Field bus PA network intelligent diagnosis method

Abstract

This invention provides an intelligent diagnostic method for PA (Power Amplifier) ​​networks, belonging to the field of industrial fieldbus technology. It can at least partially solve the problems of existing PA bus systems, such as difficulty in effectively distinguishing voltage interference from equipment faults, low diagnostic accuracy, and the inability to achieve online real-time diagnosis. The invention includes: multi-level isolation processing and data acquisition of PA bus signals; multi-level wavelet decomposition of the acquired digitized signal data to extract energy distribution vectors; inputting the energy distribution vectors into a support vector machine classifier and a neural network classifier respectively to obtain initial confidence scores; dynamically adjusting the fusion weights based on the signal-to-noise ratio to calculate the final decision confidence score; and based on the classification results, using an adaptive filter to suppress voltage interference and triggering time-domain reflection localization for equipment faults. This invention achieves online separation and diagnosis of PA bus interference and faults.

Description

Technical Field

[0001] This invention belongs to the field of industrial fieldbus technology, specifically relating to an intelligent diagnostic method for fieldbus PA networks. Background Technology

[0002] With the continuous improvement of industrial automation, fieldbus technology has been widely used in industrial control systems. PROFIBUS-PA, as a fieldbus protocol specifically designed for process automation, plays an important role in industries such as petrochemicals, power, and metallurgy. The PA bus adopts the IEC 61158-2 standard, transmitting digital signals and DC power simultaneously via a two-wire cable, providing power to field devices while enabling data communication. However, in practical applications, PA bus devices often face various electromagnetic interferences and voltage fluctuations, leading to unstable operation, degraded communication quality, and even malfunctions.

[0003] Currently, fault diagnosis in PA bus systems primarily relies on traditional diagnostic methods. In terms of efficiency, existing methods require skilled technicians to perform on-site manual testing using tools such as oscilloscopes and multimeters, which is time-consuming, labor-intensive, and inefficient. Regarding accuracy, existing methods struggle to effectively distinguish between voltage interference and equipment faults, frequently resulting in misdiagnosis and limited accuracy. In terms of real-time monitoring, existing methods cannot provide real-time monitoring and early warning, only offering reactive solutions after a fault occurs. Finally, in terms of intelligence, existing methods lack intelligent diagnostic analysis capabilities and cannot provide targeted maintenance recommendations.

[0004] While some diagnostic devices and methods for the PROFIBUS-DP bus exist in the current technology, these DP bus diagnostic techniques cannot be directly applied to PA bus systems due to significant differences between PA and DP buses at the physical and protocol layers. The PA bus uses Manchester encoding, operates on a 12-32V DC power supply, and has a communication rate of 31.25kbps, while the DP bus uses an RS485 interface, operates on a 5V differential signal, and can achieve a communication rate of up to 12Mbps. Therefore, the PA bus is more sensitive to voltage interference and noise signals, requiring specialized diagnostic techniques to address its unique technical challenges.

[0005] In recent years, with the development of artificial intelligence and big data technologies, some researchers have attempted to apply intelligent algorithms such as neural networks and fuzzy logic to fieldbus fault diagnosis, achieving some results. However, most of these technologies remain in the theoretical research stage, lacking practical diagnostic methods specifically tailored to the characteristics of the PA bus. Furthermore, existing technologies often focus only on single fault diagnosis functions, lacking comprehensive solutions that integrate power management, signal isolation, protocol conversion, and intelligent diagnostics.

[0006] Therefore, how to design an intelligent diagnostic method specifically for the characteristics of the PA bus, effectively distinguish between voltage interference and equipment faults, improve diagnostic accuracy, and achieve online real-time diagnosis is a technical problem that urgently needs to be solved in the field of industrial automation. Summary of the Invention

[0007] The present invention aims to solve at least one of the technical problems existing in the prior art, and to provide a method for intelligent diagnosis of fieldbus PA networks.

[0008] To achieve the above objectives, the present invention provides an intelligent diagnostic method for fieldbus PA networks, comprising: The PA bus signal is subjected to multi-level isolation processing and data acquisition, the PA bus signal is subjected to noise suppression, and the isolated PA bus signal is sampled to obtain digital signal data. The digitized signal data is subjected to multi-level wavelet decomposition, the decomposition coefficients of each level are extracted, and an energy distribution vector is constructed based on the decomposition coefficients of each level. The energy distribution vector is input into the support vector machine classifier and the neural network classifier respectively to obtain two primary confidence scores. The fusion weights of the support vector machine classifier and the neural network classifier are dynamically adjusted according to the current signal-to-noise ratio. The two primary confidence scores are weighted and fused to obtain the final decision confidence score. Based on the final decision confidence level, interference-fault classification is performed. If the determination result is voltage interference, an adaptive filter is activated to suppress interference in the PA bus signal. If the determination result is equipment fault, time-domain reflection localization is triggered to determine the fault distance.

[0009] Furthermore, the multi-layer wavelet decomposition is performed as follows: The digitized signal data is subjected to 5-level db4 wavelet decomposition to extract the approximation coefficient cA5 and detail coefficients cD1 to cD5; The energy value of each layer is calculated based on the approximation coefficient cA5 and the detail coefficients cD1 to cD5, and the calculation method for the energy value of each layer is as follows: ; in, Represents the energy distribution vector. Indicates the first The energy values ​​corresponding to each decomposition component, where cD1 to cD5 represent the detail coefficients of layers 1 to 5, respectively. Indicates the first All wavelet coefficients corresponding to each decomposition component; The energy distribution vector is constructed by combining the energy values ​​of each layer. .

