A bearing fault diagnosis method and system based on an analog physical neural network

Abstract

This invention discloses a bearing fault diagnosis method and system based on a simulated physical neural network, comprising: S1, preprocessing the acquired vibration signal to obtain the signal to be diagnosed; S2, executing a hybrid optimization algorithm to obtain the optimal passband parameters and optimal weight parameters; S3, configuring a feature extraction module according to the optimal passband parameters and writing the optimal weight parameters into a classification module; S4, inputting the signal to be diagnosed into the feature extraction module to obtain a simulated feature vector; S5, inputting the simulated feature vector into the classification module to output a classified voltage signal; S6, determining the fault type and outputting the result. This invention achieves optimal adaptation between feature extraction and classification tasks by employing a hybrid optimization algorithm to calculate the ratio of inter-class dispersion to intra-class aggregation of the feature vector, thereby improving diagnostic accuracy and generalization ability. By constructing a full-analog domain signal processing architecture, it eliminates the need for analog-to-digital conversion and digital computation, achieving microwatt-level power consumption and microsecond-level real-time response.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and more specifically to a bearing fault diagnosis method and system based on simulated physical neural networks. Background Technology

[0002] With the rapid development of industrial automation and intelligent manufacturing, condition monitoring and fault diagnosis technologies for rotating machinery are receiving increasing attention. Rolling bearings, as core components of rotating machinery, directly affect the safety and reliability of the entire machine due to their operating status. Currently, fault diagnosis methods based on vibration signal analysis have become the mainstream technology, widely employing an architecture of "sensor + data acquisition card + digital processor." This architecture first converts analog vibration signals into digital signals using an analog-to-digital converter, and then utilizes a digital signal processor, field-programmable gate array (FPGA), or computer to run machine learning or deep learning algorithms to complete fault identification.

[0003] However, the aforementioned digital processing methods based on the von Neumann architecture have several limitations in industrial field applications oriented towards edge computing. First, bearing fault characteristics often contain high-frequency components. According to the Nyquist sampling theorem, extremely high sampling rates are required to effectively retain fault information. This not only places stringent demands on the bandwidth and accuracy of analog-to-digital converters but also generates massive amounts of raw data, placing enormous pressure on data transmission bandwidth and storage space. Second, digital computing is inherently discrete and serial; data must be stored before processing. Complex neural network models rely on numerous multiply-accumulate instruction cycles, leading to unavoidable computational delays and making it difficult to meet the millisecond or even microsecond-level real-time response requirements of high-speed rotating equipment to sudden faults. Third, to maintain high-frequency sampling and complex floating-point operations, digital processors require high-frequency clock drives, typically resulting in watt-level power consumption. For wireless sensor nodes relying on battery power or energy harvesting technologies, this power consumption level severely restricts the deployment lifespan of monitoring nodes and the widespread application of edge devices.

[0004] To overcome the power consumption and latency bottlenecks of digital computing, some research attempts to migrate signal processing and computation tasks to the analog domain, directly performing calculations using the physical properties of voltage and current. Theoretically, this offers advantages such as extremely low power consumption and infinite time resolution. For example, patent CN120508813B discloses a bearing fault diagnosis method and system based on a multi-channel analog filter feature network. This scheme uses multiple parallel analog bandpass filter channels to extract the root mean square value of the vibration signal as a feature, and performs fault classification through an analog neural network classifier. Simultaneously, it utilizes a particle swarm optimization algorithm to jointly optimize the filter passband parameters and classifier weights. This scheme achieves signal processing and classification inference in the analog domain to a certain extent, exhibiting low power consumption and high real-time performance.

[0005] However, existing analog circuit solutions still have significant shortcomings: First, the solution disclosed in CN120508813B adopts an offline training and online inference working mode, meaning that filter parameters and classifier weights need to be pre-optimized on the computer and configured in the hardware. It cannot achieve online adaptive updates based on changes in operating conditions or performance degradation during equipment operation, and lacks lifelong learning capabilities. Second, the classifier in this solution only performs forward inference and does not build a hardware loop for error backpropagation, making it impossible to complete online gradient descent training of weights in the analog domain. Third, the hardware size of the analog neural network classifier in this solution is fixed, lacking modular expansion interfaces, making it difficult to flexibly adapt to the needs of different feature dimensions or classification tasks. Fourth, although the feature extraction module and classification module in the existing solution are jointly optimized, the optimized parameters are fixed once configured, and cannot respond to changes in signal distribution caused by equipment aging or environmental drift. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, the present invention provides a bearing fault diagnosis method and system based on a simulated physical neural network.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a bearing fault diagnosis method based on a simulated physical neural network, comprising:

[0008] S1. Collect vibration signals and perform preprocessing to obtain the signal to be diagnosed;

[0009] S2. Execute a hybrid optimization algorithm, using the passband parameters of the feature extraction module and the weight parameters of the classification module as variables to be optimized for collaborative optimization, to obtain the optimal passband parameters and the optimal weight parameters. The feature extraction module consists of multiple parallel-configured analog bandpass filter channels, and the classification module consists of an analog neural network circuit.

[0010] S3. Configure the feature extraction module according to the optimal bandwidth parameters, and write the optimal weight parameters into the classification module;

[0011] S4. Input the signal to be diagnosed into the configured feature extraction module to extract the simulated feature vector;

[0012] S5. Input the simulated feature vector into the weighted classification module, close the feedback loop to perform online gradient descent training, and output the classification voltage signal.

[0013] S6. Determine the bearing fault type based on the classified voltage signal and output the diagnostic results.

[0014] Preferably, S1 specifically involves: performing high-pass filtering on the acquired vibration signal to remove the DC component and obtain an AC vibration signal; performing amplitude normalization processing on the AC vibration signal to obtain a standardized vibration signal; and dividing the standardized vibration signal into segments according to a fixed time window to obtain the signal to be diagnosed.

[0015] Preferably, S2 specifically comprises: S21, initializing the population based on the frequency distribution range of bearing fault characteristics, with each individual in the population corresponding to a set of passband parameters from the feature extraction module and weight parameters from the classification module; S22, introducing a nonlinear convergence factor to adjust the search step size, and integrating the individual's historical best guidance mechanism and the global best guidance mechanism in the position update, iteratively updating the passband parameters and weight parameters of each individual; S23, calculating the fitness function value corresponding to each individual in the current population, wherein the fitness function is defined as the ratio of the inter-class dispersion to the intra-class aggregation of the simulated feature vector obtained after processing by the feature extraction module for samples of different fault categories; S24, iteratively executing S22 to S23 until the fitness function value converges, and outputting the passband parameters and weight parameters corresponding to the individual with the largest fitness function value as the optimal passband parameters and optimal weight parameters, respectively.

