduced. When a signal abrupt change point is detected (identified by the Canny edge detection algorithm), the fitting window is automatically reduced to 5ms to avoid edge distortion; Finally, spline interpolation is used to splice the trend terms of each segment to eliminate the discontinuity caused by segmented processing. Nonlinear distortion compensation is based on the transmission characteristics of the DP bus physical layer, establishing a mapping model between signal amplitude and phase distortion: the bus terminal impedance characteristics are obtained through frequency sweep testing, and an S-parameter matrix is ​​constructed; the nonlinear system is described using Volterra series, and the second-order term is selected as the main compensation object; a parallel computing unit is implemented in the FPGA to calculate the compensation coefficient in real time.

[0120] It is worth noting that the Scattering Parameter Matrix is ​​a common technical term in radio frequency and microwave engineering, used to describe the signal transmission characteristics of multiport networks.

[0121] In the embodiments of this application, the S-parameter matrix can be understood based on the following aspects 1)-3): Aspect 1), for a two-wire transmission system like the DP bus, construct a 2×2 S-parameter matrix: ; in, This represents the reflection coefficient of port 1, reflecting the degree of signal reflection at port 1; This represents the transmission coefficient from port 1 to port 2, reflecting the forward transmission characteristics of the signal; This represents the transmission coefficient from port 2 to port 1, reflecting the reverse transmission characteristics of the signal; This represents the reflection coefficient of port 2, reflecting the degree of signal reflection at port 2.

[0122] Aspect 2), methods for measuring the S-parameter matrix include: performing frequency sweep tests using a vector network analyzer, measuring the S-parameters at each frequency point in 1MHz steps within the DP bus operating frequency band (typically 9.6kbps-12Mbps).

[0123] Aspect 3), the application purposes of the S-parameter matrix include: using the S-parameter matrix to analyze the impedance matching of bus terminals, identifying signal reflection problems caused by impedance mismatch, and providing a basis for nonlinear distortion compensation.

[0124] Among them, S-parameters are internationally recognized radio frequency engineering terms, which are clearly defined in IEEE standards and meet the requirements of sufficient patent disclosure.

[0125] In some implementations, the fault diagnosis module uses a two-level diagnostic engine: a fast screening engine and a precise diagnosis engine. The fast screening engine employs a lightweight decision tree (no more than 5 layers deep) to initially screen the data stream per second, filtering out over 95% of normal packets. The precise diagnosis engine uses a random forest model (200 classification trees) to perform in-depth analysis of suspicious packets. These two engines work together to reduce the single-frame diagnosis time from 12ms to less than 2ms while ensuring accuracy.

[0126] In some implementations, the report generation module uses a template engine (based on Jinja2) to generate structured reports. The report includes five mandatory fields: fault timestamp (accurate to 1 μs), fault type code (following the VDI 3827 standard), fault location coordinates (device topology path description), severity score (0-100 quantified value), and confidence index (calculated based on similarity). Maintenance recommendations are generated through a rule engine; the rule base contains over 200 IF-THEN rules, covering 80% of common fault scenarios.

[0127] It is worth noting that the core function of the SURE (Stein's Unbiased Risk Estimate) unbiased estimation method in this embodiment is to adaptively calculate the optimal threshold λ for wavelet denoising. Its purpose is not to directly calculate the threshold itself, but to achieve the optimal balance between noise suppression and signal fidelity by minimizing the theoretical risk function. Unlike traditional fixed or heuristic thresholds, the SURE method can dynamically estimate the mean square error risk under different thresholds based on the actual statistical characteristics of the current signal wavelet coefficients, thereby selecting the threshold that minimizes the deviation between the reconstructed signal and the true signal. This effectively solves the problem of insufficient denoising or over-smoothing caused by threshold mismatch in scenarios with drastic fluctuations in the signal-to-noise ratio of DP bus signals in industrial settings (10dB-30dB).