[0010] Furthermore, the signal-to-noise ratio is calculated as follows: Based on the Manchester encoding characteristics of the PA bus, the rise time deviation within a consecutive preset number of bit cycles is measured, and the standard deviation of the bit timing error is calculated. ; Based on the signal amplitude of the PA bus signal and the standard deviation of the bit timing error Calculate the signal-to-noise ratio: ; in, This indicates the signal-to-noise ratio of the PA bus signal. The signal amplitude of the PA bus signal. Let be the standard deviation of the bit timing error.

[0011] Furthermore, the method for dynamically adjusting the fusion weights is as follows: When the signal-to-noise ratio is lower than a preset signal-to-noise ratio threshold, the fusion weight of the support vector machine classifier is set to a first preset high weight value, and the fusion weight of the neural network classifier is set to a first preset low weight value. When the signal-to-noise ratio is not lower than the preset signal-to-noise ratio threshold, the fusion weight of the support vector machine classifier is set to a second preset low weight value, and the fusion weight of the neural network classifier is set to a second preset high weight value. The final decision confidence level is calculated as follows: ; in, This indicates the confidence level of the final decision. and These represent the fusion weights of the support vector machine classifier and the neural network classifier, respectively. and These represent the two primary confidence levels, respectively.

[0012] Furthermore, the adaptive filter is an LMS adaptive filter, the order of which is a preset filtering order, and the step size factor is adaptively and iteratively optimized within a preset step size range according to the interference intensity.

[0013] Furthermore, the temporal domain reflection localization is performed as follows: An adjustable pulse width TDR pulse is injected into the PA bus, and the pulse width is automatically selected according to the length of the PA bus cable; Measuring the time delay of reflected waves The fault distance is calculated based on the time delay of the reflected wave. ; in, Indicates the distance to the fault. This indicates the speed at which the signal propagates in the PA cable.

[0014] Furthermore, the data acquisition, wavelet decomposition, and classification are executed using a three-thread parallel architecture: The real-time diagnostic thread processes time-domain features at a preset period. The spectrum analysis thread is triggered to execute a fast Fourier transform when the number of collected data reaches a preset sample size; The model inference thread performs a weighted fusion decision between the support vector machine classifier and the neural network classifier; The three threads are coupled through a circular message queue. When the water level of the circular message queue exceeds a preset water level threshold, a feature simplification algorithm is automatically started to reduce the feature dimension and extend the processing cycle of the real-time diagnostic thread.

[0015] Furthermore, the method also includes a step of constructing an adaptive diagnostic model based on device type: Independent feature subspaces are constructed according to device type. Each independent feature subspace adopts a transfer learning strategy to map the fault modes of other device types to the target device to expand the training samples. The transfer learning strategy includes: selecting source domain devices, establishing a mapping relationship between the feature subspaces of the source domain and the target domain through the maximum mean difference minimization method, adding the fault samples of the source domain devices to the target domain training set after mapping, and fine-tuning the pre-trained model using target device samples on the transferred training set.

[0016] Furthermore, the method also includes an edge-cloud collaborative diagnostic step: When the final decision confidence level is within a preset uncertainty range, the compressed energy distribution vector is uploaded to the cloud via a communication protocol; The cloud-based system performs a secondary verification on the energy distribution vector using a deep learning model and outputs verification weights. The edge terminal corrects the local diagnostic results based on the aforementioned verification weights.

[0017] Furthermore, the method also includes an online diagnostic trigger optimization step: The PA bus signal is acquired in real time, and the Mahalanobis distance between the current feature vector and the historical normal baseline is calculated. When the Mahalanobis distance is greater than the preset distance threshold, the multi-layer wavelet decomposition and subsequent interference-fault classification and determination steps are initiated. When the Mahalanobis distance is not greater than the preset distance threshold, the multi-layer wavelet decomposition step is skipped to save computing power.

[0018] The beneficial effects of this invention are as follows: This invention extracts the energy distribution characteristics of PA bus signals through multi-level isolation processing and multi-layer wavelet decomposition. Combined with a dual-classifier fusion decision architecture of support vector machine and neural network, it can effectively distinguish between voltage interference and equipment faults. Compared with traditional single classifier methods, the separation accuracy is significantly improved.

[0019] This invention employs a dynamic weight adjustment mechanism based on signal-to-noise ratio. In environments with strong electromagnetic interference, it prioritizes robust support vector machine classifiers, while in environments with weak interference, it prioritizes neural network classifiers with strong learning capabilities. This enables the diagnostic system to adaptively cope with working conditions of varying interference intensities.

[0020] This invention achieves real-time interference suppression by activating an LMS adaptive filter for voltage interference and determines the fault location by triggering time-domain reflection for equipment fault location, thereby realizing the classification and handling of interference and fault and improving the practicality of diagnostic results.

[0021] This invention employs a three-thread parallel architecture and a ring message queue coupling mechanism to execute real-time diagnostics, spectrum analysis, and model inference tasks in parallel, thereby meeting the real-time requirements of industrial control systems while ensuring diagnostic accuracy. Attached Figure Description

[0022] Figure 1 This is a flowchart of the intelligent diagnostic method for fieldbus PA networks according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the intelligent diagnostic module system architecture according to an embodiment of the present invention; Figure 3 This is a flowchart of the dynamic weight fusion decision-making process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a three-thread parallel architecture according to an embodiment of the present invention; Figure 5 This is a diagram of the edge-cloud collaborative diagnostic architecture according to an embodiment of the present invention.