[0016] Preferably, the convergence of the fitness function value specifically involves: recording the fitness function value of the globally optimal individual in each iteration; calculating the change range of the fitness function value of the globally optimal individual in multiple consecutive iterations; and determining that the fitness function value has converged when the change range is less than a preset threshold.

[0017] Preferably, configuring the feature extraction module according to the optimal passband parameters specifically involves: converting the optimal passband parameters into digital control words and sending them to the programmable analog filter in the feature extraction module to set the passband of each filter channel;

[0018] The step of writing the optimal weight parameters into the classification module specifically involves writing the optimal weight parameters into the lossy integrator in the classification module in the form of analog voltage, so that the weight parameters are stored in the form of charge across the feedback capacitor of the lossy integrator.

[0019] Preferably, S4 specifically involves: inputting the signal to be diagnosed into the analog bandpass filter channel in the configured feature extraction module, with each channel outputting a filtered signal within the corresponding frequency band based on the optimal passband parameters; calculating the root mean square value of the filtered signal output by each channel to obtain multiple analog feature values; and combining the multiple analog feature values ​​into an analog feature vector.

[0020] Preferably, the simulated neural network circuit comprises an array of simulated multipliers, a weighted aggregation circuit, and a lossy integrator; S5 specifically involves: inputting the simulated feature vector and the weight parameters written into the classification module in parallel into the simulated multiplier array, obtaining a weighted result through simulated multiplication; inputting the weighted result into the weighted aggregation circuit, completing the calculation of the weighted sum of neurons through the summing topology constructed by the operational amplifier, and outputting a predicted voltage signal; performing differential processing on the predicted voltage signal and the target voltage signal to calculate an error signal; performing multiplication processing on the error signal and the simulated feature vector to obtain a weight gradient; inputting the weight gradient into the lossy integrator, updating the weight parameters stored in the classification module through integration, forming a closed-loop feedback loop and outputting a classification voltage signal.

[0021] Preferably, the classification module is provided with a feedforward aggregation interface and a gradient feedback interface; the feedforward aggregation interface is used to receive the intermediate aggregation voltage signal output by the previous stage circuit board and sum it with the weighted result output by the weighted aggregation circuit in the classification module to expand the feature dimension; the gradient feedback interface is used to receive the gradient component signal output by the next stage circuit board and couple it with the gradient signal in the classification module to expand the number of classifications; through the feedforward aggregation interface and the gradient feedback interface, multiple physical neural network units are connected in a cascaded manner to form a scalable modular analog neural network architecture.

[0022] Preferably, S6 specifically involves: calculating the error energy between the classified voltage signal and the target voltage corresponding to each of the preset fault types to obtain multiple error energy values; taking the fault type corresponding to the smallest error energy value among the multiple error energy values ​​as the initial diagnosis result; calculating the ratio of the smallest error energy value to the second smallest error energy value to obtain a relative confidence index; and when the relative confidence index is greater than a preset threshold, taking the initial diagnosis result as the final fault type and outputting it.

[0023] The present invention also provides a bearing fault diagnosis system based on a simulated physical neural network, comprising: a feature extraction module, which consists of multiple parallel simulated bandpass filter channels, for receiving the signal to be diagnosed and extracting simulated feature vectors;

[0024] The classification module, consisting of an analog neural network circuit composed of an analog multiplier array, a weighted aggregation circuit, and a lossy integrator, is used to receive the analog feature vector, perform online gradient descent training, and output a classification voltage signal.

[0025] The result output module is used to determine the bearing fault type based on the classified voltage signal and output the result.

[0026] The technical effects and advantages of this invention are as follows:

[0027] A full analog domain signal processing architecture is constructed by a feature extraction module consisting of parallel analog bandpass filter channels and a classification module consisting of an analog multiplier array, a weighted aggregation circuit, and a lossy integrator. This architecture eliminates the need for analog-to-digital conversion and digital computation, achieving microwatt-level power consumption and microsecond-level real-time response, thus meeting the needs of long-term monitoring and instantaneous diagnosis of sudden faults in industrial settings.

[0028] By coupling the feedforward inference part and the error backpropagation part through a lossy integrator to form a closed-loop feedback loop, an analog domain error backpropagation loop is constructed. The weights are stored in the form of charge in the integrating capacitor and continuously updated by the gradient signal, forming a clockless adaptive closed-loop system. It can adaptively track changes in signal distribution caused by equipment aging, operating condition drift or environmental changes, and has online gradient descent training and lifelong learning capabilities.

[0029] By employing a hybrid optimization algorithm to collaboratively optimize the bandwidth parameters of the feature extraction module and the weight parameters of the classification module, and using the ratio of inter-class dispersion to intra-class aggregation of the feature vector as the fitness function, the optimal fit between feature extraction and classification tasks is achieved, thereby improving diagnostic accuracy and generalization ability.

[0030] By setting up feedforward aggregation interface and gradient feedback interface in the classification module, multiple circuit boards can be cascaded and expanded, flexibly adapting to the needs of different feature dimensions and classification tasks, and overcoming the limitation of fixed scale of simulation hardware. Attached Figure Description

[0031] Figure 1 This is a flowchart of the bearing fault diagnosis method based on simulated physical neural networks proposed in this invention. Detailed Implementation

[0032] This invention provides a bearing fault diagnosis system based on a simulated physical neural network, including a feature extraction module, a classification module, and a result output module. The feature extraction module consists of multiple parallel simulated bandpass filter channels, used to receive the signal to be diagnosed and extract simulated feature vectors. The classification module is an analog neural network circuit composed of an analog multiplier array, a weighted aggregation circuit, and a lossy integrator. It receives the analog feature vector, performs online gradient descent training, and outputs a classification voltage signal. The analog multiplier array and weighted aggregation circuit constitute a feedforward inference unit, responsible for multiplying and adding the input feature vector with the weight parameters to output a predicted voltage signal, thus realizing the forward inference function. The lossy integrator constitutes a parameter in-situ update unit. A feedback capacitor stores the weight parameters (in charge form), and a bleed resistor introduces a forgetting mechanism and suppresses noise. The lossy integrator continuously integrates after receiving the gradient signal, updating the voltage across the capacitor in real time, forming a closed loop for online updating of the weight parameters. The classification module also includes an error backpropagation unit and a modular cascaded interface unit. The error backpropagation unit is mainly responsible for calculating the error between the predicted voltage and the target voltage, multiplying the error signal with the input features to generate a weight gradient, thus realizing the error backpropagation function. The modular cascaded interface unit consists of a feedforward aggregation interface and a gradient backpropagation interface, supporting the cascaded expansion of multiple circuit boards. The feedforward aggregation interface is used to expand the feature dimension, and the gradient backpropagation interface is used to expand the number of classifications. The result output module is used to determine the bearing fault type based on the classified voltage signal and output the result.

[0033] Based on the above system composition, such as Figure 1 This invention also provides a bearing fault diagnosis method based on a simulated physical neural network, specifically including the following steps:

[0034] S1. Collect vibration signals and perform preprocessing to obtain the signal to be diagnosed.

