A non-invasive method and system for monitoring the operating state of a device based on the electromagnetic spectrum

Abstract

The present invention relates to the technical field of condition monitoring, and in particular to a non-invasive device operation condition monitoring method and system based on electromagnetic spectrum. The method includes dynamic multi-dimensional spectrum fingerprint extraction; prototype generation based on the extracted multi-dimensional spectrum fingerprints; in-prototype contrast learning and inter-prototype contrast learning respectively based on the generated prototypes; constructing a prototype contrast loss function based on the contrast learning; image modeling based on the extracted multi-dimensional spectrum fingerprints, and constructing an image contrast loss function based on the image modeling; performing function mixing based on a loss function mixing strategy to obtain an overall loss function; determining the device state based on the overall loss function; and outputting a discrimination result. It effectively solves the problem of wide-band electromagnetic signal aliasing in the dense scene of petroleum equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of condition monitoring, and in particular, to a non-invasive device operation condition monitoring method and system based on electromagnetic spectrum. Background Art

[0002] In the field of oil equipment, the health monitoring of electrical equipment and fracturing equipment is crucial for ensuring the safe and efficient production of oil and gas fields. However, existing monitoring technologies face multiple challenges in practical applications. Taking typical equipment such as the plunger pump, high-pressure pipe string, and explosion-proof motor of a fracturing truck as an example, their working environment is extremely harsh (high temperature, high pressure, strong vibration, corrosive medium). Traditional contact sensors (such as vibration and temperature sensors) need to be directly installed on the surface or inside of the equipment, resulting in complex installation processes and high maintenance costs. For example, the vibration sensor of the plunger rod of a fracturing pump needs to be frequently calibrated and is easily corroded and damaged by mud, leading to a decline in the reliability of the monitoring system; the installation of an embedded temperature sensor inside an explosion-proof motor requires shutdown and modification, seriously affecting production continuity. In addition, the extremely flammable and explosive environment at the oil field site has extremely high requirements for the explosion-proof level of sensors, further increasing the deployment cost and safety risk.

[0003] In response to the non-invasive monitoring requirement, existing technologies attempt to analyze the equipment condition through electromagnetic radiation signals. However, the electromagnetic environment of oil equipment is particularly complex: equipment such as frequency converters and high-power motors of fracturing trucks are densely distributed, and their wide-band electromagnetic signals (30 MHz - 3 GHz) are superimposed on each other. Traditional spectrum sensing technologies rely on fixed-band filtering or single-energy threshold judgment, and it is difficult to separate the characteristics of target equipment from the mixed signals. For example, harmonic interference in similar frequency bands may be generated due to bearing wear of a fracturing pump and insulation aging of a motor. Existing methods lack multi-dimensional feature correlation analysis (such as time-domain transient response and modulation sideband characteristics), and misjudge environmental noise or signals of adjacent equipment as target faults, resulting in a false alarm rate of over 30%.

[0004] In addition, oil fracturing operations are mostly located in remote areas, with strict requirements for the real-time and offline processing capabilities of the monitoring system. Existing solutions mostly rely on cloud transmission of raw spectrum data, but the network coverage at the fracturing site is unstable, and the transmission delay of high-bandwidth signals is serious, unable to meet the real-time fault warning requirements. A few local embedded terminals are limited by the computing architecture and only support single-band FFT analysis or shallow machine learning models, and it is difficult to process the high-frequency and high-dimensional electromagnetic fingerprint data of oil equipment. For example, the early wear characteristics of the power-end bearing of a fracturing truck only exist in transient pulses of 2 - 5 ms. Due to insufficient computing power, traditional embedded systems cannot complete feature extraction and matching within a millisecond time window, resulting in missed detection of minor faults.

[0005] More critically, the failure modes of oil equipment are strongly coupled. For example, the failure of the plunger pump seal may cause the lubricating oil temperature to rise, which in turn leads to a decline in the insulation performance of the motor winding. This chain of failures requires comprehensive correlation analysis of multi-physical field data such as electromagnetism, temperature, and vibration. Existing non-invasive technologies are limited to a single physical field (such as only analyzing electromagnetic radiation), lacking the ability to perform multi-dimensional modeling of the equipment operating state, making it difficult to trace the root cause of failures and predict the lifespan. They can only provide fragmented alarm information and cannot support preventive maintenance decisions.

