Industrial equipment fault early warning method and system based on pulse neural network

Abstract

The application belongs to the technical field of data processing. An industrial equipment fault early warning method and system based on a pulse neural network are provided, vibration time series signals and temperature time series signals of industrial equipment are acquired, the vibration time series signals and the temperature time series signals are input into a pulse neural network optimized by using a quantum annealing algorithm, and output states of each neuron are obtained, a connection weight matrix of each neuron of the pulse neural network is calculated according to the output states of each neuron, a global energy is determined according to the output states of each neuron and the connection weight matrix, when the global energy is lower than a first set threshold value and a maintenance time is greater than a second set threshold value, it is determined that the industrial equipment has a fault and a fault early warning signal is generated, the quantum annealing algorithm in the field of quantum computing is used, the processing capacity of the pulse neural network for long-range dependence of the time series signals of the industrial equipment is improved, and the fault early warning precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an industrial equipment fault early warning method and system based on a spiking neural network. BACKGROUND

[0002] The statements in this section merely provide background technology related to the present application and do not necessarily constitute prior art.

[0003] The neurons in the spiking neural network (SNN, Spiking Neural Network) communicate in continuous time through short pulses called spikes, which is in sharp contrast to the way neurons in artificial neural networks (ANN) exchange real-valued signals in discrete time. However, in the context of industrial equipment fault early warning, the spiking neural network has significant shortcomings when processing long-range dependent tasks, specifically:

[0004] (1) Industrial equipment time series signals have long-range dependence characteristics. The faults of devices such as motors and fans (such as bearing wear and stator winding aging) are often related to minor abnormal signals several hours or days ago (such as the long-term correlation between early minor cracks in bearings and later vibration amplitude). The existing spiking neural network is difficult to capture such cross-long-time-scale features, resulting in reduced accuracy of fault prediction; (2) Industrial site controllers (such as ARM Cortex-M series and PLC modules) have limited computing power and rely on site bus for power supply. Traditional artificial neural networks cannot be deployed due to high energy consumption (single inference cycle power consumption ≥ 50mW), while non-quantum optimized spiking neural networks have low energy consumption but insufficient long-range dependence performance, which cannot meet the reliability requirements of fault early warning. SUMMARY

[0005] To solve the problems of the prior art, the present application provides an industrial equipment fault early warning method and system based on a spiking neural network, which uses quantum annealing algorithm in the field of quantum computing to improve the processing capability of spiking neural network for long-range dependence of industrial equipment time series signals, while retaining the low energy consumption advantage of spiking neural network, adapting to the deployment requirements of industrial edge devices, and improving the fault early warning accuracy.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides an industrial equipment fault early warning method based on a spiking neural network.

[0008] An industrial equipment fault early warning method based on a spiking neural network includes the following processes:

[0009] Obtain the vibration time series signal and temperature time series signal of the industrial equipment;

[0010] The vibration time series signal and the temperature time series signal are input into the spiking neural network optimized by the quantum annealing algorithm to obtain the output state of each neuron;

[0011] The connection weight matrix of each neuron in the spiking neural network is calculated based on the output state of each neuron, and the global energy is determined based on the output state of each neuron and the connection weight matrix.

[0012] When the global energy is lower than the first set threshold and the duration is greater than the second set threshold, the industrial equipment is determined to be faulty and a fault warning signal is generated.

[0013] In one implementation of the first aspect of the present invention, a quantum annealing algorithm is used to train a spiking neural network. The total energy is set as a weighted sum of the global energy and a kinetic energy term with controllable amplitude. By adjusting the coefficients of the kinetic energy term, the objective function continuously jumps from a local minimum to a global minimum, thus obtaining the latest trained function. .

[0014] As a further limitation of the first aspect of the present invention, during the training of the spiking neural network, the connection weight matrix... Data elements for: ,in, Representative failure mode Next The output state of each neuron Representative failure mode Next The output state of each neuron.

[0015] As a further limitation of the first aspect of the invention, determining the global energy based on the output state of each neuron and the connection weight matrix includes: ,in, Representing the The output state of each neuron No. The output state of each neuron.

[0016] As a further limitation of the first aspect of the invention, the total energy is a weighted sum of the global energy and the kinetic energy term with controllable amplitude, including: ,in, It is a kinetic energy term with controllable amplitude. and These are coefficients.

