Remaining service life prediction method, device, equipment, medium and product

Abstract

The invention relates to the technical field of reliability engineering, and particularly provides a residual service life prediction method and device, equipment, a medium and a product, and the method comprises the steps: obtaining sensor data of industrial equipment; wherein the sensor data is used for indicating the operation state of the industrial equipment; performing pulse coding on the sensor data to obtain a pulse sequence after coding; wherein the pulse sequence is used for describing the dynamic change of the operation state at different time points; inputting the pulse sequence into a spiking neuron to determine a state parameter of the spiking neuron; and predicting the remaining service life of the industrial equipment based on the state parameters. Through the processing mode, the time sequence dependency relationship of the operation state can be effectively learned, and the fine-grained time sequence characteristics in the pulse sequence can be mined, so that the prediction precision and prediction accuracy of the residual service life are improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of reliability engineering, and in particular, to a remaining useful life prediction method, device, equipment, medium and product. BACKGROUND

[0002] Industrial equipment will produce performance degradation due to wear, fatigue or environmental factors during long-term operation. If the remaining useful life RUL (Remaining Useful Life) of the industrial equipment cannot be predicted in time, the industrial equipment will be shut down, and even safety accidents will be caused.

[0003] In related technologies, a traditional machine learning model and a manual feature extraction method are usually used for RUL prediction. However, this processing method cannot effectively learn the time sequence dependent features of the industrial equipment, and thus the prediction accuracy of the RUL is insufficient. SUMMARY

[0004] The present disclosure is proposed in view of the above problems. The present disclosure provides a remaining useful life prediction method, device, equipment, medium and product.

[0005] According to one aspect of the present disclosure, a remaining useful life prediction method is provided, comprising:

[0006] Obtaining sensor data of an industrial equipment; wherein the sensor data is used to indicate an operating state of the industrial equipment;

[0007] Pulse coding is performed on the sensor data, and a pulse sequence is obtained after coding; wherein the pulse sequence is used to describe dynamic changes of the operating state at different time points;

[0008] The pulse sequence is input into a pulse neuron to determine a state parameter of the pulse neuron;

[0009] Based on the state parameter, the remaining useful life of the industrial equipment is predicted.

[0010] In addition, according to the pulse sequence input into the pulse neuron to determine the state parameter of the pulse neuron according to one aspect of the present disclosure, further comprising:

[0011] Mapping the pulse sequence to a target vector space; wherein the target vector space is used to describe a membrane potential of the pulse neuron at different time points;

[0012] Determining a total number of pulse releases of the pulse neuron based on a potential change of the membrane potential;

[0013] Taking the membrane potential and the total number of pulse releases as the state parameter.

[0014] Further, the mapping the pulse sequence to a target vector space according to an aspect of the present disclosure further comprises:

[0015] acquiring a pulse release state of a plurality of pulse neurons at a first time point;

[0016] performing a weighted sum operation on the pulse release state and a target pulse sequence to obtain a target weighted result; wherein the target pulse sequence is a pulse sequence at the first time point;

[0017] determining a membrane potential decay rate of the pulse neuron at the first time point;

[0018] performing a sum operation on the target weighted result and the membrane potential decay rate to obtain a membrane potential of the pulse neuron at a second time point; wherein the second time point is after the first time point.

[0019] Further, the performing a weighted sum operation on the pulse release state and a target pulse sequence to obtain a target weighted result according to an aspect of the present disclosure further comprises:

[0020] performing a weighted operation on a first weight and the pulse release state to obtain a first weighted result; wherein the first weight is used to indicate a connection strength between the pulse neuron and other pulse neurons;

[0021] performing a weighted operation on a second weight and the target pulse sequence to obtain a second weighted result; wherein the second weight is used to indicate an influence degree of the target pulse sequence on the pulse neuron;

[0022] performing a sum operation on the first weighted result and the second weighted result to obtain the target weighted result.

[0023] Further, the determining a membrane potential decay rate of the pulse neuron at the first time point according to an aspect of the present disclosure further comprises:

[0024] determining a time difference between the first time point and the second time point;

[0025] performing an exponential operation on the time difference and a membrane time constant to obtain an operation result;

[0026] determining a product of the operation result and a membrane potential of the pulse neuron at the first time point as the membrane potential decay rate.

[0027] Further, the pulse encoding the sensor data according to an aspect of the present disclosure, and obtaining a pulse sequence after encoding further comprises:

[0028] extracting an operating state indicator in the sensor data according to at least one extraction dimension;

[0029] mapping a pre-processing result of the operating state indicator to a pulse sending time;

[0030] generating the pulse sequence based on the pulse sending time.

[0031] In addition, the mapping of the pre-processing result of the operating state indicator to the pulse sending time according to one aspect of the present disclosure further comprises:

[0032] determining the pulse sending time according to the pre-processing result and a mapping time, wherein the mapping time is a time from the pre-processing result to the pulse sending time.

[0033] In addition, the predicting of the remaining useful life of the industrial equipment based on the state parameter according to one aspect of the present disclosure further comprises:

[0034] performing weighted summation processing on the membrane potential of the pulse neuron and the total number of pulse releases to obtain a third weighted result;

[0035] predicting the remaining useful life based on the third weighted result.

[0036] In addition, according to one aspect of the present disclosure, the remaining useful life prediction method further comprises:

[0037] deriving the pulse release state function to obtain a derivation result;

[0038] optimizing the first weight and the second weight based on a product of the derivation result and a learning rate to obtain an optimized first weight and an optimized second weight, wherein the learning rate is used to indicate the adjustment amplitude of the first weight and the second weight;

[0039] taking the optimized first weight as the weight of the pulse release state, and taking the optimized second weight as the weight of the target pulse sequence.

[0040] According to another aspect of the present disclosure, a remaining useful life prediction device is provided, comprising:

[0041] an acquisition unit configured to acquire sensor data of an industrial equipment, wherein the sensor data is used to indicate an operating state of the industrial equipment;

[0042] a pulse encoding unit configured to pulse encode the sensor data, and obtain a pulse sequence after encoding, wherein the pulse sequence is used to describe dynamic changes of the operating state at different time points;

[0043] a parameter determination unit configured to input the pulse sequence into a pulse neuron to determine a state parameter of the pulse neuron;

[0044] a prediction unit configured to predict a remaining useful life of the industrial equipment based on the state parameter.