[0035] The original vibration signal of the bearing under operating conditions is collected by a vibration sensor. However, the actual vibration signal usually contains a DC component, which reflects the static offset of the signal and interferes with the subsequent fault feature extraction.

[0036] To address interference, the acquired vibration signals undergo preprocessing: First, a high-pass filter is used to remove the DC component, resulting in an AC vibration signal containing only AC components. Second, the AC vibration signal is input to an amplitude normalization circuit for amplitude normalization. Since the amplitudes of vibration signals acquired under different operating conditions or at different measuring points vary significantly, direct feature extraction would lead to inconsistent feature dimensions, affecting the stability of the diagnostic results. Amplitude normalization scales the signal amplitude to a preset range, eliminating dimensional influence and obtaining a standardized vibration signal. Next, the standardized vibration signal is segmented according to a fixed time window. This is because bearing fault diagnosis requires analyzing signal segments of a certain time length to obtain stable statistical features, and segmented processing enables continuous monitoring. Therefore, the standardized vibration signal is slidably segmented according to a preset time window length to obtain multiple signal segments to be diagnosed. Each segment serves as an independent diagnostic sample for subsequent feature extraction. After the above preprocessing, the signal to be diagnosed is obtained.

[0037] S2. Execute a hybrid optimization algorithm to jointly optimize the passband parameters of the feature extraction module and the weight parameters of the classification module as variables to be optimized, and obtain the optimal passband parameters and the optimal weight parameters; wherein, the feature extraction module is composed of multiple parallel simulated bandpass filter channels, and the classification module is composed of simulated neural network circuits.

[0038] The core of this step lies in collaboratively optimizing the parameters of the feature extraction and classification modules using a hybrid gray wolf-particle swarm optimization algorithm. This ensures that the parameters of the two modules are well-matched, thereby achieving optimal overall system performance. The specific implementation process is as follows:

[0039] First, the optimization population is initialized based on the characteristic frequency distribution range of bearing faults. Since different fault types (such as inner race faults, outer race faults, and ball faults) have their own characteristic frequencies and their harmonics, the characteristic frequency ranges for each type of fault can be pre-calculated based on the bearing's geometric parameters and operating speed. The population is initialized based on this range, with each individual in the population corresponding to a set of variables to be optimized, including the lower and upper cutoff frequencies of each simulated bandpass filter channel in the feature extraction module, and the weight parameters of each neuron node in the classification module.

[0040] Secondly, a nonlinear convergence factor is introduced to adjust the search step size of the Grey Wolf algorithm, and the individual historical optimal guidance mechanism and the global optimal guidance mechanism of the Particle Swarm Optimization (PSO) algorithm are integrated during the position update process to iteratively update the bandwidth parameters and weight parameters of each individual. This hybrid optimization strategy retains the advantages of the Grey Wolf algorithm in the global exploration phase while absorbing the fast convergence characteristics of the PSO algorithm in the local exploration phase. It can effectively avoid the algorithm getting trapped in local optima and efficiently approach the global optimum in a complex parameter space.

[0041] In each iteration, the fitness function value for each individual in the current population is calculated. The fitness function is defined as the ratio of the inter-class dispersion to the intra-class aggregation of the simulated feature vectors obtained after processing by the feature extraction module for samples of different fault categories, i.e., the Fisher discriminant ratio. Specifically, for each set of candidate parameters in the population, training samples are input into the feature extraction module configured by that set of parameters to extract simulated feature vectors. The inter-class distance between the feature centers of each category of samples and the intra-class variance of samples within each category are calculated, and the ratio of these two is used as the fitness value. That is, it is the ratio of the Euclidean distances in the feature space for samples of different fault categories after filtering and RMS processing. This fitness function directly reflects the separability of the feature space, and the optimization objective is to make the feature vectors of different fault categories far apart in space and the feature vectors of the same type of fault cluster together, providing optimal input features for the subsequent classification module. The fitness function... As shown below:

[0042]

[0043] in, It is the number of fault categories. It is the mean center of the characteristic vector of the i-th type of fault. As the global sample center, is the variance of the i-th class of samples; the fitness function aims to maximize the Fisher judgment ratio of the simulated features.

[0044] Finally, the parameter update and fitness calculation steps described above are executed iteratively until the preset number of iterations is reached or the fitness function value converges. The convergence determination method is as follows: record the fitness function value of the globally optimal individual in each iteration, calculate the change range of the fitness function value of this globally optimal individual over multiple consecutive iterations, and determine that the fitness function has converged when the change range is less than a preset threshold, at which point the iteration terminates. After the iteration terminates, the bandwidth parameters and weight parameters corresponding to the individual with the largest fitness function value are output as the optimal bandwidth parameters and optimal weight parameters, respectively, for subsequent configuration and writing.

[0045] Through the above collaborative optimization process, the parameters of the feature extraction module and the classification module are jointly optimized, ensuring that the feature space obtained by feature extraction and the decision boundary of the classification module form the best fit, thereby improving the diagnostic accuracy and generalization ability at the system level.

[0046] S3. Configure the feature extraction module according to the optimal bandwidth parameters, and write the optimal weight parameters into the classification module;

[0047] The core of this step lies in actually deploying the optimal parameters obtained through collaborative optimization into the simulated hardware circuit, ensuring that the feature extraction and classification modules operate at their best, thus providing the hardware foundation for subsequent online diagnosis. The specific implementation process is as follows:

[0048] First, the optimal passband parameters are converted into digital control words and sent to the programmable analog filters in the feature extraction module. Each analog bandpass filter channel in the feature extraction module is typically implemented using a programmable switched-capacitor filter chip. These chips use digital control words to set key parameters such as the filter's center frequency, cutoff frequency, and quality factor. Based on the optimal lower and upper cutoff frequencies obtained through co-optimization, the digital control words corresponding to each filter channel are calculated and written into the corresponding filter chip registers via a serial or parallel communication interface. After configuration, the passband of each filter channel is precisely set to the optimized frequency range, ensuring that subsequent signals can obtain the best characteristic response within this frequency range.

[0049] Secondly, the optimal weight parameters are written into the lossy integrator of the classification module in the form of analog voltage, so that the weight parameters are stored as charges across the feedback capacitor of the lossy integrator. The weight parameters in the classification module are not stored digitally, but directly as analog voltages on the feedback capacitor of the lossy integrator. Specifically, for each weight parameter, the optimal weight value obtained through co-optimization is converted into a corresponding analog voltage signal through a digital-to-analog converter circuit. This voltage signal is applied to the input of the lossy integrator through an analog switch or direct connection. Because the feedback capacitor in the lossy integrator has charge retention characteristics, the voltage across the capacitor is maintained after the external stimulus is removed, thus achieving non-volatile storage of the weights. The lossy integrator consists of an operational amplifier and a parallel-connected feedback capacitor and bleeder resistor. The feedback capacitor realizes the time-domain accumulation of gradient information and weight updates, while the bleeder resistor introduces a forgetting mechanism and suppresses high-frequency noise interference.