[0017] As a further limitation of the first aspect of the invention, during training, each fault mode Input the same number of vibration timing signals and temperature timing signals.

[0018] As a further limitation of the first aspect of the application, the failure modes include: motor bearing failure, stator failure, fan impeller imbalance, pump body cavitation and gear box wear.

[0019] In an implementation form of the first aspect of the application, the vibration time series signal and the temperature time series signal are pulse data.

[0020] The second aspect of the application provides an industrial equipment failure warning system based on a pulse neural network.

[0021] An industrial equipment failure warning system based on a pulse neural network comprises:

[0022] A time series data acquisition unit configured to acquire a vibration time series signal and a temperature time series signal of an industrial equipment;

[0023] A time series data processing unit configured to input the vibration time series signal and the temperature time series signal into a pulse neural network optimized by using a quantum annealing algorithm to obtain an output state of each neuron;

[0024] A global energy calculation unit configured to calculate a connection weight matrix of each neuron of the pulse neural network according to the output state of each neuron, and determine a global energy according to the output state of each neuron and the connection weight matrix;

[0025] A warning signal generation unit configured to determine that the industrial equipment has a failure and generate a failure warning signal when the global energy is lower than a first set threshold and a maintenance time is greater than a second set threshold.

[0026] The third aspect of the application provides a computer device comprising a processor and a computer readable storage medium.

[0027] The processor is adapted to execute a computer program.

[0028] The computer readable storage medium has a computer program stored therein, and the computer program is executed by the processor to implement the industrial equipment failure warning method based on the pulse neural network according to the first aspect of the application.

[0029] Compared with the prior art, the application has the following beneficial effects:

[0030] The application significantly improves the processing capability of the long-range dependence relationship in the time sequence signal of the industrial equipment by introducing the quantum annealing algorithm to optimize the pulse neural network. The traditional neural network is difficult to capture the correlation of the remote time point when processing long time sequence data due to the gradient disappearance or the limitation of the model complexity. The quantum annealing algorithm dynamically adjusts the kinetic term coefficient in the objective function, so that the network training process can "jump" from the local optimal solution, thereby more efficiently mining the deep time sequence pattern in the vibration and temperature data. This optimization mechanism not only retains the natural event-driven characteristics of the pulse neural network, but also enhances the modeling accuracy of the network to the long-term evolution law of the equipment state through the global search capability of quantum annealing, providing more reliable technical support for fault feature extraction in complex industrial scenarios.

[0031] The pulse neural network based on quantum annealing optimization has significant advantages in energy consumption and deployment adaptability. The pulse neural network itself takes sparse pulse coding as the core, and its calculation process is activated only when the neuron triggers the pulse, which naturally meets the requirements of low power consumption and real-time performance of industrial edge devices. The introduction of the quantum annealing algorithm does not increase the complexity of the network, but further reduces the computational overhead in the training phase by optimizing the convergence efficiency of the connection weight matrix. This lightweight design enables the optimized network to be directly deployed on resource-constrained industrial site devices without relying on cloud high-performance computing resources, reducing the system deployment cost and improving the real-time response capability of fault warning, providing an efficient solution for edge intelligence applications in industrial Internet of Things scenarios.

[0032] The synergistic optimization mechanism of global energy and kinetic term directly improves the accuracy and reliability of fault warning. By combining the global energy (reflecting the network's stable evaluation of the current equipment state) and the controllable amplitude kinetic term (driving the network to jump out of the local optimum) with a weight, the system can more accurately distinguish between normal fluctuations and fault precursors in a dynamic industrial environment. When the global energy continues to be lower than the set threshold, it indicates that the network has converged to the optimal representation of the real state of the equipment. At this time, the triggered fault warning signal can effectively avoid false positives caused by transient noise or short-term abnormalities. This energy evolution-based judgment logic, combined with the quantum annealing-optimized network parameters, significantly improves the sensitivity and specificity of fault warning, providing a more accurate risk warning means for the safe operation of industrial equipment.

[0033] The advantages of the additional aspects of the application will be partially given in the following description, partially will become apparent from the following description, or will be learned by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0034] The accompanying drawings, which form a part of this description, are included to provide further understanding of the application, and are incorporated in and constitute a part of this application.

[0035] Figure 1 A flowchart of a method for early warning of industrial equipment failure based on a spiking neural network is provided for an exemplary embodiment of the present application.

[0036] Figure 2 A schematic diagram of a neuron is provided for an exemplary embodiment of the present application.