[0023] Figure 6 This is a schematic diagram of the priority message queue configuration according to an embodiment of the present invention; Figure 7 This is an application scenario diagram of the chemical water treatment system for thermal power units according to an embodiment of the present invention; Figure 8 This is a flowchart illustrating the signal processing and feature extraction process according to an embodiment of the present invention. Figure 9 This is a schematic diagram illustrating the TDR fault location principle of an embodiment of the present invention; Figure 10 This is a circuit diagram of a three-level heterogeneous signal isolation unit according to an embodiment of the present invention; Figure 11 This is a circuit diagram of the PA / DP protocol conversion interface according to an embodiment of the present invention; Figure 12 This is a circuit diagram of a multi-stage noise suppression filter according to an embodiment of the present invention; Figure 13 This is a schematic diagram of the TDR fault location circuit according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and beneficial effects of this application clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, further illustrates this application. It should be understood that the specific embodiments described in this specification are merely for explaining this application and are not intended to limit it.

[0025] The intelligent diagnostic method for fieldbus PA networks of this invention is implemented based on an intelligent diagnostic module. (See also...) Figure 2 The intelligent diagnostic module adopts a fully modular architecture, including five core functional units: a 24V bus power supply unit, a signal isolation unit, a protocol conversion unit, a data acquisition unit, and a diagnostic processing unit.

[0026] The 24V bus power supply unit provides a stable power supply to the connected PA bus. It adopts switching power supply technology, with an input voltage range of AC85 to 264V, an output voltage of DC24V, and a maximum output current of 5A. This unit has multiple protection functions, including overvoltage protection, overcurrent protection, and short-circuit protection.

[0027] See Figure 10 The signal isolation unit adopts a three-level heterogeneous isolation architecture, including a power isolation transformer, a digital isolator, and an analog isolation amplifier. The secondary winding of the power isolation transformer provides a floating ground for the digital isolator. The output enable terminal of the digital isolator directly controls the chopper modulation clock of the analog isolation amplifier through hardwire. The ground terminals of the three are connected at a single-point star topology node.

[0028] See Figure 11 The protocol conversion unit enables bidirectional conversion between PROFIBUS-PA and PROFIBUS-DP protocols, and includes a PA interface circuit, a DP interface circuit, and a protocol conversion processor. The PA interface circuit conforms to the IEC61158-2 standard, and the DP interface circuit conforms to the RS485 standard.

[0029] The data acquisition unit employs a high-speed ADC converter with a sampling rate of up to 1Msps and a resolution of 16 bits. The ADC sample-and-hold capacitor front end integrates a second-order active anti-aliasing filter composed of operational amplifiers. The filter cutoff frequency is configured at 50kHz, with a Q value of 0.707, forming an oversampling-filtering coordinated configuration with the 1Msps sampling rate.

[0030] The diagnostic processing unit uses a high-performance embedded processor with a built-in ring hardware message queue and a three-thread parallel architecture to run the diagnostic algorithm.

[0031] Example 1 This embodiment uses the PA bus diagnostics of a pH meter device in a thermal power unit's chemical water treatment system as an application scenario. (See also...) Figure 7 The chemical water treatment system for thermal power units comprises three main parts: a condensate polishing system, a feedwater oxygenation system, and a circulating water treatment system. A pH meter is installed at the outlet of the mixed bed to monitor water quality and is connected to the DCS control system via a PA bus. The bus length is approximately 800 meters, and the typical operating environment temperature is 40 to 60 degrees Celsius, with electromagnetic interference from numerous frequency converters and motors present.

[0032] See Figure 1 The diagnostic method in this embodiment includes the following steps: Step S1: Perform multi-level isolation processing and data acquisition on the PA bus signal.

[0033] See Figure 12 Signal isolation processing: The output of the 24V bus power supply unit is connected in series with a common-mode choke and then to the signal isolation unit to form a noise suppression pre-stage circuit. The inductance of the common-mode choke is 10mH, the winding asymmetry is less than 1%, and it is paired with an X2 type safety capacitor of 0.47µF and a Y2 type capacitor of 2.2nm to form a π-type filter network, with an insertion loss greater than 40dB in the 150kHz to 30MHz frequency band.

[0034] The signal isolation unit employs a three-stage heterogeneous isolation architecture, sequentially performing power isolation, digital signal isolation, and analog signal isolation. Power isolation utilizes an SN6501 push-pull transformer driver, with an internally integrated oscillator generating a 500kHz square wave to drive a miniature transformer with a 1:1.5 turns ratio. The secondary side is rectified to output a 3.3V isolated power supply. The transformer insulation layer uses a 0.08mm thick polyimide film with a withstand voltage of 3000V per micrometer and an isolation capacitance of less than 2pF. Digital signal isolation uses the ISO7720DR chip, based on capacitive coupling technology, achieving a signal transmission delay of only 11ns and a common-mode transient immunity (CMTI) greater than 100kV per microsecond. Analog signal isolation uses the ISO124P, internally employing Σ-Δ modulation and digital isolation technology, with a gain error of ±0.05%, nonlinearity of ±0.01%, and a -3dB bandwidth of 50kHz. The RC filter network at the analog signal input has a cutoff frequency of 10kHz, providing -40dB attenuation per decibel outside the PA signal band.