[0128] In some implementations, the implementation of SURE includes the following steps S1-S7: Step S1: Wavelet decomposition and coefficient acquisition, performing db4 wavelet basis J=3-level decomposition on the FPGA preprocessed signal segment X (length N=1024 points) to obtain the detail coefficients of each level. And approximation coefficients. Thresholding is performed only on the detail coefficients. Step S2: Noise variance σ estimation, using the first level of detail coefficients. The median absolute deviation (MAD) is used to estimate the noise standard deviation: Step S3: Threshold candidate set generation, within the dynamic range. Inside, with Generate candidate threshold sequences for step size This covers the threshold selection space from conservative to aggressive. Step S4: Calculate the risk function R(λ) for each candidate threshold. Calculate the unbiased risk estimate using the SURE formula: Where # represents the number of coefficients that meet the condition. The physical meaning of this formula is: the first term N is the total coefficient energy, the second term penalizes the number of coefficients set to zero, and the third term quantizes the compression bias of the retained coefficients. Step S5: Optimal threshold selection iterates through all... The threshold that minimizes the risk is selected as the final denoising threshold: Step S6: Apply the threshold function. Substitute the improved VisuShrink threshold function to perform denoising: ; in, Indicates the candidate threshold; The corresponding unbiased risk estimate (SURE value) is used to evaluate the theoretical risk of the denoising effect. N represents the length of the signal segment (number of sampling points), which is 1024 points here. This represents the k-th detail coefficient (wavelet coefficient) of the j-th level after wavelet transform; j represents the level index of the wavelet decomposition (j=1,2,3 corresponds to different frequency scales); k represents the index of the coefficients in each level. This represents the m-th candidate threshold; The optimal threshold is selected by minimizing the risk function; J represents the total number of wavelet decomposition levels; a and b represent the adjustment parameters in the improved threshold function, controlling the transition characteristics of the denoising curve. Step S7: Iterative optimization. To improve robustness, this embodiment implements sliding window SURE estimation: every 10 frames of signal (i.e., 1 second of data) processed, σ is recalculated and updated using the statistical characteristics of the wavelet coefficients of the most recent 10 frames. This allows the threshold to track environmental changes. If the signal-to-noise ratio (SNR) changes by more than 5 dB within the window, the threshold is immediately recalculated.

[0129] The web interface in the user interface layer adopts a front-end and back-end separation architecture.

[0130] In some implementations, the backend provides a RESTful API and WebSocket real-time push based on the FastAPI framework. The Cross-Origin Resource Sharing (CORS) configuration allows access only from specified IP ranges.

[0131] In some implementations, the front end uses the Vue.js 3.0 framework and receives real-time data via WebSocket, with the data update frequency set to 10Hz to avoid browser overload.

[0132] In some implementations, the interface features an adaptive layout that supports resolutions from 1280×720 to 4K.

[0133] The intelligent diagnostic method and system for fieldbus (DP) networks presented in this application have significant technical effects and practical application value. Firstly, regarding diagnostic accuracy, by combining physical layer signal enhancement with data link layer data frame analysis, comprehensive diagnostics from the physical layer to the data link layer are achieved, with a diagnostic accuracy rate exceeding 95%. This high accuracy is mainly attributed to innovative signal enhancement technology and a multi-level fault mode library design. The signal enhancement unit effectively improves signal quality, ensuring reliable diagnostic data under various operating conditions; the multi-level fault mode library comprehensively covers various fault types, avoiding missed diagnoses and misdiagnoses. Secondly, in terms of fault location efficiency, this application has achieved a breakthrough. Traditional fault diagnosis methods often require maintenance personnel to inspect each device in the network one by one, which is time-consuming, labor-intensive, and inefficient. This application, through intelligent fault matching and location algorithms, can shorten fault location time by more than 80%. When a fault occurs, the system can immediately identify the fault type and accurately locate the fault location, providing clear guidance to maintenance personnel and greatly reducing fault diagnosis time. Thirdly, in terms of instantaneous fault capture capability, this application has unique advantages. Traditional diagnostic methods often struggle to capture transient faults lasting only milliseconds. While these faults are short-lived, they can severely impact system operation. This application's embodiments, through high-speed data acquisition technology and an innovative sliding window algorithm, can accurately capture and analyze transient faults lasting at least 10 ns, providing strong technical support for preventing and resolving intermittent faults. Fourth, regarding intelligence, this application's embodiments achieve a shift from passive diagnosis to proactive prevention. The system can not only diagnose existing faults but also predict potential fault risks and provide preventative maintenance recommendations based on statistical analysis of historical fault data. This proactive maintenance mode can significantly reduce equipment failure rates, extend equipment lifespan, and improve overall system reliability. Fifth, regarding versatility and scalability, this application's embodiments exhibit excellent adaptability. The multi-interface design allows the diagnostic module to be easily integrated into various industrial control systems without requiring large-scale modifications to existing systems. The modular design also makes the system easy to maintain and upgrade, allowing for flexible configuration of functional modules according to different application needs.

[0134] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of this application, and this application is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of this application, and these modifications and improvements are also considered to represent the scope of protection of this application.