[0006] The above problems indicate that existing monitoring technologies face four major contradictions in the scenario of oil equipment: the contradiction between the non-invasive requirement and signal aliasing interference, the contradiction between the high-dimensional feature requirement and insufficient edge computing power, the contradiction between the real-time requirement and transmission dependence, and the contradiction between the failure coupling and single analysis dimension. Therefore, there is an urgent need for a new monitoring technology for oil equipment that can achieve high-precision signal separation and multi-dimensional feature extraction in a complex electromagnetic environment and complete real-time diagnosis through an embedded terminal, thus breaking through the bottlenecks of reliability, real-time performance, and multi-dimensional analysis capabilities. Summary of the Invention

[0007] To solve the problems mentioned above, the present invention provides a non-invasive method and system for monitoring the operating state of equipment based on the electromagnetic spectrum.

[0008] In the first aspect, a non-invasive method for monitoring the operating state of equipment based on the electromagnetic spectrum provided by the present invention adopts the following technical solutions:

[0009] A non-invasive method for monitoring the operating state of equipment based on the electromagnetic spectrum includes:

[0010] Dynamic multi-dimensional spectrum fingerprint extraction;

[0011] Prototype generation based on the extracted multi-dimensional spectrum fingerprints;

[0012] High-dimensional feature generation through a time series encoder;

[0013] Intra-prototype contrast learning and inter-prototype contrast learning are respectively performed based on the generated prototypes;

[0014] Based on the contrast learning, a prototype contrast loss function is constructed;

[0015] Image modeling is performed based on the extracted multi-dimensional spectrum fingerprints, and an image contrast loss function is constructed based on the image modeling;

[0016] Based on the loss function mixing strategy, function mixing is performed to obtain an overall loss function;

[0017] Based on the overall loss function, the parameters of the time series encoder are optimized;

[0018] Device status discrimination is performed based on the high-dimensional features generated by the time series encoder;

[0019] Output the discrimination result.

[0020] Furthermore, the dynamic multi-dimensional spectrum fingerprint extraction includes synchronously collecting the original electromagnetic radiation signals within the frequency band of 0.1 MHz - 3 GHz at a sampling rate of 200 MHz through the multi-band parallel sampling circuit of the radio frequency sensor , covering the fundamental wave, harmonic and modulation sideband features of the device; among them, based on the adaptive dynamic range compression method, the dynamic range of the signal amplitude is detected in real time at the embedded terminal, and the dynamic gain G The calculation is expressed as:

[0021]

[0022] Where is the gain adjustment amount, is the peak value of the signal amplitude within the current sampling window, and the preset target amplitude range is , automatically adjusting the gain of the front-end amplifier to avoid data saturation caused by strong interference signals or loss of weak features.

[0023] Furthermore, aiming at the signal aliasing problem in the complex electromagnetic environment of the industrial field, the dynamic multi-dimensional spectrum fingerprint extraction system innovatively adopts a two-stage anti-aliasing filtering architecture: the first stage uses an analog-domain tunable band-pass filter to dynamically adjust the center frequency by real-time monitoring of the fundamental frequency characteristics of the target device Dynamically calculate the quality factor Q value:

[0024]

[0025] Where is the preset bandwidth, defaulting to , and the transfer function of the second-order active band-pass filter can be abbreviated as:

[0026]

[0027] Where s is the complex frequency variable, is the center angular frequency, K is the passband gain, and the output signal of the first-stage filter is:

[0028]

[0029] Where are the Laplace transform and its inverse transform respectively.

[0030] The second stage performs time-frequency domain energy threshold segmentation on the aliasing residual components through wavelet packet transform in the digital domain, and eliminates the vibration coupling noise from the original electromagnetic signal, which is expressed as:

[0031]

[0032] Among them, J is the number of wavelet packet decomposition layers, j is the current decomposition layer, k is the frequency band index, is the wavelet packet coefficient, is the time domain form of the bd6 wavelet basis function.