[0035] Data Acquisition: The isolated PA bus signal enters the data acquisition unit. The ADG708BR multiplexer has 8 input channels, an on-resistance of 3 ohms, and a channel switching time of 30ns. The programmable gain amplifier is set to a gain of 2 times for the PA bus signal amplitude range, ensuring the signal swing matches the ADC's full-scale range. The ADS1120 is configured in continuous conversion mode with a data rate set to 1Msps. This sampling rate, based on the Nyquist criterion, is more than 10 times the highest frequency component of the PA bus. The second-order active anti-aliasing filter has a cutoff frequency of 50kHz and a Q value of 0.707, ensuring ripple-free passband. The acquired digitized signal data is stored in a data buffer via a DMA controller. When 256 samples have accumulated, the data is automatically transferred to the diagnostic processing unit.

[0036] Step S2: Perform multi-level wavelet decomposition on the digitized signal data, extract the decomposition coefficients of each level, and construct the energy distribution vector based on the decomposition coefficients of each level.

[0037] See Figure 8 The digitized signal data obtained in step S1 is decomposed into 5-level db4 wavelet decomposition. The db4 wavelet has a support length of 8 and a regularization order of 4, achieving a good balance between time-domain and frequency-domain resolution. After decomposition, approximation coefficients cA5 and detail coefficients cD1 to cD5 are obtained. The frequency range corresponding to the approximation coefficient cA5 is 0 to 488 Hz, and the frequency ranges corresponding to the detail coefficients cD1 to cD5 are 7813 to 15625 Hz, 3906 to 7813 Hz, 1953 to 3906 Hz, 976 to 1953 Hz, and 488 to 976 Hz, respectively.

[0038] The energy value of each layer is calculated based on the decomposition coefficients of each layer. The calculation formula is as follows: ; in, This represents all the decomposition coefficients of the corresponding layer.

[0039] An energy distribution vector is constructed by combining the energy values ​​of each layer. The energy distribution vector is normalized and then used as the input to the classifier.

[0040] In this embodiment, voltage interference typically manifests as a wide-band energy distribution, with over 60% of the total energy concentrated in layers cD1 to cD3. Equipment failure manifests as a sudden drop in signal amplitude, with the energy primarily concentrated in layer cA5, accounting for over 70%.

[0041] Step S3: Input the energy distribution vector into the support vector machine classifier and the neural network classifier respectively to obtain two initial confidence scores, and dynamically adjust the fusion weights according to the current signal-to-noise ratio to calculate the final decision confidence score.

[0042] See Figure 3 First, calculate the current signal-to-noise ratio. Based on the 31.25kbps Manchester encoding characteristics of the PA bus, measure the rise time deviation over 100 consecutive bit cycles and calculate the standard deviation of the bit timing error. The standard deviation of the bit timing error is obtained by taking the standard deviation of the time difference between the rising edge and the theoretical bit center time within 100 bit periods. The formula for calculating the signal-to-noise ratio is: ; in, The signal amplitude of the PA bus signal.

[0043] The normalized energy distribution vector E is input in parallel into both the support vector machine (SVM) classifier and the neural network classifier. The SVM classifier uses a linear kernel function, and the classification hyperplane is composed of... Determine, where the weight vector and bias The data was obtained through training on historical data from the chemical water treatment system of thermal power units. The training samples included at least 15 fault modes and at least 8 interference modes. The support vector machine classifier was used based on the classification hyperplane. The positive or negative classification is either equipment failure or voltage interference, and the corresponding primary confidence level is output. The neural network classifier uses a three-layer feedforward network. The number of nodes in the input layer equals the feature dimension, the hidden layer has 20 nodes, and the output layer has 2 nodes corresponding to normal and faulty conditions, respectively. The activation function is ReLU, and the output is the initial confidence score. .

[0044] The fusion weights are dynamically adjusted based on the current signal-to-noise ratio. At that time, the fusion weights of the support vector machine classifier Fusion weights of neural network classifiers .when At that time, the fusion weights of the support vector machine classifier Fusion weights of neural network classifiers .

[0045] The formula for calculating the final decision confidence level is: ; in, Indicates the confidence level of the final decision. and These represent the fusion weights of the support vector machine classifier and the neural network classifier, respectively. and These represent two primary confidence levels.

[0046] In this embodiment, the area where the pH meter is located experiences strong inverter interference, with a measured signal-to-noise ratio of approximately 7dB, which is below the 10dB threshold. Therefore, the fusion weight is automatically set to... , It favors support vector machine classifiers, which are more robust.

[0047] Step S4: Based on the final decision confidence level, classify the interference-fault and implement corresponding handling measures.

[0048] When the final decision confidence level When the disturbance is determined to be voltage interference, a 32nd-order LMS adaptive filter is activated. The step size factor of the LMS adaptive filter is... The filter is adaptively iteratively optimized based on the interference intensity within the range of 0.005 to 0.02. In this embodiment, the initial step size factor is set to 0.01, the convergence time is approximately 50ms, and the steady-state error is less than 5%. The filter performs real-time interference suppression on subsequent PA bus signals, filtering out electromagnetic interference components and restoring signal quality.

[0049] When the final decision confidence level When a device malfunction is detected, time-domain reflectometry (TDR) positioning is triggered. An adjustable pulse width (TDR) pulse is injected into the PA bus. The pulse width is automatically selected based on the cable length: 50ns for cable lengths less than 500 meters; 100ns for cable lengths between 500 and 1000 meters; and 200ns for cable lengths of 1000 meters or more. The TDR pulse rise time is set to 20ns, and the amplitude is 5V.

[0050] In this embodiment, the pH meter PA bus cable is 800 meters long, and the pulse width is automatically selected as 100 ns. See also... Figure 9 and Figure 13 After injecting a TDR pulse, the time delay of the reflected wave is measured. The formula for calculating the fault distance is: ; in, Indicates the distance to the fault. This indicates the signal propagation speed in the PA cable, calculated at 0.8 times the speed of light. With a PA cable characteristic impedance of 100 ohms, the round-trip delay is approximately 10 ns per meter. Based on this, the fault distance can be calculated with a resolution of ±1 meter.