Claims

1. A fieldbus DP network intelligent diagnostic system, connected in the fieldbus communication link of an industrial control system, characterized in that, It includes a signal enhancement unit, a data acquisition unit, a data processing unit, and a communication interface unit, among which, The signal enhancement unit is configured with a differential amplifier circuit, and the signal enhancement unit includes an adaptive gain control module and a noise suppression module, which are used to adjust the strength of the transmitted signal of the physical layer to meet the target threshold. The data acquisition unit includes a high-speed analog-to-digital conversion module and a field-programmable gate array (FPGA) processing module, which are used to acquire the target signal of the physical layer and the data frame of the data link layer online. The data processing unit includes a multi-core processor and has a built-in fault mode library and feature extraction algorithm module, which is used to perform real-time analysis, fault detection and diagnosis on the collected target signal and data frame, and obtain the diagnosis results. The communication interface unit supports multiple industrial communication interface protocols and is used to forward the diagnostic processing results.

2. The intelligent diagnostic system for a fieldbus DP network according to claim 1, characterized in that, The diagnostic processing results include signal strength statistics, wherein, The adaptive gain control module implements multi-level gain adjustment based on a digital potentiometer, dynamically generates control logic based on the signal strength statistics, and controls the output signal amplitude to remain stable within a preset range through a closed-loop feedback mechanism.

3. The intelligent diagnostic system for a fieldbus DP network according to claim 1, characterized in that, The high-speed analog-to-digital conversion module adopts a multi-channel interleaved sampling mode; The field-programmable gate array (FPGA) processing module has a built-in first-in-first-out (FIFO) memory and adds a frame header containing timestamps, trigger types, and verification information to the acquired data before transmission.

4. The intelligent diagnostic system for a fieldbus DP network according to claim 1, characterized in that, The fault mode library includes physical layer fault modes, data link layer fault modes, and application layer fault modes. Each fault mode is represented as a multidimensional feature vector after feature extraction by the feature extraction algorithm module.

5. The intelligent diagnostic system for a fieldbus DP network according to claim 1, characterized in that, The feature extraction algorithm module employs a cascaded process of wavelet transform and Fourier transform to extract the time-domain, frequency-domain, and time-frequency-domain features of the fault signal.

6. A fieldbus DP network intelligent diagnostic method, implemented based on the fieldbus DP network intelligent diagnostic system according to any one of claims 1 to 5, characterized in that, include: A sliding window algorithm is used to continuously monitor the data flow of the fieldbus DP network. The sliding window has a preset window size and sliding step size, which is used to synchronously record the data frame transmission status of the data link layer and the signal quality parameters of the physical layer. Based on the abnormal data identified during continuous monitoring, feature extraction and anomaly classification are performed. Based on a preset fault mode library, fault matching is performed on the abnormal data after feature extraction and anomaly classification. The fault mode library includes multi-layer fault modes of the physical layer, data link layer and application layer. A fuzzy matching algorithm is used to calculate the similarity between the current features of the abnormal data and the patterns in the fault mode library, and the matching result between the current features of the abnormal data and the fault mode library is calculated. If the similarity exceeds a preset threshold, it is determined that the match is successful and the corresponding fault location information is output. If the similarity does not exceed the preset threshold, it is determined that there is no matching fault mode in the current monitoring window. Based on the matching results, a structured fault diagnosis report is generated and maintenance recommendations are provided.

7. The method according to claim 6, characterized in that, The process of feature extraction and anomaly classification based on abnormal data identified during continuous monitoring includes: Based on the cascaded processing of wavelet transform and Fourier transform, the abnormal data is analyzed and its features are extracted. Based on a recognition algorithm that uses a preset threshold and a preset anomaly pattern, the abnormal data that has undergone feature extraction processing is matched and classified with the preset anomaly pattern. Specifically, the wavelet transform performs multi-resolution analysis on the abnormal data based on a preset wavelet basis and denoises the abnormal data based on an improved threshold function for primary feature pre-extraction; the Fourier transform performs adaptive window function selection based on overlapping segmentation processing to extract the time-domain features, frequency-domain features, and time-frequency-domain features of the abnormal data for deep frequency-domain feature analysis.

8. The method according to claim 6, characterized in that, The fault diagnosis report includes at least the fault timestamp, fault type, fault location, severity, and confidence index; the maintenance recommendations are generated based on historical fault data statistical analysis and rule engine, and cover equipment adjustment, parameter optimization, network maintenance, and preventive maintenance measures.

9. The method according to claim 6, characterized in that, The generation of the structured fault diagnosis report includes: The structured fault diagnosis report is generated based on the hierarchical early warning mechanism. The structured fault diagnosis report is divided into multiple levels, and different levels correspond to different processing priorities.

10. The method according to claim 6, characterized in that, The method further includes: Intermittent faults in the matching results are identified, and the recurrence probability of the intermittent faults is assessed by fusing a time-series prediction model of historical fault event sequences with a physical model of equipment degradation, and a preventive maintenance plan is generated.

Images