[0050] Through the above configuration and writing operations, each analog bandpass filter channel of the feature extraction module is set to its optimal passband, and each weight parameter of the classification module is stored in the lossy integrator capacitor in the form of analog charge. This step lays the hardware foundation for the subsequent feature extraction in S4 and online gradient descent training in S5, ensuring that the system performs diagnostic tasks under optimal parameter conditions, and at the same time provides a physical carrier for the continuous updating of weights during online learning.

[0051] S4. Input the signal to be diagnosed into the configured feature extraction module to extract the simulated feature vector.

[0052] Because bearing vibration signals are essentially signals with specific frequency components excited by different fault sources, and different fault types will generate characteristic frequencies and their harmonics in the frequency spectrum, the core of this step lies in using a feature extraction module with pre-configured optimal passband parameters to perform frequency band decomposition and feature extraction on the signal to be diagnosed. This transforms the original vibration signal into a highly discriminative analog feature vector, providing high-quality input for the subsequent classification module. The specific implementation process is as follows:

[0053] First, the signal to be diagnosed is input into multiple parallel analog bandpass filter channels within the configured feature extraction module. The feature extraction module consists of multiple independently adjustable analog bandpass filter channels operating in parallel, with the passband parameters of each channel configured according to the optimal passband parameters obtained through collaborative optimization. The signal to be diagnosed is simultaneously input to all filter channels, and each channel filters the input signal, allowing only frequency components within its passband to pass, suppressing out-of-band noise and interference. Since the passband of each channel is set according to the optimization results to effectively distinguish characteristic frequency bands of different fault types, the filtered signals output by each channel highlight the sensitive frequency components corresponding to different fault types.

[0054] Secondly, the root mean square (RMS) value of the filtered signal output from each channel is calculated to obtain multiple analog characteristic values. The RMS value reflects the energy level of the signal. For bearing vibration signals, the energy distribution of different fault types varies significantly across different frequency bands. Therefore, the RMS value of the filtered signal from each channel constitutes an effective feature for distinguishing fault types. The RMS value is calculated using a true RMS conversion circuit, which converts the AC filtered signal into a DC voltage signal. The magnitude of this voltage value is linearly related to the RMS value of the input signal, reflecting real-time changes in signal energy. Each filter channel corresponds to a RMS value calculation unit, and these units operate in parallel, synchronously outputting multiple analog characteristic values.

[0055] Finally, the multiple simulated feature values ​​are combined into a simulated feature vector. The multiple simulated feature values ​​output by the feature extraction module physically represent multiple independent voltage signals. These voltage signals are aggregated into a simulated feature vector, which serves as the input to the classification module. The dimension of this simulated feature vector is consistent with the number of filter channels. The voltage value of each dimension represents the signal energy within the corresponding frequency band, collectively forming a representation of the energy distribution of the signal to be diagnosed in the optimized frequency band space.

[0056] Through the above feature extraction process, the signal to be diagnosed is transformed into a simulated feature vector reflecting the fault characteristics. Since the bandwidth parameters of the feature extraction module have been optimally configured through collaborative optimization, this feature vector exhibits the characteristics of clustering similar samples and dispersing dissimilar samples in the feature space.

[0057] S5. Input the simulated feature vector into the weighted classification module, close the feedback loop to perform online gradient descent training, and output the classification voltage signal.

[0058] The classification module consists of an analog multiplier array, a weighted aggregation circuit, and a lossy integrator. The analog multiplier array performs multiplication of input features and weights, the weighted aggregation circuit performs summation of neuron weights, and the lossy integrator stores weights and updates gradients. The analog multiplier array and the weighted aggregation circuit form the feedforward inference section, while the error backpropagation section and the lossy integrator form the parameter update loop. The two parts are coupled through the lossy integrator to form a closed-loop feedback.

[0059] The main processes include the following:

[0060] Feedforward inference process: The simulated feature vector extracted by S4 and the weight parameters already written into the classification module are input in parallel into the simulated multiplier array. The simulated multiplier array consists of multiple four-quadrant simulated multipliers. Each multiplier receives one simulated feature signal and the corresponding weight voltage signal, performs scalar multiplication, and outputs the weighted result. The output current signals of each multiplier are collected into a weighted aggregation circuit. The weighted aggregation circuit uses operational amplifiers to construct a multi-channel in-phase summing topology and is equipped with a precision resistor feedback network. It sums the input weighted results, completes the calculation of the neuron weighted sum, and outputs a predicted voltage signal. This predicted voltage signal reflects the preliminary judgment of the fault category of the input signal under the current network parameters.

[0061] This process is designed based on Kirchhoff's laws for the neuron activation part, and the feedforward path aims to physically implement the activation function of a linear perceptron. For the i-th neuron of class j, its mathematical model is as follows:

[0062]

[0063] in, The neuron representing class j outputs the predicted result. Represents the weight parameter. Represents input features, This represents the bias parameter.

[0064] The aforementioned four-quadrant analog multiplier can perform analog multiplication calculations in parallel. This calculation can be performed using the AD633 analog multiplier chip. Since the transfer function of actual analog devices has a scaling factor, and the transfer function of the AD633 is as follows:

[0065]

[0066] in, This represents the output voltage of the AD633 chip; , These represent the first set of differential input voltages of the multiplier, i.e., the input characteristics; , These represent the second set of differential input voltages of the multiplier, i.e., the weighting parameters.

[0067] Its scaling factor Since it is 0.1, the physical output of the neuron is ,in This represents the physical output signal of the synapse generated by a single multiplier. Represents the weight parameter. Represents input features.

[0068] This invention designs a weighted aggregation and gain compensation module for neurons to eliminate the compression of the signal dynamic range by the multiplier scaling factor and achieve weighted summation. This design utilizes an operational amplifier chip to construct a multi-channel in-phase summation topology. The system uses the TL084IDR chip as the operational amplifier, and current summation is implemented at the operational amplifier input node. To recover the signal amplitude, a precision resistor feedback network is configured, and the closed-loop gain is set as follows: .

[0069] To achieve inverse mapping compensation of the transfer function:

[0070]

[0071] in: This represents the final summation output voltage of the weighted aggregation circuit. This represents the closed-loop gain of the operational amplifier summing circuit. This represents the output voltage of a single analog multiplier. This represents the fixed scaling factor of the analog multiplier. , These represent the weight parameters and input feature values ​​for the corresponding single path, respectively.