[0051] In this embodiment, after the diagnostic steps described above, the system detected voltage interference on the pH meter PA bus, and finally determined the confidence level. The system automatically activated the LMS adaptive filter, and after interference suppression, the signal-to-noise ratio increased from 7dB to 18dB, restoring normal communication with the pH meter.

[0052] The execution process of the three-thread parallel architecture in this embodiment is as follows. (See attached document.) Figure 4 The real-time diagnostic thread processes time-domain features, including 12-dimensional time-domain feature parameters such as signal amplitude, mean, and variance, at a 10ms cycle. The spectrum analysis thread is triggered when the ADC data reaches 256 samples, performing a 1024-point FFT with a computation time of 0.8ms. The model inference thread uses a weighted fusion decision of support vector machine and neural network. The three threads are coupled through a circular message queue with a depth of 128 messages. When the level of the circular message queue exceeds 75%, a time-domain feature simplification algorithm is automatically activated, reducing the feature dimension from 12 to 6, while extending the real-time diagnostic thread cycle from 10ms to 15ms to ensure that the diagnostic update rate of the pH meter is not less than 10Hz.

[0053] The online diagnostic trigger optimization process in this embodiment is as follows: The system acquires PA bus signals in real time and calculates the Mahalanobis distance between the current feature vector and the historical normal baseline. When the Mahalanobis distance is greater than 3 times the standard deviation, wavelet decomposition and subsequent classification determination steps are initiated. When the Mahalanobis distance is not greater than 3 times the standard deviation, the wavelet decomposition step is skipped to save computational power. In this embodiment, wavelet decomposition is not required in approximately 78% of the sampling periods under normal operating conditions, effectively reducing the processor's computational load.

[0054] Example 2 This embodiment uses the PA bus diagnostics of a conductivity meter in a thermal power unit's chemical water treatment system as an application scenario. The conductivity meter is installed in the water supply pipeline to monitor ion concentration and is connected to the DCS control system via the PA bus. The bus length is approximately 400 meters, the operating environment temperature is 45 to 55 degrees Celsius, and there is electromagnetic interference from high-power motors and variable frequency speed control devices in the vicinity.

[0055] The diagnostic method in this embodiment includes the following steps: Step S1: Perform multi-level isolation processing and data acquisition on the PA bus signal.

[0056] The signal isolation process is the same as in Example 1, employing a three-stage heterogeneous isolation architecture. The output of the 24V bus power supply unit is connected in series with a common-mode choke and then to the signal isolation unit, forming a noise suppression pre-stage circuit. The signal isolation unit sequentially performs power isolation, digital signal isolation, and analog signal isolation. The secondary winding of the power isolation transformer provides a floating ground for the digital isolator. The output enable terminal of the digital isolator directly controls the chopper modulation clock of the analog isolation amplifier via hardwired connections. The grounding terminals of all three are converged at a single-point star topology node.

[0057] During data acquisition, the programmable gain amplifier was set to a gain of 2 times for the amplitude range of the conductivity meter's PA bus signal. The ADS1120 was configured in continuous conversion mode, and the data rate was set to 1Msps. The cutoff frequency of the second-order active anti-aliasing filter was configured to 50kHz, and the Q value was 0.707. The acquired digitized signal data was stored in the data buffer via a DMA controller.

[0058] Step S2: Perform multi-level wavelet decomposition on the digitized signal data to construct the energy distribution vector.

[0059] The digitized signal data obtained in step S1 is subjected to a 5-level db4 wavelet decomposition to extract the approximation coefficients cA5 and detail coefficients cD1 to cD5. The energy value of each level is calculated based on the decomposition coefficients, using the following formula: ; An energy distribution vector is constructed by combining the energy values ​​of each layer. The energy distribution vector is normalized and then used as the input to the classifier.

[0060] Step S3: Input the energy distribution vector into the dual classifier for fusion decision.

[0061] Calculate the current signal-to-noise ratio. Based on the 31.25kbps Manchester encoding characteristics of the PA bus, measure the rise time deviation over 100 consecutive bit cycles and calculate the standard deviation of the bit timing error. The formula for calculating the signal-to-noise ratio is: .

[0062] The normalized energy distribution vector E is input in parallel into both the support vector machine (SVM) classifier and the neural network classifier. The SVM classifier uses a linear kernel function and weight vectors. and bias The data was obtained through training with historical data from conductivity meters in the chemical water treatment system of thermal power units. The training samples include 5-dimensional subspace features, including impedance phase angle characteristics, and are represented as vectors. ,in to This represents the energy percentage of each layer after db4 wavelet decomposition. The impedance phase angle is given by a 1kHz excitation signal. The structure of the neural network classifier is the same as in Example 1, employing a three-layer feedforward network.

[0063] In this embodiment, the interference in the area where the conductivity meter is located is relatively weak, with a measured signal-to-noise ratio of approximately 14 dB, which is greater than the 10 dB threshold. Therefore, the fusion weight is automatically set to... , It favors neural network classifiers with strong learning capabilities. The formula for calculating the final decision confidence is: ; Step S4: Classify interference and faults based on the final decision confidence level.

