Blower real-time fault diagnosis method based on analog circuit physical neural network
By directly processing multi-source signals from blowers using analog circuit physical neural networks, the problems of delay and high power consumption in blower status monitoring at wastewater treatment plants have been solved. This has enabled low-latency, high-precision fault diagnosis, improving equipment safety and processing efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-05
- Publication Date
- 2026-05-15
AI Technical Summary
In existing technologies, the status monitoring of blowers in wastewater treatment plants suffers from signal delay, high power consumption, and insufficient diagnostic accuracy, making it difficult to achieve real-time and accurate fault diagnosis, especially the comprehensive capture of multi-dimensional information and timely early warning of potential faults.
A physical neural network with analog circuits is adopted. Multi-source signals are collected by sensors and processed directly. A physical neural network containing an input layer, a hidden layer and an output layer is built. Programmable resistors and operational amplifiers are used to perform signal weighting and nonlinear mapping to achieve fault detection, avoid the digital-to-analog conversion process, and update parameters in conjunction with a microcontroller.
It achieves near-zero latency and low power consumption real-time fault diagnosis, improves the accuracy and comprehensiveness of fault diagnosis, reduces operation and maintenance costs, ensures the safe and stable operation of the blower, and improves sewage treatment efficiency.
Smart Images
Figure CN122045894A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wastewater treatment equipment condition monitoring and fault diagnosis technology, specifically to a real-time fault diagnosis method for blowers based on analog circuit physical neural networks. Background Technology
[0002] As the core power equipment of the biochemical reaction aeration system in wastewater treatment plants, blowers need to operate continuously 24 hours a day. Their operating status directly determines aeration efficiency, wastewater treatment effect, and system energy consumption. Currently, wastewater treatment plants mostly rely on traditional electronic circuit diagnostic solutions for blower status monitoring. This involves multiple processing steps such as analog-to-digital conversion, data storage, and digital calculation, resulting in significant signal delays, high power consumption, and insufficient real-time response capabilities. Some plants use manual inspections or single-point sensor monitoring, which makes it difficult to comprehensively capture multi-dimensional operating information of the equipment, leading to a high rate of missed fault diagnosis and false alarms, and failing to provide timely warnings of potential fault risks.
[0003] While existing technologies include equipment fault diagnosis methods based on digital neural networks, these methods rely on complex algorithms and data storage units, resulting in high hardware costs and response speeds limited by data transmission and computation efficiency. In contrast, physical neural networks implemented using analog circuits possess inherent parallel processing characteristics, eliminating the need for analog-to-digital conversion and enabling direct real-time processing of analog signals. This offers advantages such as near-zero latency, low power consumption, and high efficiency. However, currently, there are no relevant technical solutions for applying analog circuit physical neural networks to multi-source signal fusion fault diagnosis of blowers in wastewater treatment plants, failing to meet the requirements for real-time, accurate, and low-power condition monitoring of blowers. Therefore, based on this requirement, this invention discloses a real-time fault diagnosis method for blowers based on analog circuit physical neural networks. Summary of the Invention
[0004] To address the problems of signal delay, high power consumption, and insufficient diagnostic accuracy in existing blower fault diagnosis methods, this invention provides a real-time blower fault diagnosis method based on analog circuit physical neural networks. First, current, temperature, and vibration signals are collected, and weight parameters are obtained through training. Finally, a physical neural network is built using analog circuits to directly process multi-source analog signals from the blower, achieving near-zero delay, low power consumption, and high-precision real-time fault diagnosis. This ensures the safe and stable operation of the blower and reduces the operation and maintenance costs of wastewater treatment plants.
[0005] The purpose of this invention is to provide a real-time classification method for motor faults based on a physical neural network using analog circuits.
[0006] To achieve the above objectives, the technical method of the present invention includes:
[0007] Step 1: Using analog components, build a trainable physical circuit neural network. The input includes three sensor input circuits and signal feature extraction and preprocessing circuits. The middle layer is a fully connected layer containing multiple neurons, represented by an operational amplifier inverse summation circuit. The activation function unit consists of a dual threshold comparator, a mid-section linear operational amplifier, and a selection switch. The confidence output is obtained by the operational amplifier summator and the output bias is superimposed. Then, it is mapped to the corresponding 0-1 confidence range by a limiting circuit.
[0008] Step 2: Collect the fault signals to be classified, build a digital model neural network corresponding to the above analog circuit, and train the parameters existing in the classification process, including weights, bias, and threshold voltage, based on the dataset and by applying physical constraints that conform to the actual circuit. Finally, obtain the converged parameter information, and set the programmable resistor and the corresponding threshold voltage through the microcontroller.