[0045] According to yet another aspect of the present disclosure, an electronic device is provided, comprising a processor, a memory, and a bus, the memory storing machine readable instructions executable by the processor, the processor and the memory communicating via the bus when the electronic device is running, the machine readable instructions being executed by the processor to perform the steps of the above-mentioned remaining useful life prediction method.

[0046] According to yet another aspect of the present disclosure, a computer readable storage medium is provided, the computer readable storage medium storing a computer program, the computer program being executed by a processor to perform the steps of the above-mentioned remaining useful life prediction method.

[0047] According to yet another aspect of the present disclosure, a computer program product is provided, the computer program product being stored in a storage medium, the program product being executed by at least one processor to implement the steps of the above-mentioned remaining useful life prediction method.

[0048] As will be described in detail below, a remaining useful life prediction method, apparatus, device, medium, and product according to embodiments of the present disclosure. In embodiments of the present disclosure, sensor data can be pulse coded, and a pulse sequence is obtained after coding, wherein the pulse sequence is used to describe dynamic changes of an operating state of an industrial equipment at different time points, then the pulse sequence is input into a pulse neuron to determine a state parameter of the pulse neuron, so that the remaining useful life of the industrial equipment can be predicted based on the state parameter. By pulse coding the sensor data, the time sequence dependency of the operating state can be effectively learned, and the remaining useful life of the industrial equipment can be predicted according to the state parameter of the pulse neuron, to mine fine-grained time sequence features in the pulse sequence, and thus improve the prediction accuracy and prediction accuracy of the remaining useful life.

[0049] It is to be understood that both the foregoing general description and the following detailed description are exemplary and intended to provide further explanation of the subject technology. BRIEF DESCRIPTION OF DRAWINGS

[0050] The above and other objects, features and advantages of the present disclosure will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which:

[0051] Figure 1 is a flowchart illustrating a remaining useful life prediction method according to an embodiment of the present disclosure.

[0052] Figure 2 is an architecture diagram further illustrating a recurrent spiking neural network RSNN in the remaining useful life prediction method of the embodiment of the present disclosure.

[0053] Figure 3 is a flowchart further illustrating a remaining useful life prediction process in the remaining useful life prediction method of the embodiment of the present disclosure.

[0054] Figure 4 is a block diagram illustrating a remaining useful life prediction apparatus according to an embodiment of the present disclosure.

[0055] Figure 5 is a hardware block diagram illustrating an electronic device according to an embodiment of the present disclosure.

[0056] Figure 6 is a schematic diagram illustrating a computer program product according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present disclosure more apparent, the following will describe example embodiments according to the present disclosure in detail with reference to the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present disclosure, rather than all the embodiments of the present disclosure, and it should be understood that the present disclosure is not limited to the example embodiments described herein.

[0058] To facilitate understanding of the present embodiment, first, a remaining useful life prediction method disclosed by the embodiment of the present disclosure is described in detail. The execution subject of the remaining useful life prediction method provided by the embodiment of the present disclosure is generally an electronic device with certain computing capability, which includes, for example, a terminal device or a server or other processing device. In some possible implementation manners, the remaining useful life prediction method can be realized by a processor calling computer readable instructions stored in a memory.

[0059] Referring to Figure 1As shown, a flowchart of a remaining useful life prediction method provided by an embodiment of the present disclosure is shown, and the method comprises steps S101-S104, wherein:

[0060] Step S101: Obtain sensor data of an industrial equipment; wherein the sensor data is used to indicate an operating state of the industrial equipment.

[0061] Here, the industrial equipment includes an aero-engine, a gas turbine, and a key unit in a manufacturing production line, and the present disclosure does not limit this.

[0062] Among them, the industrial equipment is installed with multiple sensors, such as pressure sensors, temperature sensors, and speed sensors.

[0063] Correspondingly, the sensor data is data collected from the sensors, such as pressure values output by the pressure sensors and temperature values output by the temperature sensors. Moreover, the sensor data in the present disclosure includes different working condition combinations and failure modes to cover the entire life cycle of the industrial equipment.

[0064] Embodiments of the present disclosure can preprocess the sensor data, such as normalization processing, de-redundancy, and sliding window segmentation. Among them, each sensor can correspond to one or more data transmission channels, and correspondingly, the sensor data can also correspond to one or more data transmission channels.

[0065] The present disclosure can adopt the following preprocessing flow:

[0066] Statistically calculate the variance of each data transmission channel or each sensor data in the entire life cycle, remove the channels or sensor data that are always constant; for the sensor data corresponding to one or more sensors in the same application scenario, extreme value normalization is adopted, which can weaken the distribution drift problem caused by switching different application scenarios; to avoid the problem of too large RUL prediction in the early stage, causing gradient dilution, and the problem of unclear degradation time of the industrial equipment itself, a maximum RUL threshold can be set in the early stage of the life cycle of the industrial equipment to limit the prediction value of RUL, for example, the RUL threshold can be set to 125; a sliding window with a window length T=30 and a time step s=1 is adopted to segment each sensor data, and the last time window data is selected for prediction, for example, if there are 120 time steps of data, the first 90 time steps of data can be used to predict the RUL of the 91st time step.

[0067] In addition, the execution subject of the remaining useful life prediction method in the present disclosure can be an industrial equipment or a third-party equipment, and the execution subject is integrated with a neural network model, such as a recurrent spiking neural network RSNN (Recurrent Spiking Neural Network).

[0068] In the embodiments of the present disclosure, the industrial equipment can take the sensor data collected by the sensors as the data basis, and pre-process the sensor data to ensure the continuity and real-time of the degradation trend of the sensor data.

[0069] Step S102: pulse coding is performed on the sensor data, and a pulse sequence is obtained after coding; wherein the pulse sequence is used to describe the dynamic change of the running state at different time points.

[0070] Here, each pulse in the pulse sequence represents the running state of the industrial equipment at a specified time point or a specified time period, so as to describe the dynamic change of the running state in the form of pulses.