[0064] In this embodiment, after the diagnostic steps described above, the system detected a device fault in the conductivity meter's PA bus, and finally decided on the confidence level. The system automatically triggers time-domain reflectometry (TDR) positioning. Since the cable length is 400 meters (less than 500 meters), the pulse width is automatically selected as 50 ns. After injecting the TDR pulse, the time delay of the reflected wave is measured. The calculated fault distance is as follows: ; The calculated result is greater than the total cable length of 400 meters, indicating that the reflected wave originates from the terminating resistance at the cable end, which is a normal terminating reflection. The system continued measurement and detected... An abnormal reflected signal is present at the location, and the calculated fault distance is as follows: ; On-site verification revealed a poor contact fault in the junction box approximately 300 meters from the conductivity meter, consistent with the TDR positioning results. After maintenance personnel repaired the junction box, the PA bus communication of the conductivity meter returned to normal.

[0065] This embodiment also includes the construction of an adaptive diagnostic model based on device type. For conductivity meters, a 5-dimensional subspace containing impedance phase angle features is constructed. A transfer learning strategy is employed, selecting pH meters and dissolved oxygen meters as source domain devices. A mapping relationship between the source and target domain feature subspaces is established using the maximum mean difference minimization method. For the source domain feature vector... and target domain feature vector Construct mapping function ,in For the transformation matrix, minimize Solve the problem by mapping fault samples from the source domain devices into the target domain training set and then transferring the weights. Calculated based on device similarity: ,in Let be the Euclidean distance between the feature subspaces of the two devices. The maximum distance between all devices is defined. On the transferred training set, the pre-trained model is fine-tuned using a small number of conductivity meter samples, with a learning rate of 0.0001 and 50 iterations.

[0066] In this embodiment, through the transfer learning strategy, the effective training samples of the conductivity meter were expanded from the original 800 sets to 2200 sets, and the diagnostic accuracy was improved from 91.5% to 95.8%.

[0067] Example 3 This embodiment uses the PA bus diagnostics of a dissolved oxygen meter in a thermal power unit's chemical water treatment system as an application scenario, and introduces an edge-cloud collaborative diagnostic mechanism. The dissolved oxygen meter is installed in the cooling water circuit to monitor dissolved oxygen content and is connected to the DCS control system via the PA bus. The bus length is approximately 1200 meters, the operating environment temperature is 50 to 60 degrees Celsius, and there is strong electromagnetic interference generated by numerous high-power devices.

[0068] The diagnostic method in this embodiment includes the following steps: Step S1: Perform multi-level isolation processing and data acquisition on the PA bus signal.

[0069] The signal isolation process employs the same three-stage heterogeneous isolation architecture as in Example 1. A common-mode choke is connected in series with the output of the 24V bus power supply unit and then to the signal isolation unit. The signal isolation unit sequentially performs power isolation, digital signal isolation, and analog signal isolation. The isolation capacitor of the power isolation transformer is controlled below 2pF, the common-mode transient immunity (CMTI) of the digital isolator is higher than 100kV / µs, and the gain error of the analog isolation amplifier is controlled within ±0.05%.

[0070] During data acquisition, the programmable gain amplifier is set to a gain of 2 times for the amplitude range of the dissolved oxygen analyzer's PA bus signal. The ADS1120 is configured in continuous conversion mode, and the data rate is set to 1Msps. The acquired digitized signal data is stored in the data buffer via the DMA controller.

[0071] Step S2: Perform multi-level wavelet decomposition on the digitized signal data to construct the energy distribution vector.

[0072] The digitized signal data obtained in step S1 is subjected to a 5-level db4 wavelet decomposition to extract the approximation coefficient cA5 and detail coefficients cD1 to cD5. The energy value of each level is calculated based on the decomposition coefficients. ; An energy distribution vector is constructed by combining the energy values ​​of each layer. .

[0073] For dissolved oxygen analyzer equipment, a 7-dimensional subspace including polarization voltage stability is constructed, which is represented by vector as follows: ,in to The energy percentage of each layer after 5-layer decomposition of the db4 wavelet. This represents the polarization voltage fluctuation range. The dimensions of each subspace were determined using principal component analysis, retaining over 95% of the variance information.

[0074] Step S3: Input the energy distribution vector into the dual classifier for fusion decision.

[0075] Calculate the current signal-to-noise ratio. Based on the Manchester encoding characteristics of the PA bus, measure the rise time deviation over 100 consecutive bit cycles and calculate the standard deviation of the bit timing error. The formula for calculating the signal-to-noise ratio is: .

[0076] In this embodiment, strong electromagnetic interference exists in the area where the dissolved oxygen analyzer is located, with a measured signal-to-noise ratio of approximately 5dB, which is lower than the 10dB threshold. Therefore, the fusion weight is automatically set to... , The normalized energy distribution vector E is input in parallel into the support vector machine (SVM) classifier and the neural network classifier. The weight vector of the SVM classifier... and bias The data was obtained through training with historical data from dissolved oxygen analyzers. The structure of the neural network classifier is the same as in Example 1. The formula for calculating the final decision confidence is: ; In this embodiment, after preliminary judgment, the final decision confidence level is determined. The uncertainty level falls within the preset range of 0.5 to 0.7. At this point, the confidence level of the edge decision is insufficient for reliable judgment, triggering edge-cloud collaborative diagnosis.

[0077] Step S4: Perform edge-cloud collaborative diagnostics.

[0078] See Figure 5 When the final decision confidence level When the value is within the uncertainty range of 0.5 to 0.7, the compressed energy distribution vector E is uploaded to the cloud server via the MQTT protocol.