[0009] Step 3: Using the trained physical neural network classification circuit, multiple analog signals are collected by sensors during actual motor operation and directly fed into the circuit to complete fault identification. This classification method does not involve any digital-to-analog conversion process, offering significant real-time advantages.
[0010] Preferably, in step one, building the analog circuit physical neural network includes using analog circuit components such as operational amplifiers and programmable resistors to construct a physical neural network architecture containing an input layer, a hidden layer, and an output layer; the input layer is equipped with multiple signal receiving units, which correspond to the access channels for blower current, vibration, and temperature signals respectively; the hidden layer realizes the weighted operation of analog signals through a network composed of programmable resistors and operational amplifiers; and the output layer outputs fault diagnosis results through a voltage comparator.
[0011] Preferably, in step two, the three-channel sensor signal acquisition and preprocessing method includes: deploying distributed sensors at key locations such as the blower motor stator, bearing housing, and air inlet to collect multi-source analog signals such as current, voltage, vibration acceleration, and surface temperature in real time, and performing analog conditioning and characteristic processing to output three normalized characteristic voltages: temperature characteristic voltage X T Vibration characteristic voltage X V Current characteristic voltage X I Furthermore, these characteristic voltages are limited to a preset range;
[0012] Preferably, in steps two and three, the construction and parameter training of the physical neural network includes: selecting multi-source analog signals under the normal operating state of the blower and seven typical fault states (C0: normal operation, C1: bearing wear - slight, C2: bearing wear - severe, C3: impeller imbalance - slight, C4: impeller imbalance - severe, C5: motor winding overheating, C6: voltage abnormality) for feature processing as training samples; each fault discrimination channel C k include:
[0013] (1) Linear weighted summation unit (intermediate layer): Consists of an operational amplifier inverting summator and several input resistors / feedback resistors, which sum the values of X... T X V X I The weighted sums are then applied, along with the bias, to obtain the hidden node input. The weights are determined by the resistance ratio, and the bias is formed by injecting a reference voltage through a bias resistor; the weighting resistor and / or the bias resistor are programmable resistors.
[0014] (2) Nonlinear activation unit: Input to hidden node Perform nonlinear mapping to output hidden node voltages The activation unit is a hardware-mode Hard-Sigmoid circuit consisting of a dual threshold comparator, a mid-range linear operational amplifier, and a selection switch, to form a nonlinear response that approximates a sigmoid.
[0015] (3) Output layer confidence generation unit: outputs multiple hidden nodes The weighted sums are then added together with the output bias to generate the fault confidence voltage. And through the limiting circuit The equivalent voltage range is constrained to 0 to 1.
[0016] (4) Threshold comparison alarm unit: With configurable threshold Compare and output alarm signals. The comparator is configured to suppress jitter due to hysteresis. This is based on the output of seven channels. Determine the fault type, severity level, and complex fault conditions.
[0017] Preferably, the fault type determination strategy is that the fault comprehensive decision module supports at least one or a combination of the following decision methods:
[0018] (1) Threshold-triggered decision: when > Output the fault alarm at the time. .
[0019] (2) Composite fault determination: When multiple channels simultaneously meet the conditions > Output the composite fault set in real time, and The faults are sorted by size to indicate their primary and secondary importance.
[0020] Preferably, the weight parameters, bias parameters, and threshold parameters used for writing or updating each fault discrimination channel in the real-time fault diagnosis of this invention are completed by the above-mentioned training process, and the parameters are implemented by a programmable resistor and a threshold reference circuit. The multi-source preprocessed analog signals collected in real time during the blower's operation are directly input into the physical neural network; the neural network performs parallel computation through the analog circuit to analyze and process the signals in real time, outputting equipment health status assessment values and fault type determination results; when the assessment value is lower than a preset threshold or a fault characteristic is detected, an alarm signal is triggered.
[0021] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0022] 1. The core identification is completed by the operational amplifier weighted summation and activation circuit, which does not require complex digital calculations. It is suitable for online real-time monitoring and edge deployment. It adopts hardware real-time inference and has the characteristics of low latency and low power consumption.
[0023] 2. Weights, biases, and thresholds are set by programmable resistors and threshold references, and can be updated on-site via a microcontroller, enabling classification to be achieved simply by replacing parameters for different situations.