[0037] Figure 3 A schematic diagram of a system for early warning of industrial equipment failure based on a spiking neural network is provided for an exemplary embodiment of the present application.

[0038] Figure 4 A schematic diagram of a computer device is provided for an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0039] The present application will be further described below with reference to the accompanying drawings and examples.

[0040] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0041] Early warning of industrial equipment failure is one of the core applications of industrial Internet of Things. According to industry statistics, the annual industrial loss caused by unplanned equipment downtime exceeds 100 billion yuan. Therefore, real-time modeling and anomaly recognition of time series signals such as equipment vibration, temperature, and current are crucial. As described in the background art, the traditional technical path has the following problems: Artificial neural network (ANN): such as the fault prediction model based on Transformer, although it can capture long-range dependencies, the single reasoning cycle power consumption is ≥50mW, and it needs to rely on cloud or high-performance edge servers, which cannot be deployed on on-site controllers (such as PLC, embedded modules) with limited computing power; although the optimized spiking neural network (SNN) has low energy consumption (≤20mW), it is insufficient in processing power for long-range dependencies, and the capture rate of pre-fault signals across hours is low, resulting in delayed or false early warning; the method of optimizing the SNN connection matrix through classical algorithms relies on thermal fluctuations and is easily trapped in local optimum. In the high-dimensional energy landscape of industrial signals (such as multi-dimensional signals of vibration + temperature + current), the completeness of fault feature capture is insufficient, and the energy consumption increases after increasing the number of iterations, which reverses the advantage of SNN.

[0042] In summary, there is a clear demand in current industrial settings for fault warning solutions that combine edge deployment, long-range dependency, and low energy consumption, which existing technologies cannot simultaneously meet. Therefore, this paper proposes a fault warning method for industrial equipment based on spiking neural networks. Employing quantum annealing technology, the spiking neural network achieves the dual objectives of "low energy consumption + high performance with long-range dependency" in industrial equipment fault warning scenarios. Specifically, as follows... Figure 1 As shown, the process includes the following:

[0043] S101: Acquire vibration timing signals and temperature timing signals of industrial equipment;

[0044] S102: Input the vibration timing signal and the temperature timing signal into the spiking neural network optimized by the quantum annealing algorithm to obtain the output state of each neuron;

[0045] S103: Calculate the connection weight matrix of each neuron in the spiking neural network based on the output state of each neuron, and determine the global energy based on the output state of each neuron and the connection weight matrix.

[0046] S104: When the global energy is lower than the first set threshold and the duration is greater than the second set threshold, the industrial equipment is determined to be faulty and a fault warning signal is generated.

[0047] In this implementation, the neuron model of the spiking neural network is as follows: Figure 2 As shown, each neuron receives input signals from other neurons. arrive Each neuron contains weights. arrive and bias ,in, arrive The importance of each input signal was controlled, and It controls how easily neurons are activated. Inside the neuron, the received signals are... arrive Multiply by the corresponding weight arrive Add them together and add a bias. The result is:

[0048] (1);

[0049] Then, the activation function is applied to... The process is performed to obtain the output state of the neuron.

[0050] More specifically, taking 256 input signals from industrial sensors as an example, the neuron receives 256 input signals from the industrial sensors (denoted as...). To Each neuron contains weights (denoted as To ) and bias (denoted as ), and the parameters are defined as follows:

[0051] Weights To are used to control the importance of each input signal, and the value range is [-1, 1]; among them, the absolute value of the weight corresponding to the 128 vibration signals (such as To ) is set to 0.6-1.0 (because the vibration signal has a high correlation with the fault), and the absolute value of the weight corresponding to the 128 temperature signals (such as To ) is set to 0.3-0.5 (the temperature signal is mainly used to assist in fault judgment);

[0052] Bias is used to control the ease of neuron activation, and the value range is [0.8, 1.2]; in the high-noise environment of the industrial field (such as workshop dust and electromagnetic interference), the bias can be adjusted to 1.0-1.2 to reduce the invalid pulses triggered by noise;

[0053] The input signals To are all pulse signals, and the value is 0 or 1, where "1" indicates that the sensor detects an abnormal event (such as a vibration peak exceeding 1.2 times the normal threshold of the device or a temperature exceeding the normal working range of the device), and "0" indicates no abnormal event;

[0054] The number of input channels is fixed at 256, including 128 vibration event signals and 128 temperature event signals, which meets the needs of multi-dimensional collection of industrial sensors.