[0079] A deep learning model with a dual-branch LSTM network architecture is deployed on the cloud server. The model's input layer receives a compressed energy vector E. The main branch is a two-layer LSTM network with 64 units per layer; the auxiliary branches are three-layer fully connected networks with 32, 16, and 8 neurons respectively. The outputs of the two branches are concatenated at a feature fusion layer, and weights are calculated using an attention mechanism. Finally, the weights are verified by a sigmoid activation function. The model was trained using six months of historical data from thermal power units, including 8000 normal samples and 2000 fault samples. The mean squared error loss function was used, with Adam as the optimizer, a learning rate of 0.001, a batch size of 32, and 200 training epochs. L2 regularization and dropout layers were added to prevent overfitting.

[0080] Cloud model output verification weights The calculation formula is: ; in, For edge prediction results, This represents the prediction results from the cloud-based model. When the final decision confidence level is below 0.7, the validation weights are adjusted. When the final decision confidence level is not lower than 0.7, the verification weight is... The edge end is determined according to the formula. Correct the diagnostic results. If the corrected output value is greater than 0.75, a fault is confirmed.

[0081] In this embodiment, the cloud model determines that the abnormal PA bus signal of the dissolved oxygen meter originates from a device malfunction, and verifies the weights. The corrected diagnostic output value is 0.81, which is greater than 0.75, confirming a device malfunction.

[0082] The system then triggered time-domain reflectometry (TDR) positioning. Since the dissolved oxygen analyzer's PA bus cable length is 1200 meters (not less than 1000 meters), the pulse width was automatically selected as 200 ns. After injecting the TDR pulse, the time delay of the reflected wave was measured, and it was detected that... An abnormal reflected signal is present at the location, and the calculated fault distance is as follows: ; On-site verification revealed a short circuit caused by insulation aging in the cable approximately 720 meters from the dissolved oxygen analyzer, consistent with the TDR location results. After maintenance personnel replaced the faulty section of cable, the PA bus communication of the dissolved oxygen analyzer returned to normal.

[0083] See Figure 6 In this embodiment, the three-thread parallel architecture is configured with a three-level priority message queue. The first level is for pH meter signal processing, with a queue depth of 128 and a response latency of less than 2ms. The second level is for conductivity meter signal processing, with a queue depth of 64 and a response latency of less than 5ms. The third level is for dissolved oxygen meter signal processing, with a queue depth of 32 and a response latency of less than 10ms. This three-level priority configuration ensures that the pH value diagnostic update rate remains no less than 10Hz when the PA bus load rate is greater than 85%.

[0084] The online diagnostic trigger optimization process in this embodiment is the same as in Embodiment 1. The system acquires PA bus signals in real time and calculates the Mahalanobis distance between the current feature vector and the historical normal baseline. When the Mahalanobis distance is greater than 3 times the standard deviation, wavelet decomposition and subsequent classification steps are initiated. When the Mahalanobis distance is not greater than 3 times the standard deviation, the wavelet decomposition step is skipped to save computing power.

[0085] This embodiment also includes the construction of an adaptive diagnostic model based on device type. For dissolved oxygen meters, a 7-dimensional subspace including polarization voltage stability is constructed. A transfer learning strategy is employed, selecting pH meters and conductivity meters as source domain devices. A mapping relationship between the feature subspaces of the source and target domains is established using the maximum mean difference minimization method. Fault samples from the source domain devices are then added to the target domain training set after mapping. Transfer weights... Calculated based on device similarity: On the transferred training set, the pre-trained model was fine-tuned using a small number of dissolved oxygen meter samples. Through the transfer learning strategy, the effective training samples for dissolved oxygen meters were expanded from the original 600 sets to 1800 sets, and the diagnostic accuracy improved from 89.7% to 94.2%.

[0086] In summary, the embodiments of the present invention have at least the following technical effects: This invention extracts the energy distribution characteristics of PA bus signals through multi-level isolation processing and multi-layer wavelet decomposition. Combined with a dual-classifier fusion decision architecture of support vector machine and neural network, it can effectively distinguish between voltage interference and equipment failure under different signal-to-noise ratio conditions. The average separation accuracy of the three embodiments reaches 95.2%.

[0087] This invention employs a dynamic weight adjustment mechanism based on signal-to-noise ratio (SNR). It prioritizes the support vector machine classifier in environments with strong electromagnetic interference and the neural network classifier in environments with weak interference, enabling the diagnostic system to adaptively cope with different interference intensities. Even in strong interference scenarios with an SNR below 5 dB, the separation accuracy can still be maintained above 92%.

[0088] This invention uses a three-thread parallel architecture and a ring message queue coupling mechanism to execute real-time diagnostics, spectrum analysis and model inference tasks in parallel, with the single-frame data processing time controlled within 2ms, thus meeting the real-time requirements of industrial control systems.

[0089] This invention improves diagnostic reliability by employing an edge-cloud collaborative diagnostic mechanism. For scenarios where edge-side decision-making confidence is insufficient, a cloud-based deep learning model is used for secondary verification. In long-term operational tests of the three embodiments, the system provided an average warning 52 hours before a fault occurred, achieving a 100% timely maintenance response rate.

[0090] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for intelligent diagnostics of a fieldbus PA network, characterized in that, include: The PA bus signal is subjected to multi-level isolation processing and data acquisition, the PA bus signal is subjected to noise suppression, and the isolated PA bus signal is sampled to obtain digital signal data. The digitized signal data is subjected to multi-level wavelet decomposition, the decomposition coefficients of each level are extracted, and an energy distribution vector is constructed based on the decomposition coefficients of each level. The energy distribution vector is input into the support vector machine classifier and the neural network classifier respectively to obtain two primary confidence scores. The fusion weights of the support vector machine classifier and the neural network classifier are dynamically adjusted according to the current signal-to-noise ratio. The two primary confidence scores are weighted and fused to obtain the final decision confidence score. Based on the final decision confidence level, interference-fault classification is performed. If the determination result is voltage interference, an adaptive filter is activated to suppress interference in the PA bus signal. If the determination result is equipment fault, time-domain reflection localization is triggered to determine the fault distance.