[0033] Furthermore, the dynamic multi-dimensional spectrum fingerprint extraction further includes spatio-temporal alignment of multi-modal signals. Among them, the rising edge of the electromagnetic signal detected by the radio frequency sensor is used as the global trigger source, and the acquisition windows of the infrared thermal imager and the microphone are synchronously started through the FPGA hardware trigger pin to achieve microsecond-level time alignment of cross-modal data. Taking the electromagnetic trigger timestamp as the reference, the time window of the multi-modal signal is intercepted:

[0034]

[0035] Among them, is the temperature signal, is the acoustic vibration signal, is the time window width.

[0036] Furthermore, the prototype generation according to the extracted multi-dimensional spectrum fingerprints includes randomly generating two different enhanced views for each spectrum fingerprint using each enhancement method in the data enhancement library containing G types of enhancement types, obtaining two sets of enhanced view sets; then inputting the enhanced views into the temporal encoder to obtain its high-dimensional hidden representation , :

[0037]

[0038] Among them are the learnable parameters of the time encoder, The prototype calculation formula is:

[0039]

[0040] Among them , G represents the number of enhancement types, represents the non-linear projection function.

[0041] Furthermore, the in-prototype contrast learning and inter-prototype contrast learning are respectively performed based on the generated prototypes, including achieving a uniform distribution of different spectral fingerprint prototypes in the representation space through in-prototype contrast learning, where the value controls the penalty intensity of negative samples, and the distance metric function is used to calculate the distance j between the k th and the th spectral enhancement views , and then the softmax function is used to map the distance to a value . The calculation formula is:

[0042] .

[0043] Furthermore, the prototype contrast loss function is constructed based on contrast learning, including using various data augmentation means through in-prototype contrast learning and inter-prototype contrast learning to construct a strongly generalized representation and improve the classification performance of downstream tasks. The overall prototype-based contrast loss is defined as:

[0044]

[0045] where α is a hyperparameter.

[0046] Furthermore, the image contrast loss function is constructed based on image modeling, including obtaining a representation through an image encoder: , and extracting the corresponding representation through a time series encoder; by screening the information suitable for sequence-image contrast learning, more appropriate time series representations and and image representations and are obtained from

[0047]

[0048] . The spectral sequence-image contrast loss function is expressed as: where i represents the contrast loss between the th image representation and all spectral sequence representations in the batch, while i represents the contrast between the

[0049] th spectral sequence and the image representation. , temporal encoder via back-propagation The update of the parameters, the update parameter gradient is:

[0050]

[0051] in Indicates encoder parameters, symbol Represents partial derivative, which is used to describe the rate of change of a variable in a multivariable function.

[0052] In a second aspect, a non-intrusive equipment operation status monitoring system based on electromagnetic spectrum includes:

[0053] The data acquisition module is configured to extract dynamic multi-dimensional spectrum fingerprints;

[0054] The prototype module is configured to generate a prototype according to the extracted multi-dimensional spectrum fingerprint; perform intra-prototype contrast learning and inter-prototype contrast learning based on the generated prototypes; and construct a prototype contrast loss function based on contrast learning;

[0055] The image module is configured to perform image modeling according to the extracted multi-dimensional spectrum fingerprint, and construct an image contrast loss function based on the image modeling;

[0056] A mixing module is configured to perform function mixing based on a loss function mixing strategy to obtain an overall loss function;

[0057] The discrimination module is configured to discriminate the device state based on the overall loss function and output a discrimination result.

[0058] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for example, a non-intrusive device operation status monitoring method based on an electromagnetic spectrum.

[0059] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being suitable for being loaded by the processor and executing the aforementioned electromagnetic spectrum-based non-invasive device operation status monitoring method.