[0055] The internal calculation process of the neuron is divided into two steps:

[0056] First step: input weighted accumulation, multiply To by the corresponding weight To , add all the product results, and then add the bias , to get the intermediate result , as shown in formula (1).

[0057] Second step: activation function processing, using ReLU-like pulse activation function adapted to industrial scenarios, when ​When the threshold is set to adapt to industrial noise environment, the neuron outputs 1 (generating a pulse, representing the detection of potential fault characteristics); otherwise, it outputs 0, and finally obtains the neuron's output signal.

[0058] In this implementation, when multiple neurons form a network layer, and multiple network layers further form a complex SNN, the connection strength between neurons is determined by the weight matrix (denoted as...). ) definition, matrix elements in Representing the The first neuron pairs with the first The influence strength of each neuron ( , Each corresponds to a processing neuron for 256 input signals. To ensure the stability and fault feature capture capability of the SNN in industrial scenarios, the matrix... The following two core constraints must be followed:

[0059] Symmetric connection ( This means that the effect of a neuron on a pair is the same as the effect on a pair. This condition is not merely for convenience; it is a mathematical prerequisite for ensuring that the network can define a global energy function (Lyapunov function) and that the network dynamics always evolve in the direction of decreasing energy.

[0060] No self-connection ( Neurons do not connect to themselves. This rule is also crucial because non-zero self-connections ( This can lead to individual neurons getting stuck in a "self-locked" stable state, thus preventing the network as a whole from exploring and reaching a meaningful collective stable state.

[0061] The value of is the sum of the correlations between and state across all failure modes. A large positive value means that in most memories, and tend to be in the same state, thus forming an excitatory connection between them. Conversely, a large negative value indicates that they tend to be in different states, forming an inhibitory connection. Through this simple local rule, the global structural information of all failure modes is cleverly encoded into the network's connection weights.

[0062] Suppose we want to store One fault mode , Defined as:

[0063] (2);

[0064] In equation (2) Will State mapping If in a specific failure mode In the equation, if neurons A and B have the same state (both 1 or both 0), then the value of A is 1; if their states are different, then the value is -1.

[0065] The value is the sum of the state correlations across all failure modes. If A large positive number (e.g., ≥0.6) indicates that the neuron... and In most fault modes, the states are the same (e.g., in a certain type of fault, vibration abnormalities and temperature abnormalities occur simultaneously), and the two form an excitatory connection; if A large negative number (such as less than or equal to -0.6) indicates that the neuron... and In most fault modes, the states are opposite (e.g., in a certain type of fault, only vibration is abnormal, while temperature remains unchanged), forming an inhibitory connection. Using this calculation rule, the global characteristics of the five types of faults (including the temporal correlation of fault signals and multi-dimensional collaborative characteristics) can be fully encoded into the connection matrix. .

[0066] For a symmetric connection matrix For a given network, a global energy function can be defined as the system's objective optimization function:

[0067] (3);

[0068] in, This refers to the global energy.

[0069] Neural network training requires finding the global minimum of the function. The parameters are defined in the industrial scenario as follows:

[0070] Total number of neurons 512: The SNN adopts a 2-layer hidden layer design, with 256 neurons in each layer, for a total of 512 neurons, which can adapt to the feature expansion requirements of 256 input signals.

[0071] Representing the The state of each neuron is converted from the original pulse signal to -1 or 1 through mapping;

[0072] Global energy Value range and fault diagnosis: The value range is [-100, 50]. In industrial equipment fault early warning scenarios, when... At that time, it is determined that the SNN has stably captured the fault characteristics;

[0073] The core objective of SNN training is to find global capabilities. global minimum, ensuring stable capture of long-range dependent failure signals of industrial equipment.

[0074] In the present implementation, the quantum annealing technique is adopted to obtain the globally optimal solution of long-range correlation, the pulse neural network is trained on a quantum machine, and quantum computing is used to solve complex problems that are difficult for classical computers to handle. Quantum annealing is an optimization method based on quantum tunneling effect. Its core mechanism is to evolve the quantum state from the initial high-energy state to the low-energy state solution of the target problem by adjusting the Hamiltonian of the system. Unlike classical annealing which relies on thermal fluctuations, quantum annealing uses quantum tunneling to penetrate energy barriers, thereby more efficiently avoiding local optimal solutions. Quantum tunneling is a non-local effect that has the probability of skipping local optimal solutions and finding the globally optimal solution, i.e., the global energy minimum. A major advantage of quantum annealing is that it can guarantee finding the global minimum of the target function under ideal conditions. Traditional methods such as Monte Carlo, simulated annealing, genetic algorithms, and machine learning-based methods cannot guarantee this and are prone to falling into local minima in complex or high-dimensional energy landscapes.