[0072] The above process is designed as a pooling layer, which physically and accurately replicates the linear weighting function of neurons. Following the pooling layer, a cascaded unity-gain adder circuit serves as a bias injection layer, introducing a bias voltage. The translation control of the classification hyperplane is achieved through the principle of voltage superposition, and the predicted voltage is ultimately output. : ,in: This represents the final predicted output voltage signal of the neuron. , These represent the weighted parameter voltage signal and the input characteristic voltage signal, respectively. This represents the bias parameter voltage signal.

[0073] Error backpropagation process: The predicted voltage signal and the preset target voltage signal are differentially processed to calculate the error signal. The target voltage signal is preset according to the actual fault type of the input sample, reflecting the expected classification output. The differential processing is completed in real time by a precision subtraction circuit constructed from operational amplifiers, and the output error signal represents the deviation between the current prediction and the expectation. Subsequently, the error signal is multiplied with the simulated feature vector, and the multiplication operation is completed by the simulated multiplier array of the reverse path to obtain the parameter gradient. This gradient signal reflects the direction and magnitude of the partial derivatives of the loss function with respect to the current weight parameters and the current bias parameters, and is the core basis for parameter updates.

[0074] This process employs the physical generation of gradient flow, aiming to calculate the loss function in real time via hardware circuitry. : ,in: , These represent the predicted output voltage signal and the target voltage signal, respectively.

[0075] Calculate the gradient of the loss function with respect to the network parameters, i.e.: and ,in, This represents the partial derivative of the loss function with respect to the weight parameters. This represents the partial derivative of the loss function with respect to the bias parameters. Represents the real-time error signal. This represents the input characteristic voltage signal.

[0076] The specific design involves first designing a differential error dynamics unit, then using an operational amplifier to construct a precision subtraction circuit that responds to the output voltage in real time. With target voltage The difference is used to generate an instantaneous error signal. : .

[0077] To avoid the complex gradient calculation network generated in subsequent stages from loading the error signal and causing a decrease in calculation accuracy, an impedance buffer and signal isolation stage based on a TL084IDR voltage follower is configured at the error output. This unit provides high input impedance. With low output impedance ,Right now and This ensures the signal integrity of the error gradient flow.

[0078] The buffered instantaneous error signal The instantaneous error signal is input to the multiplier array in the reverse path. Compared with the original input features Perform simulated multiplication to physically calculate the projection of the local gradient: ,in: Represents the local gradient. Represents the instantaneous error signal. This represents the original input features. The signal then passes through this stage of the compensation circuit to obtain the true weight gradient. .

[0079] Parameter update process: The lossy integrator consists of an operational amplifier, a feedback capacitor, and a bleeder resistor. The voltage stored across the feedback capacitor is the current weight value. The weight gradient signal is input to the lossy integrator, and the voltage value stored on the feedback capacitor is updated through integration, thus achieving online weight adjustment. The feedback capacitor accumulates the gradient signal in the time domain, while the bleeder resistor introduces a forgetting mechanism and suppresses high-frequency noise interference. The output of the integrator is directly connected to the weight input of the analog multiplier in the feedforward inference section, forming an adaptive closed-loop control system without digital clock drive. When the input characteristics change or errors occur, the weights are adjusted immediately and continuously until the error signal converges to a preset threshold range.

[0080] It is worth noting that this invention designs the calculation and storage of parameters as an integrated storage and computation structure, and the system updates the weight parameters. Instead of being stored digitally, the charge is directly stored across the feedback capacitor C of the integrator as analog charge. The input to the integrator is the gradient signal of the parameters. Its time-domain dynamic equation is: ;in The values ​​of the system's weight parameters at time t. That is, the learning rate. , Represents the initial weight value. The infinitesimal time variable representing the integral.

[0081] This invention specifically introduces a bleed resistor connected in parallel with the capacitor. This constitutes a lossy integral topology. This not only prevents integral saturation but also introduces a forgetting mechanism. In the frequency domain, this is equivalent to a low-pass filter, effectively suppressing high-frequency thermal noise in the circuit and utilizing noise-assisted stochastic resonance to help the system escape local minima, thus improving the robustness of optimization.

[0082] In this invention, the classification module is equipped with a feedforward aggregation interface and a gradient feedback interface to support multi-level cascaded expansion. The feedforward aggregation interface receives the intermediate aggregation voltage signal output from the previous stage circuit board and sums it with the weighted result output from the current stage weighted aggregation circuit, thereby expanding the feature dimensions. When multiple circuit boards are cascaded, a portion of the sum and voltage signal from the previous stage circuit board can be connected to this node, utilizing the virtual ground characteristic of the operational amplifier to achieve cross-board voltage summation, i.e.: ,in: This represents the total weighted aggregate voltage at the virtual ground node after multiple circuit boards are cascaded. and These represent the weighted aggregate voltage of the previous circuit board and the weighted aggregate voltage of the current circuit board, respectively.

[0083] The gradient backpropagation interface receives the gradient component signals output from the next-level circuit board and couples them with the gradient signal of the current level, enabling an expansion of the number of classifications. Through this interface, multiple physical neural network units are cascaded to form a scalable, modular analog neural network architecture that can flexibly adapt to the needs of different feature dimensions and classification tasks. This interface allows the instantaneous error signal generated by the final-level circuit board to be processed. The gradient calculation unit is fed back to the front-end circuit board to ensure that all weight parameters distributed on different circuit boards can be updated according to a unified global loss function.

[0084] Through the continuous operation of the aforementioned closed-loop feedback circuit, the classification module performs online gradient descent training of the weights while executing forward inference, outputting a classification voltage signal reflecting the fault classification result for subsequent steps to determine the fault type. This process is completed entirely in the analog domain, without the need for digital clock driving, achieving low-power, high-real-time online adaptive diagnosis.

[0085] S6. Determine and output the bearing fault type based on the classified voltage signal. The determination process is as follows:

[0086] First, the error energy between the classified voltage signal and the preset target voltages corresponding to each type of fault is calculated, resulting in multiple error energy values. The classified voltage signal output by the classification module is a continuous analog voltage value, the magnitude of which reflects the proximity of the current input sample to the feature space of each type of fault. For each fault type, a target voltage value is preset, which is calibrated according to the physical characteristics of the fault category and the circuit operating range. For example, a low level corresponds to a normal state, a second-low level to an inner ring fault, a second-high level to an outer ring fault, and a high level to a ball bearing fault. Sufficient voltage margin is maintained between adjacent categories to form clear classification boundaries. The error energy between the classified voltage signal and each target voltage is calculated, typically using Euclidean distance or absolute difference as the error energy metric. The smaller the error energy, the closer the classified voltage signal is to the expected output of the corresponding fault type.

[0087] Secondly, the fault category corresponding to the smallest error energy value among the multiple error energy values ​​is taken as the initial diagnostic result. Since the classification module has converged to the optimal parameter state during online gradient descent training, for the test sample, the classified voltage signal should be closest to the target voltage corresponding to the true fault category of the sample. Therefore, the fault category corresponding to the smallest error energy is the preliminary result of the network's determination.