[0060] In summary, the present invention has the following beneficial technical effects:

[0061] The technical solution proposed by the present invention has achieved multiple breakthroughs in the field of non-invasive monitoring, and its significant beneficial effects are reflected in multiple key technical dimensions. In terms of high-precision feature extraction in a complex electromagnetic environment, through the original dynamic multi-dimensional spectrum fingerprint technology, which organically integrates the time-domain transient waveform matching algorithm, the frequency-domain harmonic energy dynamic weighting strategy, and the modulation-domain sideband symmetry depth analysis, the industry problem of wide-band electromagnetic signal aliasing in the dense scenario of petroleum equipment has been successfully overcome. This core technology has reduced the misjudgment rate of target device feature extraction from the industry average of 30% to less than 5% in a cliff-like manner, and can be achieved only by making full use of the data collected by existing RF sensors, achieving significant improvements in both feature extraction accuracy and economy.

[0062] Aiming at the bottleneck problem of limited edge computing resources, the lightweight prototype enhancement architecture carefully designed in this patent has successfully achieved the real-time parsing ability of high-dimensional spectrum fingerprints through innovative feature view calculation optimization on a general ARM / RISC-V processor platform, and strictly controlled the end-to-end full-process processing delay within the technical index of 50ms, fully meeting the real-time processing performance requirements of high-frequency transient signals (such as 2ms-level short pulses). Compared with traditional embedded terminal solutions, this technical solution has increased the computing efficiency by up to 8 times, and at the same time completely gets rid of the dependence on dedicated hardware acceleration modules, achieving the best balance between computing performance and deployment cost, providing stable and reliable edge computing technical support for the health monitoring of petroleum equipment.

[0063] In the field of real-time diagnosis under offline working conditions, the end-side full-process diagnosis algorithm chain innovatively constructed in this patent has perfectly realized the localized closed-loop processing mechanism of signal acquisition, spectrum fingerprint extraction, and fault matching. The data volume of a single diagnosis result has been optimized and compressed to less than 1KB technical standard, and the communication bandwidth requirement has been reduced by more than 99% year-on-year. This major technical breakthrough enables the system to operate completely autonomously at the fracturing operation site with poor network coverage, completely eliminating the dependence on cloud data transmission, and ensuring the absolute reliability and millisecond-level timeliness of the real-time warning function.

[0064] In addition, the present invention has also creatively broken through the technical limitations of traditional single-field data analysis. Through the advanced algorithm of contrast learning based on the enhanced prototype, the electromagnetic fingerprint features, infrared temperature field, and acoustic vibration signals are accurately aligned in time and space, and a leading multi-physical field coupling analysis ability in the industry is constructed. This innovative technology has significantly improved the accuracy of equipment operation state recognition by more than 25%, while fully exploiting the value of existing sensor data and completely avoiding additional hardware investment, providing unprecedented comprehensive technical support for the comprehensive fault diagnosis of petroleum equipment.

[0065] Generally speaking, the technical solution of this patent not only achieves a breakthrough at the core algorithm level, but also significantly improves the monitoring efficiency, reduces the deployment cost, enhances the adaptability and reliability of the system in the actual application scenario, providing an innovative solution for the field of oil equipment health monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a schematic diagram of a non-invasive device operating state monitoring method based on electromagnetic spectrum in Embodiment 1 of the present invention.

[0067] Figure 2 is a schematic diagram of the method in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0068] The present invention will be further described in detail below with reference to the accompanying drawings.

[0069] Embodiment 1

[0070] Referring to Figure 1 , a non-invasive device operating state monitoring method based on electromagnetic spectrum in this embodiment includes:

[0071] Dynamic multi-dimensional spectrum fingerprint extraction;

[0072] Prototype generation according to the extracted multi-dimensional spectrum fingerprints;

[0073] High-dimensional feature generation through a time series encoder;

[0074] Prototype intra-comparison learning and prototype inter-comparison learning are respectively performed based on the generated prototypes;

[0075] Based on the comparison learning, a prototype comparison loss function is constructed;

[0076] Image modeling is performed according to the extracted multi-dimensional spectrum fingerprints, and an image comparison loss function is constructed based on the image modeling;

[0077] Based on the loss function mixing strategy, function mixing is performed to obtain an overall loss function;

[0078] Based on the overall loss function, the parameters of the time series encoder are optimized;

[0079] Based on the high-dimensional features generated by the time series encoder, the device state is discriminated;

[0080] The discrimination result is output.