[0071] Since the sensor data is derived from multiple sensors, the sensor data corresponding to each sensor can be pulse coded to obtain multiple pulse sequences, or the sensor data corresponding to each sensor can be integrated, and the integrated result can be pulse coded to obtain a pulse sequence. The present disclosure does not make any requirements in this regard.

[0072] The present disclosure does not limit the way of pulse coding, for example, time sequence coding, frequency coding or delay coding, and an exemplary scheme will be given later.

[0073] In the embodiments of the present disclosure, the industrial equipment can convert the continuous sensor data into discrete pulse events by the above-mentioned manner, so as to maintain the sparsity of the pulse events, thereby avoiding redundant calculation.

[0074] Step S103: inputting the pulse sequence into the pulse neuron to determine the state parameter of the pulse neuron.

[0075] Here, the state parameter is used to describe the activity state of the pulse neuron, for example, membrane potential and total pulse release times, which will be further described later.

[0076] As described above, the neural network model adopted by the present disclosure is RSNN, and the pulse neurons in the RSNN are arranged in layers. The original RSNN architecture includes an input coding layer, a hidden layer and an output decoding layer, wherein there can be one or more hidden layers between the input coding layer and the output decoding layer.

[0077] For the hidden layer, the disclosure adopts a sparse recurrent connection leaky integrate and fire (LIF) neuron to form a liquid layer, which can be understood as a liquid state machine (LSM) architecture. The LIF neuron model is a biological neuron-based model used to simulate the dynamic behavior of neurons, which can store and amplify the timing differences of the pulse sequence through the dynamics of the LSM, and realize dynamic characterization of the running state of the industrial equipment.

[0078] In the embodiments of the disclosure, the industrial equipment can convert the pulse sequence into the state parameters of the pulse neuron, realize natural coding of the timing information, and thus better capture and process the dynamic features in the pulse sequence.

[0079] Step S104: predicting the remaining useful life of the industrial equipment based on the state parameters.

[0080] In the embodiments of the disclosure, the industrial equipment can predict the length of time from the current time point to the failure or the need for major maintenance of the industrial equipment, i.e., the remaining useful life, by means of the state parameters of the pulse neuron, so as to reduce the downtime of the industrial equipment and reduce the maintenance cost of the industrial equipment.

[0081] In the above embodiments, the pulse sequence can be obtained by pulse coding of the sensor data, the timing dependence of the running state can be characterized by the pulse sequence, and the pulse sequence can be input to the pulse neuron, so as to predict the remaining useful life of the industrial equipment according to the state parameters of the pulse neuron, to mine the fine-grained timing features in the pulse sequence, and thus improve the prediction accuracy and prediction accuracy of the remaining useful life.

[0082] In an optional embodiment, the above step of inputting the pulse sequence to the pulse neuron to determine the state parameters of the pulse neuron specifically includes the following steps:

[0083] mapping the pulse sequence to a target vector space; wherein the target vector space is used to describe the membrane potential of the pulse neuron at different time points;

[0084] determining the total number of pulse releases of the pulse neuron based on the potential change of the membrane potential;

[0085] taking the membrane potential and the total number of pulse releases as the state parameters.

[0086] As mentioned above, the present disclosure introduces LSM into the RSNN architecture. Based on the dynamic idea of ​​LSM, the present disclosure first projects the pulse sequence to the input synapses of the liquid layer. This layer is composed of heterogeneous LIF neurons of the order of magnitude M (usually much larger than the input dimension N). This can be understood as the pulse neurons of the liquid layer, and a dynamic system is formed through random and sparse recursive connections.

[0087] Among them, the input dimension can be understood as the number of pulse neurons in the input coding layer. Each pulse neuron in the input coding layer can correspond to one or more pulse sequences. Correspondingly, each pulse neuron in the input coding layer can also correspond to one or more pulse neurons in the liquid layer. The correspondence between the pulse sequences, the pulse neurons in the input coding layer and the pulse neurons in the liquid layer can be adjusted as needed.

[0088] When the pulse train enters the spiking neurons in the liquid layer, the membrane potential of the spiking neurons in each liquid layer produces transient fluctuations under the combined action of leakage, pulse train and loop feedback, thus forming a liquid morphology that evolves over time.

[0089] Formally, if the pulse sequence is u(t), the cyclically connected liquid layer will form a nonlinear mapping from low dimension to high dimension, that is, a mapping process from the input dimension N to the order of magnitude M: M (t) = L M (u(t)), where L M is the mapping function for mapping the pulse sequence, x M (t) is the high-dimensional liquid state vector of the liquid layer at time point t. This high-dimensional liquid state vector can essentially be understood as the membrane potential of the pulse neuron in the liquid layer at time point t. This will be described exemplarily later and no further explanation will be given here.

[0090] High-dimensional liquid state vectors at different time points can form a vector space, also known as the target vector space, which can map the pulse sequence to significantly separated trajectories. At the same time, the natural decay of the membrane potential in the target vector space can ensure that the historical information used to describe the dynamic changes in the operating state at historical moments gradually dissipates, thereby realizing fading memory rather than static storage.

[0091] The present disclosure is directed to potential changes of membrane potential, which can be described by using differential equations to take into account both biological interpretability and computational simplicity: Among them, τ m Represents the membrane time constant that controls the membrane potential decay rate, V(t) is the membrane potential at time t, V rest is the reset potential of the spiking neuron in the liquid layer, R m is the membrane resistance, and I(t) is the synaptic current, i.e., the pulse train.

[0092] By the above formula, the integral of the synaptic current is continuously accumulated to the membrane potential. When the membrane potential is raised by the pulse sequence or recursive feedback and reaches a preset potential threshold, the pulse neuron will immediately release a pulse and reset the membrane potential to a reference level, i.e., a reset potential, and enter a natural decay stage determined by the membrane time constant. Based on the continuous cycle of the above process, the number of pulse releases is accumulated to obtain the total number of pulse releases of the pulse neuron.