[0075] The entire quantum annealing process can be described by the equation

[0076] (4);

[0077] where is the time-dependent wave function of the pulse neural network system, which is used to describe the quantumized expression of industrial failure features; i is the imaginary unit ( ); is the reduced Planck constant (the actual value is approximately equal to , and a normalized value of 1.0 can be used in industrial scene calculations to simplify operations); is the time, with ms as the unit, and is fixed at 1000 ms in industrial scenarios (adapted to the time scale of equipment failure signals); is the total Hamiltonian of the system at time t, which controls the evolution process of the quantum state.

[0078] The quantum annealing algorithm model consists of two parts. One part is the global energy , which is the target function to be optimized, and the second part is the kinetic energy, which is introduced by the amplitude-controllable kinetic energy term , i.e., the system temperature is regulated. As a perturbation of the system, its role is to make the target function jump from the local minimum to the global minimum. The total Hamiltonian (i.e., the total energy) is:

[0079] (5);

[0080] where ​The potential energy coefficient is dynamically adjusted over time. Its initial value is 1.0, and it decreases to 0.1 at a rate of 0.001 / ms (the adjustment is completed within 1000ms). Its function is to reduce the system's dependence on initial noise. The coefficient representing the kinetic energy term is dynamically adjusted over time. Its initial value is 0.1, and it increases to 1.0 at a rate of 0.001 / ms (the adjustment is completed within 1000ms). Its function is to drive the system to break through the local energy barrier through quantum kinetic energy.

[0081] Amplitude controllable kinetic energy term The calculation method is as follows:

[0082] (6);

[0083] in, For the first The quantum momentum of each neuron follows a normal distribution N(0,1) and is used to simulate the random fluctuation characteristics of industrial signals, constantly changing... This causes the system to move away from its original local minimum, changing its potential energy. The corresponding connection matrix... It also dynamically adjusts, cooling the neural network system composed of neurons to the ground state, that is, the lowest energy state. Since it is a quantum system, the system has the probability of jumping to a distant place with a lower potential energy, so it can capture relatively stable long-distance correlations.

[0084] The greatest advantage of spiking neural networks is that they can make full use of information based on spatiotemporal events. The spiking neural network using quantum annealing technology in this invention can not only consume less energy than artificial methods, but also achieve a great leap in performance in capturing long-range semantic dependencies.

[0085] In this implementation, more specifically, the co-training process of quantum annealing and SNN consists of four steps, including:

[0086] Step 1: Collect historical data of industrial equipment operating under five types of fault conditions (collect 1000 sets of time-series signals for each type of fault, with each set containing 256 input features), initialize the connection matrix T of the SNN using the Hebbian rule; at the same time, set the initial parameters of quantum annealing, i.e., A(t)=1.0, B(t)=0.1, and start the quantum annealing process;

[0087] The second step is to simultaneously reduce the potential energy coefficient A(t) and increase the kinetic energy coefficient B(t) within 1000ms, and use the quantum tunneling effect to explore the high-dimensional energy landscape of industrial signals, breaking through the local energy minimum caused by noise (such as false vibration signals caused by electromagnetic interference in the workshop).

[0088] Step 3: Global Energy The connection matrix T is dynamically adjusted in real time with quantum evolution. When the long-range characteristics of a certain type of fault (such as vibration period changes caused by bearing wear) are detected, the connection weights of the corresponding neurons in T are automatically strengthened (for example, from 0.4 to 0.8;

[0089] Step 4: When the peak value of the vibration signal drops to 0.1, rises to 1.0, stop the quantum evolution, and cool the system to the ground state (i.e., the lowest energy state), at which point the connection matrix is the optimal matrix, and the SNN can stably capture the long-range dependence characteristics of equipment faults;

[0090] Step 5: solidify the optimal matrix to an industrial edge controller (such as an ARM Cortex-M4), complete the on-site deployment of the SNN model, and no longer need to perform quantum calculations on site, only need to load the optimal matrix to perform inference, greatly reducing the on-site computing power requirements.