[0088] Finally, the ratio of the minimum error energy value to the second smallest error energy value is calculated to obtain the relative confidence index. When the relative confidence index is greater than a preset threshold, the initial diagnosis result is taken as the final fault type and output. The relative confidence index is defined as the ratio of the minimum error energy to the second smallest error energy. This index reflects the certainty of the classification decision: when the minimum error energy is much smaller than the second smallest error energy, the ratio approaches zero, indicating that the decision has high confidence; conversely, if the two are close, the ratio approaches 1, indicating that the classification voltage signal is in the ambiguous region of the classification boundary. The system presets a threshold, for example, 0.5. When the relative confidence index is less than this threshold, it indicates that the minimum error energy is significantly lower than the second smallest error energy, the classification decision has high confidence, and the initial diagnosis result is output as the final fault type; when the relative confidence index is greater than or equal to the threshold, it indicates that the classification voltage signal is in the ambiguous region of the boundary, the system determines that the diagnosis is uncertain, and can output a warning signal or require re-collection of samples for secondary diagnosis.

[0089] By employing the dual judgment mechanism of the aforementioned minimum error energy criterion and relative confidence index, this invention effectively shields against boundary misjudgments caused by environmental noise, non-ideal circuit characteristics, or extremely degraded samples, ensuring the reliability and robustness of the diagnostic results. The final output fault type can be digitally encoded into indicator lights, displays, or communication interface signals for use by on-site maintenance personnel or higher-level monitoring systems.

[0090] The above method will be explained below with reference to specific embodiments:

[0091] Example 1

[0092] This embodiment verifies the effectiveness of the above method based on the standard public dataset provided by the Bearing Data Center of Case Western Reserve University (CWRU).

[0093] (1) Selection of experimental data and signal generation

[0094] Vibration data from the drive-end bearing, sampled at 12 kHz, were selected from the CWRU dataset. Four typical operating states were chosen as classification targets:

[0095] Normal state: Motor load 0HP, speed 1797rpm.

[0096] Inner Race Fault: Fault diameter 0.007 inches.

[0097] OuterRaceFault: Fault diameter 0.007 inches, fault location at 6 o'clock.

[0098] Ball fault: fault diameter 0.007 inches.

[0099] This invention simulates a hardware system. In the experiment, an overlap sampling method was used to slice the original vibration signal, with a time window length of 2048 points and an overlap rate of 50%. A total of 400 training samples and 100 test samples were generated. Then, the NIUSB-6289 data acquisition card was used to convert the vibration signal into a continuous analog voltage signal, which served as the system's input signal.

[0100] (2) Parameter optimization and hardware configuration

[0101] Before inputting the signal into the analog neural network, a front-end analog filtering network is constructed and its parameters are adaptively optimized. This invention constructs an analog feature extraction network with four channels and uses an optimization algorithm to obtain the optimal cutoff frequency. Then, the RMS value of the signal is extracted to achieve feature extraction. The hardware uses a MAX261 microprocessor-based programmable general-purpose switched-capacitor filter chip, with each channel configured as a fourth-order Butterworth bandpass filter. Subsequently, the filtered signal is input to an AD637 true RMS converter chip to convert the AC signal into a DC eigenvector voltage value. The specific steps are as follows:

[0102] Step 1: Calculate the fault characteristic frequency and sampling theorem based on the bearing's geometric parameters, and initialize the center frequency of the MAX261. The quality factor Q is set to ensure coverage of the low-frequency, mid-frequency, and high-frequency ranges. The optimization variable is set as the lower cutoff frequency of the four bandpass filter channels. and bandwidth Construct the particle swarm position vector: The spatial ranges are as follows: , .

[0103] Step 2: Filter parameter optimization based on the H-GWO-PSO algorithm

[0104] To maximize the intra-class cohesion and inter-class dispersion of the signal, parameter optimization based on the H-GWO-PSO algorithm is performed during the training phase:

[0105] This invention uses the feature space Fisher ratio as the fitness function instead of directly using accuracy, aiming to maximize the inter-class scatter matrix. With the intra-class scatter matrix The ratio of traces, that is: ;

[0106] in, Let represent the fitness value of the k-th particle in the particle swarm optimization algorithm, and let represent the inter-class discrete matrix. : The intra-class scatter matrix is : ,in: Let be the four-dimensional RMS feature vector of the j-th sample in class c after processing by the parameters of the current filter. The total number of fault categories. N is the sample index within the c-th type of fault. c Let be the total number of samples for type c faults. Let be the feature mean vector of all samples of the c-th type of fault. This represents the global average vector for all categories and all samples.

[0107] Step 3: Perform H-GWO-PSO algorithm optimization and hardware configuration

[0108] The H-GWO-PSO algorithm is run in the computation, iterated 100 times, automatically simulates and calculates the fitness value of the CWRU training set under different parameters, locks a set of optimal parameters, and writes this set of parameters into the MAX261 filter array of the corresponding hardware circuit through SPI communication to complete the hardware configuration.

[0109] (3) Simulated neural network online training

[0110] Step 1: Initialize the connection of the analog circuit system

[0111] The training process at this stage requires no computer intervention and is conducted entirely in a closed loop within the on-chip analog circuit. First, the integrator capacitor is short-circuited and reset, discharging it to the reference zero potential, thereby assigning initial weights to the neural network.

[0112] Next, by closing the error feedback loop using a physical jumper cap, the output of the subtractor module is directly connected to the input of the inverting multiplier array. Finally, a signal generator adaptively sets the target voltage for different categories of training samples.

[0113] This invention abandons the conventional discrete logic labels in digital networks, and instead performs multi-level analog level calibration based on the "energy gradient" and "feature space distance" of the bearing vibration signal. For example:

[0114] When the input is a normal sample, set the target output vector. When the input is an inner ring fault sample, the target output voltage is set to... .

[0115] The arrangement of the target voltages described above essentially maps the high-dimensional physical feature space of the bearing onto a one-dimensional voltage coordinate axis through the circuit. By maintaining a constant classification margin between adjacent categories, not only is the tolerance of the analog circuit to feature fluctuations of similar samples maximized, but the initial configuration of the analog training architecture is also completed.

[0116] Step 2: Perform closed-loop gradient calculation for training until the system reaches steady state.

[0117] Turn on the signal generator, start the NIUSB-6289 data acquisition card to input the CWRU dataset signal. After the signal passes through the MAX261 array and AD637 module, a characteristic voltage signal is generated and input into the neural network circuit. In the initial stage, the network parameters have not been trained, and the network... and The significant differences between the parameters result in a high-amplitude error signal. This error drives the inverting circuit to generate a gradient current, which charges the integrating capacitor. After 200ms, the error signal amplitude decays exponentially and stabilizes within the noise floor. At this point, the voltage across the integrator capacitor no longer changes significantly, and the parameters have converged to the optimal solution.