[0093] Since the above calculation process occurs when the pulse is triggered and the potential is updated, the pulse neuron has the characteristics of event-driven sparse computing, which can significantly reduce the number of operations and computing resources. Moreover, the pulse sequence carries dynamic characteristics, which captures the fine-grained changes reflected by temperature, pressure, or speed sensors during the degradation process of industrial equipment, and lays a foundation for high-fidelity timing input for the mapping process of the pulse sequence to the target vector space, further providing neural dynamics support for accurate characterization of degradation rules.

[0094] In the above embodiment, the low-dimensional pulse sequence is nonlinearly mapped to a high-dimensional liquid state vector containing fading memory by the dynamics of LSM, to fully explain the degradation process and long-short term timing dependence of industrial equipment. Moreover, the membrane potential and the total number of pulse releases are selected as the state parameters of the pulse neuron, which takes into account the macro timing characteristics carried by the discrete pulse sequence and the micro sub-threshold changes reflected by the membrane potential, to achieve high-precision and stable life prediction.

[0095] In an optional embodiment, the above step of mapping the pulse sequence to the target vector space specifically includes the following steps:

[0096] Obtaining the pulse release state of a plurality of pulse neurons at a first time point;

[0097] Performing weighted sum processing on the pulse release state and a target pulse sequence to obtain a target weighted result; wherein the target pulse sequence is a pulse sequence at the first time point;

[0098] Determining the membrane potential decay rate of the pulse neuron at the first time point;

[0099] Performing sum processing on the target weighted result and the membrane potential decay rate to obtain the membrane potential of the pulse neuron at a second time point; wherein the second time point is after the first time point.

[0100] Regarding the mapping process of the pulse sequence, it can be calculated by the following formula:

[0101] wherein s j(t) is the pulse release state of the jth pulse neuron in the liquid layer at the first time point t, s j (t) = 1 when the pulse neuron releases a pulse, s j (t) = 0; x k (t) is the pulse sequence received by the kth pulse neuron in the input encoding layer at the first time point t, and the pulse sequences received by all pulse neurons in the input encoding layer at time t are integrated together, that is, the target pulse sequence; is the recursive connection weight between the pulse neurons inside the liquid layer, that is, the first weight mentioned later; is used to describe the degree of influence of the target pulse sequence on the pulse neuron, that is, the second weight mentioned later; M is the number of pulse neurons in the liquid layer, and N is the number of pulse neurons in the input encoding layer; is the target weighted result; V i (t) is the membrane potential of the ith pulse neuron in the liquid layer at time t; Δt is the time difference between the first time point and the second time point t+Δt; τ m is the membrane time constant representing the control of the decay rate of the membrane potential; is the decay rate of the membrane potential; V i (t+Δt) is the membrane potential of the ith pulse neuron in the liquid layer at the second time point t+Δt. The above processing process will be further described later.

[0102] In the above embodiment, the potential that does not trigger the preset potential threshold can be ensured to decrease exponentially with time, so that the pulse neuron has the function of memory forgetting balance, and can retain short-term historical information without excessive accumulation. Moreover, through the cyclic interaction between the pulse neurons, the nonlinear integration of the historical input data can be promoted, so as to realize the linear separability of the membrane potential changing with time, thereby improving the prediction accuracy of the remaining useful life.

[0103] In an optional embodiment, the above step of weighting and summing the pulse release state and the target pulse sequence to obtain a target weighted result comprises the following steps:

[0104] The first weight corresponding to the pulse release state is weighted to obtain a first weighted result; wherein the first weight is used to indicate the connection strength between the pulse neuron and other pulse neurons;

[0105] The second weight corresponding to the target pulse sequence is weighted to obtain a second weighted result; wherein the second weight is used to indicate the degree of influence of the target pulse sequence on the pulse neuron;

[0106] The first weighting result and the second weighting result are summed to obtain the target weighting result.

[0107] Here, the first weighting result is where s j (t) is the pulse release state of the jth pulse neuron in the liquid layer at the first time point t, M is the number of pulse neurons in the liquid layer, is the first weight.

[0108] For the second time point, the pulse release state of each pulse neuron in the liquid layer and the respective corresponding first weight can be weighted, and the weighted result is integrated together, that is, the first weighting result.

[0109] Here, the second weighting result is where x k (t) is the pulse sequence received by the kth pulse neuron in the input encoding layer at the first time point t, N is the number of pulse neurons in the input encoding layer, is the second weight.

[0110] For the second time point, the pulse sequence received by each pulse neuron in the input encoding layer and the respective corresponding second weight can be weighted, and the weighted result is integrated together, that is, the second weighting result.

[0111] In the embodiments of the present disclosure, the pulse release state and the first weight are weighted to obtain the first weighting result, and the target pulse sequence and the second weight are weighted to obtain the second weighting result, so as to obtain the target weighting result. In this way, the activity state of other pulse neurons and the pulse sequence of other pulse neurons can be comprehensively considered, so as to divide these two aspects into the update process of the membrane potential, so as to improve the accuracy of the membrane potential update.

[0112] In an optional implementation, the above step of determining the membrane potential decay rate of the pulse neuron at the first time point specifically comprises the following steps:

[0113] Determine the time difference between the first time point and the second time point;

[0114] Exponentially operate the time difference and the membrane time constant to obtain an operation result;

[0115] The product of the operation result and the membrane potential of the pulse neuron at the first time point is determined as the membrane potential decay rate.

[0116] The present disclosure describes the process of calculating the membrane potential decay rate, that is, where Δt is the time difference between the first time point and the second time point, Vi (t) is the membrane potential of the i-th spiking neuron in the liquid layer at time t, τ m is a membrane time constant representing a control rate of decay of the membrane potential, is the operation result.

[0117] In the embodiments of the present disclosure, the decay rate of the membrane potential is determined according to the time difference, the membrane time constant and the membrane potential, and the decay rate of the membrane potential is integrated into the process of updating the membrane potential, so that the accuracy of updating the membrane potential can be further improved.

[0118] In the above embodiments, the membrane potential at the historical moment, the pulse release state and the input pulse sequence are combined to update the membrane potential at the current moment, and the historical input is continuously integrated by means of the cyclic weight coefficient while maintaining the sparsity of the pulse event, so that the key information is amplified and separated in the target vector space.