[0091] More specifically, the present implementation proposes an industrial equipment fault warning system, which has a four-layer hardware architecture, and the core components and functions of each layer are as follows:

[0092] Sensor layer: the core components are a neural state vibration sensor (model Inivation DVS240) and a digital temperature sensor (model DS18B20). The vibration sensor uses a 1 kHz sampling frequency to detect the peak value of the equipment vibration in real time, and outputs a pulse (X=1) when the peak value exceeds 1.2 times the rated vibration of the equipment; the temperature sensor has a 100 ms sampling interval, and outputs a pulse (X=1) when the detected temperature exceeds the normal working temperature of the equipment (such as the upper limit of the normal working temperature of the motor 80℃);

[0093] Preprocessing layer: the core component is the filter module built into the industrial edge controller (model ARM Cortex-M4). This module uses a sliding window filtering algorithm with a window size of 5 frames (each frame has a duration of 10 ms), which can effectively eliminate isolated noise pulses; at the same time, the vibration signal and the temperature signal are fused into 256 input features, which are transmitted to the SNN calculation layer;

[0094] SNN calculation layer: the core component is the SNN inference module built into the edge controller. This module loads the connection matrix optimized by quantum, performs neuron accumulation calculation ( ) and activation function operation, and finally outputs the global energy ;

[0095] Warning layer: the core components are the on-site sound and light alarm and the remote 5G module. When the system energy ​And the state lasts for 10 frames (total duration 100 ms), the sound and light alarm triggers the scene alarm; at the same time, the 4G module uploads the fault characteristics (including global energy and pulse sequence) to the industrial cloud platform, realizing remote early warning and data retention.

[0096] Based on the above hierarchical division, the specific workflow is as follows:

[0097] The sensor real-time collects the equipment vibration and temperature signals, and the preprocessing layer generates 1 frame of 256-way pulse characteristics every 10 ms. After filtering and removing noise through a sliding window, the feature data is input into the SNN calculation layer.

[0098] The SNN calculation layer loads the optimal connection matrix T, calculates the system energy corresponding to each frame of characteristics , and if , the frame is marked as a "potential failure frame";

[0099] If 10 frames of "potential failure frames" are continuously detected (total duration 100 ms), the system determines that there is a failure precursor in the equipment, and immediately triggers the on-site alarm and remote data upload of the early warning layer;

[0100] The industrial cloud platform collects new failure data generated on site every month (if any), and uses the cloud quantum computing device to re-optimize the connection matrix through quantum annealing, and then pushes the updated matrix to the edge controller through the 5G module, realizing model iteration and ensuring long-term early warning accuracy.

[0101] Optionally, in some other implementations, to solve the problem of traditional failure early warning "only identifying failure, unable to determine mode", combined with pre-trained failure benchmark characteristics and real-time signals, the failure mode is determined through "real-time energy matching + neuron state distribution verification" dual-dimension, the core is to use the optimal connection matrix optimized by quantum and the output state of each neuron, to build the accurate matching logic of "real-time characteristics-benchmark characteristics".

[0102] When the following criteria are met, it is determined to be the kth failure mode:

[0103] Criterion 1:

[0104] (7);

[0105] Criterion 2:

[0106] (8);

[0107] Wherein,

[0108] (9);

[0109] This represents the global energy of the SNN during real-time operation of the device; The pre-training baseline energy represents the k-th type of fault mode; This represents the energy matching threshold, with a value of 5. This represents the total number of SNN neurons, which is fixed at 512. Representing the The real-time state of each neuron (taking a value of -1 or 1) is calculated from the real-time sensor signal. The value is obtained by quantization. Representing the Class of failure modes The typical state of a neuron; This represents the state distribution matching threshold, with a value of 0.2. This represents the optimal connection matrix after quantum optimization.