[0118] (4) Testing and performance verification of quasi-neural networks

[0119] Relying on the charge retention characteristics of the feedback capacitor, the trained parameters are stored, and the NIUSB-6289 data acquisition card is started to input the CWRU dataset test signal. To illustrate the system performance, the NIUSB-6289 data acquisition card is used to acquire the output node voltage of the test signal, and a robust decision algorithm based on minimum error energy is designed.

[0120] For the output voltage vector obtained from actual measurement in, and Let each of the two output nodes represent the actual voltage components obtained from actual measurements, and define their error energy relative to the k-th class center. for: ,in This represents the target voltage vector pre-defined by the system for the k-th type of fault. and These represent two specific reference voltage components in the target voltage vector of the kth class.

[0121] The error energy between the actual tested node voltage and the target set voltage is used to calculate the final classification result of the system. The category with the smallest error energy is determined as follows: .

[0122] In addition, to quantify the reliability of the classification, the Relative Confidence Index (RCI) is introduced: ,in To minimize error energy, For the second smallest energy error; when Only when the time is right is the diagnosis considered correct. In this case, the threshold is designed to be [value missing]. This mechanism effectively shields against boundary misjudgment caused by environmental interference or extremely degraded samples.

[0123] To further illustrate the inference and diagnostic mechanism based on single-ended analog voltage, let's assume that during a certain time period, the system receives an unknown bearing vibration signal. After this signal passes through the analog forward inference module, the actual voltage value read at the output node is... .

[0124] The system compares the output voltage with the target set voltages for the two categories respectively, and calculates the error energy corresponding to each category:

[0125] Error energy compared to normal state: ;

[0126] Error energy relative to the inner circle state: ;

[0127] The system compares the two calculation results above and finds that the minimum error energy is Therefore, the output voltage is determined to be spatially closest to the target value of the inner fault; to prevent the signal from being in the ambiguous boundary zone, the system further calculates the relative confidence index. At this point, the minimum error energy is... The second smallest error energy is The relative confidence level is: .

[0128] Clearly, the confidence index of 0.997 for this test sample is far greater than the set judgment threshold, indicating that the 2.7V output of this signal clearly points to an inner race fault, and the probability of it being interfered with by circuit noise or other fault characteristics is extremely low. Ultimately, the system diagnoses and outputs that the bearing is in an "inner race fault" state, completing a highly robust closed-loop diagnosis.

[0129] Example 2

[0130] This embodiment demonstrates cascading training and testing for 4-class classification.

[0131] This embodiment aims to verify the hardware scalability of the present invention when processing multi-classification tasks. By cascading two identical analog neural network computing units (hereinafter referred to as Unit A and Unit B), the simultaneous identification of signals from four different health states in the Case Western Reserve University dataset is achieved.

[0132] (1) Cascaded hardware topology

[0133] The system is designed as a distributed, cascaded architecture with class-specific outputs. It employs two cascaded circuits to extend functionality from binary to four classifications. Unit A is responsible for identifying normal state signals and inner-ring fault signals. The node output is defined as... and ,in and These represent the actual predicted voltages of the two output nodes under the current input, respectively, while Unit B is responsible for identifying outer race fault signals and ball fault signals. The node outputs are defined as follows: and ,in and These represent the actual predicted voltages of the two output nodes under the current input, respectively, which are the responsibility of Unit B. The front-end feature extraction module outputs simulated voltages and connects them to the feature input ports of Unit A and Unit B via terminals. Then, the bias adjustment ports of Unit A and Unit B are connected together via jumper caps to ensure that the two computational units share the same reference potential, so that the four classification hyperplanes are under the same voltage reference system, avoiding misalignment of the classification surface due to ground potential drift.

[0134] (2) Goal setting and data training

[0135] Four state data points at 1797 rpm were selected from the CWRU dataset as the training set. To adapt to the analog hardware architecture of a single output node, this embodiment abandons the one-hot encoding required by traditional digital neural networks and innovatively adopts the "equidistant multi-level level mapping" strategy to set the analog target voltage.

[0136] Taking into account the linear operating range of the operational amplifier module and the inherent noise floor of the analog circuit, this invention uses "maximizing inter-class classification margin" and "avoiding deep saturation of the inverting integrator" as joint optimization conditions, and designs the target signal voltages for the four health states as follows:

[0137] Under normal sample conditions, set the target output vector. Under the inner ring fault sample, the target output voltage is set to Under the ball bearing failure sample, set the target output vector. Under the outer ring fault sample, the target output voltage is set to This design ensures that the voltage difference between any two adjacent states reaches 2V. This significant voltage difference is much larger than the peak-to-peak thermal noise of the analog multiplier and operational amplifier module, enabling accurate high-noise-resistance classification to be completed with only a simple voltage comparator during the testing phase, thus greatly reducing the hardware overhead of the edge computing terminal.

[0138] (3) Distributed parallel training

[0139] Step 1: Gradient Reset

[0140] Remove the jumpers used for weight preservation from the integrator modules on both UnitA and UnitB circuit boards, and close the jumpers used for error feedback propagation loops.

[0141] Step 2: Orthogonalization training and gradient coupling calculation

[0142] The third extended input port in the backpropagation process is used as the input for the backpropagation of the first two types of errors, ensuring that the circuit performs gradient coupling calculation and update of four types of signals during backpropagation. To prevent coupling oscillations in the weight updates of the two independent boards, an orthogonal training method is designed, introducing the local error energy signal generated by UnitA into the extended input of UnitB. This physical connection forms a lateral suppression mechanism in the analog domain, forcing UnitA and UnitB to extract uncorrelated feature components.

[0143] Step 3: Signal Training Process

[0144] The signal generator outputs a training sample sequence that mixes four types of faults. , , , These represent four pre-defined ideal target output components for different fault categories. Unit A is responsible for calculating and... , The errors of the first two target components are used to drive the eight AD633 multipliers on the board to update the weight matrix via backpropagation. UnitB synchronous computing , Based on the errors of the latter two target components, the eight AD633 multipliers on the driver board update the weight matrix. Because the weight parameters are shared and time-synchronized, the two boards are actually performing a parallel gradient descent optimization of a 4×4 weight matrix. The total error of the detection system... for: in, This represents the local error calculated on the Unit A circuit board. This represents the local error calculated synchronously on the Unit B circuit board.

[0145] Actual experimental verification shows that, under the cascaded four-class classification task, the system takes about 350ms to reach a steady state, which is slightly slower than single-board binary classification, but still belongs to the sub-second real-time training system.

[0146] (4) System testing and result judgment

[0147] To verify the generalization inference, test set samples were input into the inference module for verification. The output voltages of the four channels on the oscilloscope were observed, and the signal category was determined using the minimum error energy determination algorithm.