[0024] 3. By collecting multi-source analog signals of blower current, voltage, vibration, and temperature for fusion diagnosis, and combining the parallel processing capabilities of physical neural networks, the characteristics of equipment operation status are fully captured, significantly improving the accuracy and comprehensiveness of fault diagnosis and reducing the risk of missed or misdiagnosed cases.
[0025] 4. The output results of this invention are interpretable and scalable. Specifically, each type of fault has an independent confidence level and alarm, which facilitates display, recording, sorting and complex fault analysis; the number of channels and the number of hidden nodes can be expanded to adapt to different devices.
[0026] 5. The method is adapted to the harsh operating environment of sewage treatment plants. The analog circuit has low energy consumption and high stability, and can operate reliably for a long time. It provides technical support for the safe and uninterrupted operation of the blower, indirectly improving the sewage aeration efficiency and treatment effect, and achieving cost reduction and efficiency improvement. Attached Figure Description
[0027] Appendix Figure 1 This is a flowchart of the motor fault signal classification method of the present invention;
[0028] Appendix Figure 2 This is a block diagram of the overall structure of the patented technology of this invention, including signal acquisition, feature generation, parameter training, building an analog circuit neural network, and real-time classification process;
[0029] Appendix Figure 3 This is a schematic diagram of signal transmission inside the hidden layer;
[0030] Appendix Figure 4 A diagram illustrating the parameter setting process. Detailed Implementation
[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0032] Example: The present invention will be clearly described below with illustrations and detailed description. Any person skilled in the art who understands the examples of the present invention can make changes and modifications based on the technology taught in the present invention without departing from the spirit and scope of the present invention.
[0033] In this example, a physical neural network circuit with three inputs and seven analog outputs is used, taking a hidden node with M=5. It includes: a three-channel sensor acquisition module (temperature, vibration, and current); a feature generation and normalization module that outputs three feature voltages; and seven parallel fault discrimination channels, each with the same input and outputting a identifiable attribute. The fault comprehensive judgment module outputs fault codes and alarm levels; the parameter setting module is used to write the weights, biases, and thresholds of each channel.
[0034] In this example, the sensor output first passes through a buffered operational amplifier to form a low-impedance node, then branches into multiple feature extraction branches for parallel processing, and finally is weighted and synthesized into a single path. Feature voltage signal acquisition and feature generation include: temperature feature X. T The temperature sensor output, after input protection, enters the operational amplifier buffer; then, it undergoes a second-order low-pass filter to extract the mean value V. T,avg The temperature rise trend V is approximated by a high-pass filter. T,tr And it is limited to avoid noise amplification. Then it is fused through a resistor-weighted network:
[0035]
[0036] Finally, through proportional amplification or attenuation and clamping circuitry, Limited to 0–1V; vibration characteristic X V The vibration sensor signal, after being buffered by the input protection operational amplifier, enters the bandpass filter to select the key mechanical frequency band of the equipment; then, after precision rectification and low-pass filtering, the envelope amplitude V is obtained. env V is obtained by adding a high-frequency enhancement branch. hf Through weighted fusion:
[0037]
[0038] And obtained after normalization and amplitude limiting Bearing wear is more dependent on Impeller imbalance is more dependent on the energy near the rotational frequency. Or low-frequency weighting; current characteristic X I The current signal can be obtained by a shunt resistor and differential amplification. The V signal is then calculated using an RMSD (root mean square) calculation circuit. I,rms The ripple is extracted by a high-pass filter and then rectified by a low-pass filter to obtain the ripple index V. I,rb , merged into:
[0039]
[0040] In this example, the intermediate layer implements linear weighted summation. Each hidden node uses an operational amplifier inverting summer, and the three inputs are respectively connected through programmable resistors R. T ,R V ,R I Connect to the summation point, feedback resistor R f Gain is generated; bias is formed by the reference voltage through a programmable bias resistor R. b Injection. This allows for the configurability of weights and biases for different channels and nodes.
[0041] In this example, the output of the hidden node enters the activation unit circuit. A dual threshold comparator is used to determine whether the output falls into the low or high range by comparing the relationship between the input and the threshold. A linear operational amplifier in the middle section performs differential amplification, and the output is a linear pure analog voltage. Finally, a circuit composed of analog switches achieves a three-way output selection, so that the output gradually saturates near the high and low levels, and maintains a large slope in the middle region, thereby achieving nonlinear enhancement in the abnormal range and finally forming a nonlinear response that approximates a sigmoid.