[0119] In an optional embodiment, the above step pulse encodes the sensor data, and the encoded pulse sequence is obtained, and the specific steps include the following steps:

[0120] According to at least one extraction dimension, an operating state indicator in the sensor data is extracted;

[0121] The preprocessing result of the operating state indicator is mapped to a pulse sending time;

[0122] The pulse sequence is generated based on the pulse sending time.

[0123] Here, the extraction dimension can be adjusted according to the actual application scene. Specifically, the extraction dimension can include a data change amplitude, a statistical feature or a keyword capable of reflecting a key performance of the industrial equipment, wherein the statistical feature can be a mean value, a maximum value and a variance.

[0124] In the embodiments of the present disclosure, the sensor data can be preprocessed first, and then the process of extracting the operating state indicator from the sensor data is performed, so that the preprocessing result of the operating state indicator can be directly obtained.

[0125] In addition, the operating state indicator can be extracted from the sensor data first, and then the operating state indicator is preprocessed, so that the preprocessing result of the operating state indicator can also be obtained. The preprocessing process has been mentioned in the foregoing of the present disclosure, and will not be described more herein, and the present disclosure does not limit the way of obtaining the preprocessing result.

[0126] In the embodiments of the present disclosure, a delay coding manner is adopted to extract an operation state indicator from sensor data and map the operation state indicator to a pulse dimension to generate a coded pulse sequence based on a pulse sending time, so that the pulse sequence can carry more accurate dynamic characteristics, and high-precision prediction can be performed based on the pulse sequence.

[0127] In an optional embodiment, the step of mapping the pre-processing result of the operation state indicator to the pulse sending time comprises the following steps:

[0128] determining the pulse sending time according to the pre-processing result and a mapping time, wherein the mapping time is a time from the pre-processing result to the pulse sending time.

[0129] wherein, is the pre-processing result corresponding to the i-th pulse neuron in the input coding layer, t i is the pulse sending time of the i-th neuron in the input coding layer, T enc is the mapping time, which can also be the total duration of the time window mentioned in the pre-processing process.

[0130] In a case where one sensor corresponds to one neuron in the input coding layer, is the pre-processing result of the sensor data of the i-th sensor, and the pre-processing result is 0-1. The higher the pre-processing result is, the earlier the pulse is triggered, that is, the earlier the pulse sending time is, so as to accurately carry the operation state indicator by using the pulse sending time, and ensure that only a small number of neurons in the entire input coding layer are in an active state at any time.

[0131] In addition, for a case where the dynamic range of the operation state indicator or the pre-processing result is large, a group delay coding can be further adopted: a same operation state indicator or pre-processing result drives a plurality of threshold-incremental pulse neurons in parallel, so that the pulse density is nonlinearly increased with the change of the numerical granularity, thereby improving the resolution while maintaining sparsity. By using this delay coding manner, the time dimension can be automatically embedded, the rhythm characteristics of the degradation curve can be completely retained without additional time embedding or position coding, and other data other than the operation state indicator do not generate pulses, thereby significantly reducing the subsequent calculation amount.

[0132] In an optional embodiment, the step of predicting the remaining service life of the industrial equipment based on the state parameter comprises the following steps:

[0133] performing weighted summation processing on the membrane potential and the total release number of the pulse neuron to obtain a third weighted result;

[0134] predict the remaining useful life based on the third weighted result.

[0135] The present disclosure adopts a (membrane potential-pulse total release times) hybrid decoding strategy in the output decoding layer to obtain a coherent and high-precision life estimation. In the prediction process, the pulse neuron of the output decoding layer accumulates its own pulse release times on one hand to obtain the pulse total release times to reflect the macroscopic timing activity pattern; on the other hand, reads the instantaneous membrane potential of the pulse neuron to capture the subtle wobble at the sub-threshold level.

[0136] In the embodiments of the present disclosure, for the above two aspects, linear fusion is performed according to the weight coefficients (0.3-0.6), and then a single full connection is performed to obtain a third weighted result, and the third weighted result is mapped to the RUL result.

[0137] In the above implementation, the pulse total release times guarantee the monotonic change of the output to the overall trend of the historical pulse stream, and the membrane potential compensates for the quantization error and jitter that easily occurs when only relying on the pulse total release times, thereby realizing smooth and continuous prediction under the driving of a sparse event stream; at the same time, the hybrid method is helpful to the stable training of the continuous transmission of the membrane potential gradient in the backpropagation. The entire decoding process is easy to implement on a field programmable gate array (FPGA) or a neuromorphic chip, and can ensure real-time performance and energy saving rate in a resource-limited scenario, for example, millisecond-level low latency and microwatt-level energy consumption.

[0138] In an optional implementation, the above steps further include the following steps:

[0139] deriving the derivative of the pulse release state function to obtain a derivative result;

[0140] optimizing the first weight and the second weight based on the product of the derivative result and a learning rate to obtain an optimized first weight and an optimized second weight; wherein the learning rate is used to indicate the adjustment amplitude of the first weight and the second weight;

[0141] taking the optimized first weight as the weight of the pulse release state, and taking the optimized second weight as the weight of the target pulse sequence.

[0142] Since the pulse release state function of the pulse neuron is in the form of a step, it is naturally not derivable and cannot be directly used with the traditional gradient descent strategy. Therefore, the present disclosure introduces the idea of alternative gradient in the backpropagation-through-time (BPTT) process: the forward propagation still adopts the real LIF firing mechanism, and the derivative of the step function is replaced by a differentiable smooth curve in the backpropagation phase, thereby realizing the chain transmission of errors while maintaining the accurate pulse timing.

[0143] Specifically, the following formula can be used for approximation in the back propagation phase:

[0144] Where S(V) is the pulse release state function, gamma is an adjustable hyperparameter, controlling the slope and smoothness of the surrogate gradient, V is the membrane potential, V threshold is the preset potential threshold. This smooth approximation avoids the zero gradient problem and suppresses the excessive gradient shock at the pulse edge, making the gradient design calculation stable and reliable.

[0145] Based on this, the product of the learning rate and the derivative result can be determined, and the first weight and the second weight are respectively reduced by the product to obtain the optimized first weight and the optimized second weight.