[0148] This system successfully expanded the classification capability from 2 classes to 4 classes by cascading two circuit boards without redesigning the PCB layout, verifying the scalability of the in-situ computing architecture. In the CWRU four-class classification test set, this cascaded system demonstrated that the distributed gradient descent algorithm implemented through a physical interface still exhibits good convergence in multi-task scenarios, validating the advantages of the bearing fault diagnosis method proposed in this invention.

[0149] It should be noted that the specific chip models (MAX261, AD637, AD633, TL084IDR) mentioned in the above embodiments are merely preferred examples of implementation schemes and are not intended to limit the present invention. Those skilled in the art can select other functionally equivalent chips or discrete components to build circuits based on factors such as performance, cost, and availability.

[0150] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A bearing fault diagnosis method based on simulated physical neural networks, characterized in that: include: S1. Collect vibration signals and perform preprocessing to obtain the signal to be diagnosed; S2. Execute a hybrid optimization algorithm, using the passband parameters of the feature extraction module and the weight parameters of the classification module as variables to be optimized for collaborative optimization, to obtain the optimal passband parameters and the optimal weight parameters. The feature extraction module consists of multiple parallel-configured analog bandpass filter channels, and the classification module consists of an analog neural network circuit. S3. Configure the feature extraction module according to the optimal bandwidth parameters, and write the optimal weight parameters into the classification module; S4. Input the signal to be diagnosed into the configured feature extraction module to extract the simulated feature vector; S5. Input the simulated feature vector into the weighted classification module, close the feedback loop to perform online gradient descent training, and output the classification voltage signal. The simulated neural network circuit consists of an array of simulated multipliers, a weighted aggregation circuit, and a lossy integrator. Specifically, S5 involves: inputting the simulated feature vector and the weight parameters written into the classification module in parallel into the simulated multiplier array; obtaining a weighted result through simulated multiplication; inputting the weighted result into the weighted aggregation circuit; calculating the weighted sum of neurons using a summation topology constructed by operational amplifiers; outputting a predicted voltage signal; performing differential processing between the predicted voltage signal and the target voltage signal to calculate an error signal; multiplying the error signal with the simulated feature vector to obtain a weight gradient; inputting the weight gradient into the lossy integrator; updating the weight parameters stored in the classification module through integration operations; forming a closed-loop feedback loop; and outputting a classification voltage signal. S6. Determine the bearing fault type based on the classified voltage signal and output the diagnostic results.

2. The bearing fault diagnosis method based on simulated physical neural networks according to claim 1, characterized in that: S1 specifically involves: performing high-pass filtering on the acquired vibration signal to remove the DC component and obtain an AC vibration signal; performing amplitude normalization processing on the AC vibration signal to obtain a standardized vibration signal; and dividing the standardized vibration signal into segments according to a fixed time window to obtain the signal to be diagnosed.

3. The bearing fault diagnosis method based on simulated physical neural networks according to claim 1, characterized in that: S2 specifically consists of: S21, initializing the population based on the frequency distribution range of bearing fault characteristics, with each individual in the population corresponding to a set of passband parameters from the feature extraction module and weight parameters from the classification module; S22, introducing a nonlinear convergence factor to adjust the search step size, and integrating the individual's historical best guidance mechanism and the global best guidance mechanism in the position update, iteratively updating the passband parameters and weight parameters of each individual; S23, calculating the fitness function value corresponding to each individual in the current population, where the fitness function is defined as the ratio of the inter-class dispersion to the intra-class aggregation of the simulated feature vectors obtained after processing by the feature extraction module for samples of different fault categories; S24, iteratively executing S22 to S23 until the fitness function value converges, and outputting the passband parameters and weight parameters corresponding to the individual with the largest fitness function value as the optimal passband parameters and optimal weight parameters, respectively.

4. The bearing fault diagnosis method based on simulated physical neural networks according to claim 3, characterized in that: The convergence of the fitness function value specifically involves recording the fitness function value of the globally optimal individual in each iteration. Calculate the change range of the fitness function value of the globally optimal individual in multiple consecutive iterations; when the change range is less than a preset threshold, it is determined that the fitness function value has converged.

5. The bearing fault diagnosis method based on simulated physical neural networks according to claim 1, characterized in that: The specific steps of configuring the feature extraction module according to the optimal passband parameters are as follows: converting the optimal passband parameters into digital control words and sending them to the programmable analog filter in the feature extraction module to set the passband of each filter channel; The step of writing the optimal weight parameters into the classification module specifically involves writing the optimal weight parameters into the lossy integrator in the classification module in the form of analog voltage, so that the weight parameters are stored in the form of charge across the feedback capacitor of the lossy integrator.

6. The bearing fault diagnosis method based on simulated physical neural networks according to claim 1, characterized in that: S4 specifically involves: inputting the signal to be diagnosed into the analog bandpass filter channel in the configured feature extraction module; each channel outputs a filtered signal within the corresponding frequency band based on the optimal passband parameters; calculating the root mean square value of the filtered signal output by each channel to obtain multiple analog feature values; and combining the multiple analog feature values ​​into an analog feature vector.

7. The bearing fault diagnosis method based on simulated physical neural networks according to claim 1, characterized in that: The classification module is equipped with a feedforward aggregation interface and a gradient feedback interface. The feedforward aggregation interface receives the intermediate aggregation voltage signal output from the previous stage circuit board and sums it with the weighted result output from the weighted aggregation circuit in the classification module to expand the feature dimension. The gradient feedback interface receives the gradient component signal output from the next stage circuit board and couples it with the gradient signal in the classification module to expand the number of classifications. Through the feedforward aggregation interface and the gradient feedback interface, multiple physical neural network units are connected in a cascaded manner to form a scalable modular analog neural network architecture.

8. The bearing fault diagnosis method based on simulated physical neural networks according to claim 1, characterized in that: S6 specifically involves: calculating the error energy between the classified voltage signal and the target voltage corresponding to each of the preset fault types to obtain multiple error energy values; and taking the fault type corresponding to the smallest error energy value among the multiple error energy values ​​as the initial diagnostic result. The ratio of the minimum error energy value to the second smallest error energy value is calculated to obtain the relative confidence index. When the relative confidence index is greater than a preset threshold, the initial diagnosis result is taken as the final fault type and output.

9. A bearing fault diagnosis system based on a simulated physical neural network, used to implement the bearing fault diagnosis method based on a simulated physical neural network as described in any one of claims 1-8, characterized in that, include: The feature extraction module consists of multiple parallel analog bandpass filter channels, used to receive the signal to be diagnosed and extract analog feature vectors. The classification module, consisting of an analog neural network circuit composed of an analog multiplier array, a weighted aggregation circuit, and a lossy integrator, is used to receive the analog feature vector, perform online gradient descent training, and output a classification voltage signal. The result output module is used to determine the bearing fault type based on the classified voltage signal and output the result.

Images