[0081] Specifically:

[0082] (1) Define,

[0083] The spectrum sequence data is defined as where Mis the number of data dimensions, T is the number of time steps. When M = 1, the sequence is called a one-dimensional time series. When M > 1, the sequence is called a multi-dimensional time series. In the dense scenario of oil equipment, the wideband electromagnetic signals generated by equipment such as fracturing trucks and frequency converters are superimposed on each other, and the spectrum sequence data is M a multi-dimensional time series with X being a frequency band fingerprint sample.

[0084] The spectrum fingerprint recognition task aims to assign predefined class labels to each spectrum fingerprint, and each label represents a different operating state of the oil equipment. Given a dataset , where each spectrum fingerprint , n is the number of samples. The goal of this work is to learn a mapping function that corresponds each spectrum fingerprint to the operating state of the oil equipment, C where

[0085] is the total number of classes of the equipment operating state.

[0086] The data acquisition module of the present invention is based on existing sensors (including radio frequency sensors, infrared thermal imagers, industrial microphones, etc.), and adopts a multi-channel adaptive synchronous triggering mechanism and an anti-aliasing dynamic sampling strategy to achieve precise capture of multi-physical field signals.

[0087] (2.1) Wideband electromagnetic signal acquisition,

[0088] Through the multi-band parallel sampling circuit of the radio frequency sensor, the original electromagnetic radiation signals in the frequency band of 0.1 MHz - 3 GHz are synchronously acquired at a sampling rate of 200 MHz , covering the fundamental wave, harmonic and modulation sideband characteristics of the equipment. The adaptive dynamic range compression (ADRC) technology is adopted to detect the dynamic range of the signal amplitude in real time at the embedded terminal, and the dynamic gain G is calculated as follows:

[0089]

[0090] where is the gain adjustment amount, is the peak value of the signal amplitude within the current sampling window, and the preset target amplitude range is . The gain of the front-end amplifier is automatically adjusted (adjustment step ±6 dB / μs) to avoid data saturation caused by strong interference signals or loss of weak features.

[0091] The normalized electromagnetic radiation signal is:

[0092]

[0093] And based on the operating conditions of the device (such as the motor startup stage of the fracturing truck), dynamically switch the sampling mode: continuous sampling with a 1 MHz bandwidth is adopted in the steady state stage, and the instantaneous sampling rate is increased to 500 MHz when triggered by a transient pulse (current mutation at the 2 ms level) to ensure complete capture of high-frequency transient characteristics.

[0094] (2.2) Anti-aliasing and noise suppression,

[0095] Aiming at the problem of electromagnetic signal aliasing in the industrial field, a two-stage anti-aliasing filter is preset at the acquisition end. The first stage uses an analog-domain tunable band-pass filter ( Q The value dynamic range is 5 - 100), and according to the fundamental frequency of the target device such as the fundamental frequency of the motor of the fracturing truck is 50 Hz), dynamically calculate Q value:

[0096]

[0097] where is the preset bandwidth, and the default value is 10% The transfer function of the second-order active band-pass filter can be abbreviated as:

[0098]

[0099] where s is the complex frequency variable, is the center angular frequency, K is the passband gain, and the output signal of the first-stage filter is:

[0100]

[0101] where are the Laplace transform and its inverse transform respectively.

[0102] The second stage performs time-frequency domain energy threshold segmentation on the aliasing residual components through wavelet packet transform (selecting the db6 wavelet basis) in the digital domain to eliminate vibration coupling noise from the original electromagnetic signal:

[0103]

[0104] where J is the wavelet packet decomposition layer number, j is the current decomposition layer number, k is the frequency band index, is the wavelet packet coefficient, is the time domain form of the bd6 wavelet basis function.