[0146] After using the above surrogate gradient, the weight coefficient can be updated iteratively by standard BPTT. The loss function mainly uses the mean square error (MSE) of the RUL predicted value and the RUL true value, and an asymmetric penalty term can be selected to emphasize the punishment for excessively optimistic prediction.

[0147] The basic form is: Where B is the batch size, i.e. the number of data samples processed at a time during the training process, y b and represent the RUL true value and the RUL predicted value of the bth data sample, respectively.

[0148] During the training process, the Adam optimizer can be used in combination with gradient norm clipping to prevent gradient explosion, and a sparse regularization constraint can be applied to the output average firing rate to further reduce training energy consumption while ensuring expression capability and providing a reliable parameter basis for subsequent low-power real-time inference.

[0149] Based on this, the present disclosure considers both prediction accuracy and computational economy when evaluating the pros and cons of the model. In terms of prediction accuracy, the root mean square error (RMSE) is used as the overall error measure. This index is more sensitive to large errors because it takes the square root of the mean of the squared deviations.

[0150] The specific expression formula is: An asymmetric penalty function is also introduced, and the specific mathematical expression is: Where B is the batch size, i.e. the number of data samples processed at a time during the training process, y b and represent the RUL true value and the RUL predicted value of the bth data sample, respectively.

[0151] This penalty function imposes a more severe penalty for overoptimism (i.e., overestimation of RUL), which is consistent with maintenance principles in industrial safety scenarios. Regarding computational efficiency, this disclosure can calculate the floating-point operations required for model inference to measure theoretical computational complexity; calculate video memory usage to reflect the demand for hardware storage resources during deployment; and record single-sample inference latency to evaluate the response speed of real-time applications. By simultaneously examining these performance and efficiency indicators, the applicability of each algorithm in actual application scenarios can be fully revealed.

[0152] Taking into account the results of both accuracy and efficiency, RSNN can significantly reduce computing and storage overhead while ensuring higher prediction accuracy, providing a practical solution for online device health management in resource-constrained scenarios.

[0153] Reference Figure 2 As shown, this is an architecture diagram of RSNN in the remaining useful life prediction process provided by an embodiment of the present disclosure. The architecture diagram includes an input coding layer, a hidden layer (liquid layer) and an output decoding layer. For the processing process of each layer in the architecture diagram, please refer to the previous text and no further description will be given here.

[0154] The following combination Figure 3 The above remaining useful life prediction process is described as follows:

[0155] S301: Acquire sensor data of industrial equipment.

[0156] Among them, sensor data is used to indicate the operating status of industrial equipment.

[0157] S302: Pulse encoding is performed on the sensor data to obtain a pulse sequence after encoding.

[0158] Among them, the pulse sequence is used to describe the dynamic changes of the operating status at different time points.

[0159] S303: Map the pulse sequence to the target vector space.

[0160] The target vector space is used to describe the membrane potential of spiking neurons at different time points.

[0161] S304: Determine the total number of pulse releases of the spiking neuron based on the potential change of the membrane potential.

[0162] S305: The membrane potential and the total number of pulse releases are used as state parameters.

[0163] S306: Predicting the remaining useful life of the industrial equipment based on the state parameters.

[0164] From the above description, it can be seen that the technical solution disclosed in this disclosure has the following advantages:

[0165] (1) High prediction accuracy: High-dimensional liquid state combined with hybrid decoding fully excavates the fine-grained timing characteristics of degenerative sequences, and the prediction accuracy is significantly higher than existing deep learning and traditional machine learning models.

[0166] (2) Real-time and energy-efficient: Event-driven, sparse pulse computing significantly reduces floating point operation and storage requirements; millisecond-level inference can be achieved on GPUs (Graphics Processing Units), FPGAs or neuromorphic chips, meeting online monitoring and edge deployment requirements.

[0167] (3) Strong model generalization and adaptability: Liquid state machine structure does not need to reconstruct neural network for specific equipment, only needs to adjust a small number of hyperparameters to migrate to different working conditions or new equipment, reducing development and maintenance costs.

[0168] (4) Stable training and easy implementation: The alternative gradient method is used to solve the non-differentiable problem of pulse network, and the end-to-end training process is used to ensure convergence stability and avoid complex manual design and feature engineering.

[0169] (5) Explainability and visualization friendly: Hybrid decoding preserves continuous information of membrane potential, combined with liquid layer pulse mode, can intuitively present dynamic evolution of degenerative process, facilitating operation and maintenance personnel to understand and make decisions.

[0170] (6) Adapt to low-power hardware: Network sparsity and event-based computing characteristics are compatible with neuromorphic chips and other low-power processors, enabling long-term operation on device side without frequent maintenance or power replacement.

[0171] Based on the same inventive concept, the present disclosure also provides a remaining useful life prediction device corresponding to the remaining useful life prediction method. Since the device in the present disclosure solves the problem by similar principles as the above-mentioned remaining useful life prediction method, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described again.

[0172] Referring to Figure 4 FIG. 1 is a schematic diagram of a remaining useful life prediction device provided by the present disclosure, which comprises an acquisition unit 40, a pulse encoding unit 41, a parameter determination unit 42, and a prediction unit 43.

[0173] The acquisition unit is configured to acquire sensor data of an industrial device, wherein the sensor data is used to indicate the running state of the industrial device.

[0174] The pulse encoding unit is configured to pulse encode the sensor data, and obtain a pulse sequence after encoding, wherein the pulse sequence is used to describe the dynamic change of the running state at different time points.

[0175] determine a state parameter of the pulse neuron based on the pulse sequence;

[0176] predict a remaining service life of the industrial equipment based on the state parameter.

[0177] In a possible implementation, the apparatus is further configured to:

[0178] map the pulse sequence to a target vector space, wherein the target vector space is used to describe a membrane potential of the pulse neuron at different time points;

[0179] determine a total number of pulse releases of the pulse neuron based on a potential change of the membrane potential;

[0180] use the membrane potential and the total number of pulse releases as the state parameter.