[0110] Optionally, in other implementations, to address the issue that traditional SNN neuron state quantization (simply mapping 0 / 1 to -1 / 1) cannot accurately capture subtle precursors to industrial equipment faults, the fault sensitivity thresholds of industrial sensors (such as vibration exceeding 1.2 times the rated value or temperature exceeding 80°C) are incorporated into the neurons. Value calculation and state mapping, combined with the global energy constraint of quantum annealing, ensure that the neuron state accurately reflects the "early signs of device malfunction," rather than simply the presence or absence of a signal. Specifically, the neuron's output state is set as follows:

[0111] (10);

[0112] in,

[0113] (11);

[0114] Representing the The final state (-1 / 1) of each neuron is used in the subsequent connection matrix. Computation and fault identification; The weighted summation result of the i-th neuron reflects the degree of anomaly in the industrial signal; This represents the weight of the neuron on the vibration signal; This represents the weight of the neuron on the temperature signal; Representing the Road vibration input signal; Representing the Road temperature input signal; Represents neuron bias; Represents the industrial fault-sensitive activation threshold; Represents the global energy of the SNN; representing a fault precursor energy threshold value;

[0115] Optionally, in other implementations, a fixed evolution rate (potential energy coefficient , kinetic energy coefficient constant rate) is adopted for traditional quantum annealing, which cannot adapt to the "long-term, gradual" characteristics of industrial equipment fault precursors (such as bearing wear precursors lasting for several hours). The evolution rate of quantum annealing is associated with the actual duration of industrial fault precursors, and the change rate of , is dynamically adjusted through a timing adaptive factor, ensuring that quantum tunneling can fully capture long-range dependencies of faults and avoid missing long-term characteristics due to fixed rates. Specifically, a timing-adaptive quantum annealing coefficient evolution formula is given, including:

[0116] (12);

[0117] (13);

[0118] wherein the timing adaptive factor is:

[0119] (14);

[0120] represents the initial value of , represents the target value of , represents the initial value of , t represents the current time of quantum annealing, represents the total duration of quantum annealing, represents the actual duration of the industrial equipment fault precursor, represents the standard fault precursor duration.

[0121] Figure 3 shows an industrial equipment fault warning system based on a pulse neural network, comprising:

[0122] a timing data acquisition unit configured to acquire vibration timing signals and temperature timing signals of the industrial equipment;

[0123] a timing data processing unit configured to input the vibration timing signals and the temperature timing signals into a pulse neural network optimized by a quantum annealing algorithm to obtain output states of each neuron;

[0124] The global energy calculation unit is configured to calculate the connection weight matrix of each neuron of the spiking neural network according to the output state of each neuron, and determine the global energy according to the output state of each neuron and the connection weight matrix.

[0125] The early warning signal generation unit is configured to determine that the industrial equipment has a fault and generate a fault early warning signal when the global energy is lower than the first set threshold and the maintenance time is greater than the second set threshold.

[0126] It can be understood that the above-mentioned various units can be combined into one or several other units respectively or entirely, or some of the units can be further split into a plurality of units with smaller functions to constitute, which can achieve the same operation without affecting the implementation of the technical effects of the embodiments of the present application. The above-mentioned units are divided based on logical functions, and the functions of one unit can also be implemented by multiple units, or the functions of multiple units can be implemented by one unit. In other embodiments of the present application, the system can also include other units, and these functions can also be assisted by other units in actual application, and can be implemented by multiple units.

[0127] According to another embodiment of the present application, the system described in the embodiment can be constructed by running a computer program (including program code) capable of performing each step involved in the corresponding method of the present application on a general computing device such as a computer including processing elements and storage elements such as a central processing unit (CPU), a random access memory (RAM), a read-only memory (ROM), etc., the computer program can be recorded on a computer readable recording medium, and loaded into the above-mentioned computing device through the computer readable recording medium and run therein.

[0128] Figure 4 A computer device is shown, which includes a processor 401, a communication interface 402, and a computer readable storage medium 403. Wherein the processor 401, the communication interface 402 and the computer readable storage medium 403 can be connected through a bus or other means.

[0129] Wherein, the communication interface 402 is used for receiving and sending data, the computer readable storage medium 403 can be stored in the memory of the electronic device, the computer readable storage medium 403 is used for storing computer programs, the computer programs include program instructions, and the processor 401 is used for executing the program instructions stored in the computer readable storage medium 403.

[0130] The processor 401 is a computing core and a control core of the electronic device, and is adapted to implement one or more instructions, and is specifically adapted to load and execute one or more instructions to implement a corresponding method flow or a corresponding function.

[0131] The processor 401 is configured to perform the following process:

[0132] Obtain a vibration time sequence signal and a temperature time sequence signal of the industrial equipment;

[0133] Input the vibration time sequence signal and the temperature time sequence signal into a pulse neural network optimized by using a quantum annealing algorithm to obtain an output state of each neuron;

[0134] Calculate a connection weight matrix of each neuron of the pulse neural network according to the output state of each neuron, and determine a global energy according to the output state of each neuron and the connection weight matrix;

[0135] When the global energy is lower than a first set threshold and a maintenance time is greater than a second set threshold, it is determined that the industrial equipment has a fault and a fault warning signal is generated.