[0105] (2.3) Spatiotemporal alignment of multi-modal signals,

[0106] For the asynchronous data streams of the infrared thermal imager and the acoustic vibration sensor, the rising edge of the electromagnetic signal detected by the radio frequency sensor (accuracy ±10 μs) is used as the global trigger source. Through the FPGA hardware trigger pin, the acquisition windows of the infrared thermal imager (temperature field) and the microphone (acoustic vibration signal) are synchronously started to achieve microsecond-level time alignment of cross-modal data. Taking the electromagnetic trigger timestamp as the reference, the time windows of the multi-modal signals are intercepted:

[0107]

[0108] where is the temperature signal, is the acoustic vibration signal, is the time window width. The multi-dimensional spectral fingerprint output fuses the electromagnetic frequency band, temperature, and acoustic vibration characteristics into a unified matrix to provide a reference for subsequent calculations:

[0109]

[0110] where represents the energy sequence of the m th electromagnetic frequency band.

[0111] (3) Encoder design

[0112] Views generated from different samples can capture the characteristics of spectral fingerprints under various transformations. In the present invention, the enhanced views of spectral fingerprints are aggregated into prototypes to minimize the potential negative impacts brought by the semantic changes that may be introduced by specific enhancements. At the same time, aggregating enhanced views can generate more stable sample representations, reduce the randomness of the representations, and highlight the essential features of spectral fingerprints. For each spectral fingerprint, two different enhanced views are randomly generated for each enhancement method in the data enhancement library containing G types of enhancements, obtaining two sets of enhanced view sets:

[0113]

[0114] where and respectively represent the two enhanced views generated for the k th sample i using different random parameters by the th enhancement method.

[0115] Subsequently, the enhanced views are input into the temporal encoder to obtain its high-dimensional hidden representation , :

[0116]

[0117] Among them are the learnable parameters of the time encoder, where is the specific content of model training. Through the loss function designed later, the parameters are corrected to better obtain the hidden representation of the input spectrum fingerprint for the downstream device operating state monitoring task. The encoder of the present invention is a hierarchical Transformer encoder. First, the input is sequence-blocked. For convenience of representation, hereinafter is used to represent i.e., a certain enhanced view:

[0118]

[0119] Among them p is the block length, s is the sliding step ( s < p ), and after linear projection, is obtained:

[0120]

[0121] Among them is the learnable weight, is the positional encoding. The l layer of the hierarchical Transformer encoding is:

[0122]

[0123] Among them , the single-layer calculation details are as follows:

[0124]

[0125] After that, the hierarchical outputs are concatenated to obtain R :

[0126]

[0127] The hierarchical outputs are passed through max pooling to obtain the hidden representation r :

[0128]

[0129] (4) Intra-prototype contrast learning,

[0130] To achieve a uniform distribution of different spectrum fingerprint prototypes in the representation space, the present invention designs intra-prototype contrast learning with adaptive parameters. Through the value controls the penalty intensity of negative samples. Through the distance metric function Calculate the j rd and k th spectrum enhancement views and the distance between them , and then use the softmax function to map the distance to a value. The calculation formula is:

[0131]

[0132] Performing contrastive learning on multiple enhanced views within the prototype helps prevent views with significantly different amplitudes from dominating the prototype, thereby indirectly optimizing the aggregation process from views to the prototype. To output a low-dimensional representation suitable for contrastive learning, , generate the corresponding representation through non-linear projection:

[0133]

[0134] For , the in-prototype contrast loss is defined as:

[0135]

[0136] where and . Before calculating , set to negative infinity to ensure that positive sample pairs are close to each other.

[0137] (5) Inter-prototype contrastive learning,

[0138] The prototype calculation formula is:

[0139]

[0140] where , G represents the number of enhancement types, represents the non-linear projection function:

[0141]

[0142] where represents the input, represents the first-layer weight, represents the first-layer bias vector, represents the second-layer weight, represents the second-layer bias vector, represents the activation function.

[0143] For the i th sample, and are positive samples of each other, and they regard the prototypes of other samples in the batch as negative samples. The contrast loss between the prototypes of

[0144]

[0145] is defined as: B where

[0146] denotes the batch size.