2. The intelligent diagnostic method for fieldbus PA networks according to claim 1, characterized in that, The multi-layer wavelet decomposition is performed as follows: The digitized signal data is subjected to 5-level db4 wavelet decomposition to extract the approximation coefficient cA5 and detail coefficients cD1 to cD5; The energy value of each layer is calculated based on the approximation coefficient cA5 and the detail coefficients cD1 to cD5, and the calculation method for the energy value of each layer is as follows: ； in, Represents the energy distribution vector. Indicates the first The energy values ​​corresponding to each decomposition component, where cD1 to cD5 represent the detail coefficients of layers 1 to 5, respectively. Indicates the first All wavelet coefficients corresponding to each decomposition component; The energy distribution vector is constructed by combining the energy values ​​of each layer. .

3. The intelligent diagnostic method for fieldbus PA networks according to claim 1, characterized in that, The signal-to-noise ratio is calculated as follows: Based on the Manchester encoding characteristics of the PA bus, the rise time deviation within a consecutive preset number of bit cycles is measured, and the standard deviation of the bit timing error is calculated. ; Based on the signal amplitude of the PA bus signal and the standard deviation of the bit timing error Calculate the signal-to-noise ratio: ； in, This indicates the signal-to-noise ratio of the PA bus signal. The signal amplitude of the PA bus signal. Let be the standard deviation of the bit timing error.

4. The intelligent diagnostic method for fieldbus PA networks according to claim 3, characterized in that, The method for dynamically adjusting the fusion weights is as follows: When the signal-to-noise ratio is lower than a preset signal-to-noise ratio threshold, the fusion weight of the support vector machine classifier is set to a first preset high weight value, and the fusion weight of the neural network classifier is set to a first preset low weight value. When the signal-to-noise ratio is not lower than the preset signal-to-noise ratio threshold, the fusion weight of the support vector machine classifier is set to a second preset low weight value, and the fusion weight of the neural network classifier is set to a second preset high weight value. The final decision confidence level is calculated as follows: ； in, This indicates the confidence level of the final decision. and These represent the fusion weights of the support vector machine classifier and the neural network classifier, respectively. and These represent the two primary confidence levels, respectively.

5. The intelligent diagnostic method for fieldbus PA networks according to claim 1, characterized in that, The adaptive filter is an LMS adaptive filter, and the order of the LMS adaptive filter is a preset filtering order. The step size factor is adaptively and iteratively optimized within a preset step size range according to the interference intensity.

6. The intelligent diagnostic method for fieldbus PA networks according to claim 1, characterized in that, The temporal reflection localization is performed as follows: An adjustable pulse width TDR pulse is injected into the PA bus, and the pulse width is automatically selected according to the length of the PA bus cable; Measuring the time delay of reflected waves The fault distance is calculated based on the time delay of the reflected wave. ； in, Indicates the distance to the fault. This indicates the speed at which the signal propagates in the PA cable.

7. The intelligent diagnostic method for fieldbus PA networks according to claim 1, characterized in that, The data acquisition, wavelet decomposition, and classification are executed through a three-thread parallel architecture: The real-time diagnostic thread processes time-domain features at a preset period. The spectrum analysis thread is triggered to execute a fast Fourier transform when the number of collected data reaches a preset sample size; The model inference thread performs a weighted fusion decision between the support vector machine classifier and the neural network classifier; The three threads are coupled through a circular message queue. When the water level of the circular message queue exceeds a preset water level threshold, a feature simplification algorithm is automatically started to reduce the feature dimension and extend the processing cycle of the real-time diagnostic thread.

8. The intelligent diagnostic method for fieldbus PA networks according to claim 1, characterized in that, It also includes the steps for building an adaptive diagnostic model based on device type: Independent feature subspaces are constructed according to device type. Each independent feature subspace adopts a transfer learning strategy to map the fault modes of other device types to the target device to expand the training samples. The transfer learning strategy includes: selecting source domain devices, establishing a mapping relationship between the feature subspaces of the source domain and the target domain through the maximum mean difference minimization method, adding the fault samples of the source domain devices to the target domain training set after mapping, and fine-tuning the pre-trained model using target device samples on the transferred training set.

9. The intelligent diagnostic method for fieldbus PA networks according to claim 1, characterized in that, It also includes edge-cloud collaborative diagnostic steps: When the final decision confidence level is within a preset uncertainty range, the compressed energy distribution vector is uploaded to the cloud via a communication protocol; The cloud-based system performs a secondary verification on the energy distribution vector using a deep learning model and outputs verification weights. The edge terminal corrects the local diagnostic results based on the aforementioned verification weights.

10. The intelligent diagnostic method for fieldbus PA networks according to any one of claims 1 to 9, characterized in that, It also includes online diagnostic trigger optimization steps: The PA bus signal is acquired in real time, and the Mahalanobis distance between the current feature vector and the historical normal baseline is calculated. When the Mahalanobis distance is greater than the preset distance threshold, the multi-layer wavelet decomposition and subsequent interference-fault classification and determination steps are initiated. When the Mahalanobis distance is not greater than the preset distance threshold, the multi-layer wavelet decomposition step is skipped to save computing power.

Images