[0136] Those of ordinary skill in the art can be aware that units and algorithm steps of each example described in combination with the embodiments disclosed in the application can be implemented, for example, by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed by hardware or software depends on the specific application and design constraints of the technical solution. Those of ordinary skill in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the application.

[0137] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in or transmitted by a computer readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired (for example, coaxial cable, optical fiber, digital line) or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer readable storage medium can be any available medium that can be accessed by a computer or a data processing device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (for example, floppy disk, hard disk, magnetic tape), optical media (for example, DVD), or semiconductor media (for example, solid state disk) and the like.

[0138] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for early warning of industrial equipment failure based on a spiking neural network, characterized in that, The method comprises the following steps: obtaining a vibration time series signal and a temperature time series signal of an industrial equipment; inputting the vibration time series signal and the temperature time series signal into a pulse neural network optimized by a quantum annealing algorithm to obtain an output state of each neuron; The quantum annealing algorithm is used for training the pulse neural network. The total energy is set as the weighted sum of the global energy and the amplitude-controllable kinetic energy term. By adjusting the coefficient of the kinetic energy term, the target function is continuously transferred from the local minimum to the global minimum, and the latest training result is obtained , is a connection weight matrix calculating a connection weight matrix of each neuron of the pulse neural network according to the output state of each neuron, and determining a global energy according to the output state of each neuron and the connection weight matrix; when the global energy is lower than a first set threshold and a maintenance time is greater than a second set threshold, determining that the industrial equipment has a fault and generating a fault warning signal.

2. The industrial equipment fault warning method based on a pulse neural network according to claim 1, wherein During the training of a spiking neural network, the connection weight matrix... Data elements for: ,in, Representative failure mode Next The output state of each neuron Representative failure mode Next The output state of each neuron.

3. The industrial equipment fault warning method based on a pulse neural network according to claim 1, wherein A global energy is determined from the output states of the individual neurons and the connection weight matrix, including: wherein represents the output state of the th neuron, the output state of the th neuron.

4. The industrial equipment fault warning method based on a pulse neural network according to claim 3, wherein The total energy is a weighted sum of the global energy and the amplitude- controllable kinetic energy term, including: wherein, is the amplitude-controllable kinetic energy term, and are coefficients, respectively.

5. The industrial equipment fault warning method based on a pulse neural network according to claim 4, wherein Each fault mode at training time The same number of vibration and temperature time series signals are input.

6. The industrial equipment fault warning method based on a pulse neural network according to claim 1, wherein the fault modes include motor bearing failure, stator failure, fan impeller imbalance, pump body cavitation, and gear box wear.

7. The industrial equipment fault warning method based on a pulse neural network according to any one of claims 1-6, wherein the vibration time series signal and the temperature time series signal are both pulse data.

8. A pulse neural network-based industrial equipment failure early warning system, characterized by, The method comprises: a time series data acquisition unit configured to obtain a vibration time series signal and a temperature time series signal of an industrial equipment; a time series data processing unit configured to input the vibration time series signal and the temperature time series signal into a pulse neural network optimized by a quantum annealing algorithm to obtain an output state of each neuron; The quantum annealing algorithm is used for training the pulse neural network. The total energy is set as the weighted sum of the global energy and the amplitude-controllable kinetic energy term. By adjusting the coefficient of the kinetic energy term, the target function is continuously transferred from the local minimum to the global minimum, and the latest training result is obtained , is a connection weight matrix a global energy calculation unit configured to calculate a connection weight matrix of each neuron of the pulse neural network according to the output state of each neuron, and determine a global energy according to the output state of each neuron and the connection weight matrix; a warning signal generation unit configured to, when the global energy is lower than a first set threshold and a maintenance time is greater than a second set threshold, determine that the industrial equipment has a fault and generate a fault warning signal.

9. A computer device, comprising: The method comprises: a processor and a computer readable storage medium; the processor is adapted to execute a computer program; the computer readable storage medium has a computer program stored therein, and the computer program is executed by the processor to implement the industrial equipment fault warning method based on a pulse neural network according to any one of claims 1-7.

Images