[0181] In a possible implementation, the apparatus is further configured to:

[0182] obtain a pulse release state of a plurality of pulse neurons at a first time point;

[0183] perform weighted summation processing on the pulse release state and a target pulse sequence to obtain a target weighted result, wherein the target pulse sequence is a pulse sequence at the first time point;

[0184] determine a membrane potential decay rate of the pulse neuron at the first time point;

[0185] perform summation processing on the target weighted result and the membrane potential decay rate to obtain a membrane potential of the pulse neuron at a second time point, wherein the second time point is after the first time point.

[0186] In a possible implementation, the apparatus is further configured to:

[0187] perform weighted processing on a first weight and the pulse release state to obtain a first weighted result, wherein the first weight is used to indicate a connection strength between the pulse neuron and other pulse neurons;

[0188] perform weighted processing on a second weight and the target pulse sequence to obtain a second weighted result, wherein the second weight is used to indicate an influence degree of the target pulse sequence on the pulse neuron;

[0189] perform summation processing on the first weighted result and the second weighted result to obtain the target weighted result.

[0190] In a possible implementation, the apparatus is further configured to:

[0191] determining a time difference between the first time point and the second time point;

[0192] performing an exponential operation on the time difference and a membrane time constant to obtain an operation result;

[0193] determining a product of the operation result and a membrane potential of the spiking neuron at the first time point as the membrane potential decay rate.

[0194] In a possible implementation, the apparatus further includes:

[0195] extracting an operating state indicator in the sensor data according to at least one extraction dimension;

[0196] mapping a preprocessing result of the operating state indicator to a pulse sending time;

[0197] generating the pulse sequence based on the pulse sending time.

[0198] In a possible implementation, the apparatus further includes:

[0199] determining the pulse sending time according to the preprocessing result and a mapping time, where the mapping time is a time from the preprocessing result to the pulse sending time.

[0200] In a possible implementation, the apparatus further includes:

[0201] performing weighted summation processing on the membrane potential of the spiking neuron and the total number of pulse releases to obtain a third weighted result;

[0202] predicting the remaining useful life based on the third weighted result.

[0203] In a possible implementation, the apparatus further includes:

[0204] performing derivation on the pulse release state function to obtain a derivation result;

[0205] optimizing the first weight and the second weight based on a product of the derivation result and a learning rate to obtain an optimized first weight and an optimized second weight, where the learning rate is used to indicate an adjustment range of the first weight and the second weight;

[0206] using the optimized first weight as the weight of the pulse release state, and using the optimized second weight as the weight of the target pulse sequence.

[0207] The description of the processing flow of each module in the apparatus and the interaction flow between the modules can refer to the related description in the above method embodiments, and will not be repeated here.

[0208] Corresponding to Figure 1 The remaining useful life prediction method in the method, the embodiment of the disclosure also provides an electronic device 50, as shown in Figure 5 The electronic device 50 provided by the embodiment of the disclosure is shown in the structure schematic diagram, which includes:

[0209] The processor 51, the memory 52, and the bus 53; the memory 52 is used to store the execution instructions, including the memory 521 and the external memory 522; the memory 521 here is also called the internal memory, which is used to temporarily store the operation data in the processor 51 and exchange data with the external memory 522 such as a hard disk; the processor 51 exchanges data with the external memory 522 through the memory 521; when the electronic device 50 is running, the processor 51 and the memory 52 communicate through the bus 53, so that the processor 51 executes the following instructions:

[0210] Obtain sensor data of an industrial device; wherein the sensor data is used to indicate the running state of the industrial device;

[0211] Pulse code the sensor data, and obtain a pulse sequence after coding; wherein the pulse sequence is used to describe the dynamic change of the running state at different time points;

[0212] Input the pulse sequence into a pulse neuron to determine a state parameter of the pulse neuron;

[0213] Based on the state parameter, predict the remaining useful life of the industrial device.

[0214] The embodiment of the disclosure also provides a computer readable storage medium, which stores a computer program, and the computer program is run by a processor to execute the steps of the remaining useful life prediction method described in the method embodiment. Wherein the storage medium can be a volatile or non-volatile computer readable storage medium.

[0215] The embodiment of the disclosure also provides a computer program product 60, as shown in Figure 6 The computer program product 60 provided by the embodiment of the disclosure is shown in the structure schematic diagram, which carries a computer program 61, and the computer program 61 includes a program that can be used to execute the steps of the remaining useful life prediction method described in the method embodiment, which can be specifically referred to in the above method embodiment, and will not be repeated here.

[0216] In the above, the remaining useful life prediction method, device, equipment, medium and product according to the embodiments of the present disclosure are described with reference to the drawings. First, sensor data of an industrial equipment is acquired, wherein the sensor data is used to indicate the running state of the industrial equipment. Then, the sensor data is pulse coded, and a pulse sequence is obtained after coding, wherein the pulse sequence is used to describe the dynamic change of the running state at different time points. The pulse sequence is input into a pulse neuron to determine a state parameter of the pulse neuron. Finally, the remaining useful life of the industrial equipment is predicted based on the state parameter. By pulse coding the sensor data, the time sequence dependency of the running state can be effectively learned, and the remaining useful life of the industrial equipment can be predicted according to the state parameter of the pulse neuron, so as to mine the fine-grained time sequence features in the pulse sequence, and further improve the prediction accuracy and prediction accuracy of the remaining useful life.

[0217] The basic principles of the present disclosure are described above in combination with specific embodiments, but it should be pointed out that the advantages, advantages, effects and the like mentioned in the present disclosure are only examples and not limitations, and these advantages, advantages, effects and the like cannot be considered as the must-have of each embodiment of the present disclosure. In addition, the specific details of the above disclosure are only for the purpose of example and for the purpose of understanding, and are not limited to the above specific details, and the present disclosure is not limited to the above specific details.

[0218] The block diagrams of the devices, apparatuses, equipment, systems involved in the present disclosure are only illustrative examples and are not intended to require or imply that the connection, arrangement, configuration must be as shown in the block diagram. As those skilled in the art will recognize, these devices, apparatuses, equipment, systems can be connected, arranged, configured in any manner. Words such as "include", "contain", "have" and the like are open-ended words, mean "including but not limited to", and can be used interchangeably. The words "or" and "and" used herein mean the word "and / or", and can be used interchangeably unless the context clearly indicates otherwise. The word "such as" used herein means the phrase "such as but not limited to", and can be used interchangeably.