[0042] Next, output H from the 5 activated hidden nodes. k,1 ~H k,5 The weighted summation is then performed again using an op-amp summer, and the output bias c is added. k , obtain confidence level Y k It is then mapped to the corresponding 0-1 confidence range after amplitude limiting.
[0043] The threshold alarm output section of this example will use Y. k Input hysteresis comparator and threshold θ k Comparison yields alarm A k The threshold reference voltage is provided by a digital-to-analog converter. Finally, A is determined... k Perform a comprehensive fault diagnosis.
Claims
1. A real-time fault diagnosis circuit for blowers based on analog circuit physical neural networks, characterized in that, include: A) A three-channel sensor input module for acquiring temperature signals, vibration signals, and current signals; B) A three-input feature generation module is used to condition and normalize the temperature signal, vibration signal, and current signal into three input feature voltages. , , ; C) Seven parallel fault detection channels C1~C7, each fault detection channel is... , , As the only three inputs, output the confidence level corresponding to the fault. With / or alarm signals The seven parallel fault discrimination channels correspond to the normal operation state and at least six fault states, respectively. D) Fault comprehensive judgment module, used to determine the fault type and severity based on the output of seven fault judgment channels; E) Parameter setting module, used to write or update parameters such as weight, bias and threshold of each fault discrimination channel. The weight parameter is implemented by a programmable resistor and the threshold parameter is implemented by digital-to-analog conversion.
2. The circuit according to claim 1, characterized in that, The three input characteristic voltages are: This is the temperature characteristic voltage, which characterizes the average temperature and / or the temperature rise trend; The vibration characteristic voltage characterizes the vibration intensity and / or energy at a specific frequency band. The voltage characteristic of the current characterizes the current level and / or ripple characteristics; and , , After normalization, the voltage is limited to a preset range.
3. The circuit according to claim 1, characterized in that, The number of fault discrimination channels is the same as the number of signals to be classified. Each fault discrimination channel includes a hidden layer and an output layer. The hidden layer consists of M hidden nodes and includes a weighted summation unit and a nonlinear activation unit to... , , The hidden node voltage is output after weighted summation and nonlinear mapping.
4. The circuit according to claim 3, characterized in that, The j-th hidden node of the k-th fault detection channel satisfies: Furthermore, the weighted summation is achieved through an operational amplifier inverting summer, with the weights determined by the feedback resistor. With input resistance The ratio is determined by the reference voltage. Formed via resistive injection, satisfying:
5. The circuit according to claim 3, characterized in that, The nonlinear activation circuit is a three-segment nonlinear circuit structure consisting of a dual threshold comparator, a mid-segment linear operational amplifier, and a selection switch, to form a piecewise linear nonlinear mapping, i.e., to realize Hard-Sigmoid, satisfying: Where k is the linear magnification factor of the middle section. The lower threshold reference voltage, This is the upper threshold reference voltage.
6. The circuit according to claim 3, characterized in that, The output layer of the k-th fault detection channel outputs the fault confidence voltage: The normalized confidence level of 0 to 1 was obtained by limiting the amplitude.
7. The circuit according to claim 1, characterized in that, The fault comprehensive judgment module includes: A) Threshold-triggered decision: When > Output the fault alarm at the time. ; B) Composite Fault Decision: When multiple channels simultaneously meet the requirements... > Output the composite fault set in real time, and The faults are sorted by size to indicate their primary and secondary importance.
8. The circuit according to claim 1, characterized in that, The parameter setting module includes a microcontroller and a programmable resistor array. The microcontroller is used to write the weights and bias parameters of each channel, so that different faulty channels correspond to different parameter sets.
9. A real-time fault diagnosis method for a blower based on the circuit described in any one of claims 1-8, characterized in that, include: Collect temperature, vibration, and current signals; generate Three-input characteristic voltage , , The characteristic voltage is input in parallel into seven fault discrimination channels to obtain... ~ The fault type and alarm level are determined based on threshold and maximum value selection. The method has no analog-to-digital conversion process throughout, which has great real-time advantages.
10. The method according to claim 9, characterized in that, The weights, biases, and threshold parameters of each channel of the circuit are obtained through offline training or calibration. The offline training includes: establishing an equivalent digital model consistent with the structure of the analog circuit, and introducing physical constraints corresponding to the actual circuit during the training process to obtain a parameter set that satisfies the circuit's realizability; and having the microcontroller write the parameter code values into the programmable resistor array and threshold reference circuit to complete field deployment and updates; after the parameters are written, low-latency real-time classification of blower faults can be completed using only the analog circuit.