[0219] In addition, as used herein, "or" used in the list of items starting with "at least one of indicates a separate list, so that, for example, "at least one of A, B or C" means A or B or C, or AB or AC or BC, or ABC (i.e. A and B and C). In addition, the word "exemplary" does not mean that the described example is preferred or better than other examples.

[0220] It should also be noted that in the systems and methods of the present disclosure, each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions of the present disclosure.

[0221] Various changes, modifications, and improvements in the technologies described herein can be made without departing from the teachings of the technology defined by the appended claims. Moreover, the scope of the claims of this disclosure is not limited to the specific aspects described above. Rather, the specific aspects are provided for illustration purposes only to assist in understanding the claims, and the scope of the claims is not limited to the specific aspects described above. Accordingly, the appended claims include within their scope all processes, machines, manufactures, compositions of matter, means, methods, or steps, whether now known or later developed, that fall within the scope of the claims.

[0222] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other aspects without departing from the scope of the disclosure. Thus, the present disclosure is not intended to be limited to the aspects shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0223] The above description has been presented for the purpose of illustration and description. Furthermore, this description is not intended to limit the embodiments of the disclosure to the forms disclosed herein. Although several example aspects and embodiments have been discussed above, those of ordinary skill in the art will appreciate a variety of modifications, alternatives, permutations, additions, and sub-combinations of the aspects and embodiments disclosed herein.

Claims

1. A method for predicting remaining useful life, characterized in that: include: Acquiring sensor data of industrial equipment; wherein the sensor data is used to indicate the operating status of the industrial equipment; Pulse encoding is performed on the sensor data to obtain a pulse sequence after encoding; wherein the pulse sequence is used to describe the dynamic changes of the operating state at different time points; inputting the pulse sequence into a spiking neuron to determine a state parameter of the spiking neuron; Based on the state parameters, a remaining useful life of the industrial equipment is predicted.

2. The method according to claim 1, characterized in that Inputting the pulse sequence into a spiking neuron to determine a state parameter of the spiking neuron includes: Mapping the pulse sequence to a target vector space; wherein the target vector space is used to describe the membrane potential of the spiking neuron at different time points; determining the total number of pulse releases of the spiking neuron based on the potential change of the membrane potential; The membrane potential and the total number of pulse releases are used as the state parameters.

3. The method according to claim 2, characterized in that Mapping the pulse sequence to a target vector space includes: Obtaining the pulse release states of multiple pulse neurons at a first time point; Performing weighted summation processing on the pulse release state and the target pulse sequence to obtain a target weighted result; wherein the target pulse sequence is the pulse sequence at the first time point; determining a membrane potential decay rate of the spiking neuron at the first time point; The target weighted result and the membrane potential decay rate are summed to obtain the membrane potential of the spiking neuron at a second time point; wherein the second time point is after the first time point.

4. The method according to claim 3, characterized in that The weighted summing process of the pulse release state and the target pulse sequence to obtain a target weighted result includes: Performing weighted processing on the first weight and the corresponding pulse release state to obtain a first weighted result; wherein the first weight is used to indicate the connection strength between the spiking neuron and other spiking neurons; Performing weighted processing on the second weight and the corresponding target pulse sequence to obtain a second weighted result; wherein the second weight is used to indicate the degree of influence of the target pulse sequence on the spiking neuron; The first weighted result and the second weighted result are summed to obtain the target weighted result.

5. The method according to claim 3, characterized in that Determining the membrane potential decay rate of the spiking neuron at the first time point includes: determining a time difference between the first time point and the second time point; performing an exponential operation on the time difference and the membrane time constant to obtain an operation result; The product of the calculation result and the membrane potential of the spiking neuron at the first time point is determined as the membrane potential decay rate.

6. The method according to claim 1, characterized in that The pulse encoding of the sensor data to obtain a pulse sequence after encoding includes: extracting an operating status indicator from the sensor data according to at least one extraction dimension; Mapping the preprocessing result of the operating status indicator to a pulse sending time; The pulse sequence is generated based on the pulse transmission time.

7. The method according to claim 6, characterized in that Mapping the preprocessing result of the operating status indicator to the pulse sending time includes: The pulse sending time is determined according to the preprocessing result and the mapping time; wherein the mapping time is the time from mapping the preprocessing result to the pulse sending time.

8. The method according to claim 1, characterized in that The predicting of the remaining service life of the industrial equipment based on the state parameter includes: Performing weighted sum processing on the membrane potential of the spiking neuron and the total number of pulse releases to obtain a third weighted result; The remaining useful life is predicted based on the third weighted result.

9. The method according to claim 4, characterized in that The method further comprises: Derivative the pulse release state function to obtain the derivative result; Optimizing the first weight and the second weight based on the product of the derivation result and the learning rate to obtain an optimized first weight and an optimized second weight; wherein the learning rate is used to indicate an adjustment range of the first weight and the second weight; The optimized first weight is used as the weight of the pulse release state, and the optimized second weight is used as the weight of the target pulse sequence.

10. A device for predicting remaining useful life, characterized in that: include: An acquisition unit, configured to acquire sensor data of an industrial device; wherein the sensor data is used to indicate an operating status of the industrial device; A pulse encoding unit, configured to pulse encode the sensor data to obtain a pulse sequence after encoding; wherein the pulse sequence is used to describe the dynamic changes of the operating state at different time points; a parameter determination unit, configured to input the pulse sequence into a spiking neuron to determine a state parameter of the spiking neuron; A prediction unit is used to predict the remaining service life of the industrial equipment based on the state parameter.

11. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the steps of the remaining useful life prediction method as described in any one of claims 1 to 9 are performed.

12. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the steps of the remaining useful life prediction method according to any one of claims 1 to 9.

13. A computer program product, characterized in that The computer program product is stored in a storage medium, and the program product is executed by at least one processor to implement the remaining useful life prediction method according to any one of claims 1 to 9.

Images