[0147]

[0148] where α is a hyperparameter.

[0149] (6) Spectral fingerprint image modeling

[0150] It is difficult for numerical modeling-based spectral fingerprint mode enhancement to capture the structural information crucial for class recognition. Time series data in different fields are usually composed of line segments or curve segments. Therefore, compared with numerical values, it is more direct to capture the structural information of spectral fingerprints based on shape. To capture the structural features of spectral fingerprints, the present invention introduces the image modality and proposes sequence-image contrast learning using a hybrid strategy to simultaneously obtain numerical and structural information.

[0151] First, each time series sample is transformed into an image. Subsequently, the time series sample and its corresponding image form a positive sample pair, while the time series and images of other samples in the same batch are used as negative sample pairs. Although basic contrast learning can establish the correspondence between time series and images, since the negative samples do not consider the numerical characteristics of the time series, such differences are still insufficient to distinguish different time series samples. Therefore, we designed a geodesic hybrid method to generate hybrid representations located between the subspace representations of the two modalities. Using these hybrid representations as negative samples in contrast learning can expand the effective subspace of the learned representations and make the samples easier to classify.

[0152] Visualizing spectral fingerprint data through a line chart is a natural intuition for transforming numerical data into the image modality. In the line chart, x the y axis represents the timestamp, For different dimensions, since each variable has a different dimension, a line chart is plotted for each variable, and the images of different variables are standardized to the same square size. In addition, different variables are identified with different colors, and the sub-images of each variable are finally stitched into a complete image, denoted as , where H , W represents the standardized square size, and C represents the number of channels. Subsequently, the representation is obtained through an image encoder:

[0153]

[0154] where the structure of the image encoder is:

[0155]

[0156] where is the convolutional kernel weight, is the convolution operation, is the activation function, and is the convolutional bias vector.

[0157] Global pooling is represented as:

[0158]

[0159] (7) Image contrast loss function,

[0160] Extract the corresponding representation through the time series encoder . Due to the characteristics of different modalities, there is incomparable information between the time series and the image modality. For example, the contrast of an image has nothing to do with the inherent properties of time series data, and such information is not suitable for cross-modal contrast learning. Therefore, to filter out the incomparable information, the present invention performs a non-linear projection on the representations of each modality so that they can be compared during training. By screening the information suitable for sequence-image contrast learning, a more suitable time series representation and are obtained from and the image representation . The sequence-image contrast loss of the i th sample in the batch can be expressed as:

[0161]

[0162] where, represents the contrast loss between the i th image representation and all spectral sequence representations in the batch, while represents the contrast between the i th spectral sequence and the image representation, and the spectral sequence-image contrast loss function is defined as:

[0163]

[0164] (8)Loss function hybrid strategy

[0165] During the model training process, the hybrid loss function can comprehensively consider the balance relationship between different optimization objectives in the design of the hybrid loss function. The present invention adopts a hybrid loss strategy based on dynamic adaptive weights, which can automatically adjust the weight ratio according to the performance of different loss functions during the training process, thereby effectively improving the overall performance of the model:

[0166]

[0167] where is the overall loss function, is a dynamic weight coefficient that changes with the time step, and is defined as where The relative convergence degree of the loss is :

[0168]

[0169] indicates loss stagnation, indicates loss decline, indicates loss oscillation.

[0170] Dynamically adjust the loss weight coefficient based on the convergence degree:

[0171]

[0172] where is used to dynamically adjust the sensitivity, indicates the number of loss terms, which is 2 here, is , is .

[0173] (9)Device operation status classification

[0174] Based on the overall loss function , the parameters of the time series encoder are updated through backpropagation, and the parameter gradient for the update is:

[0175]

[0176] where represents the encoder parameters, and the symbol represents the partial derivative, which is used to describe the rate of change of a certain variable in a multivariable function.

[0177] Chain rule decomposition:

[0178]

[0179] where is the prototype representation, is the total number of samples, and the parameter update rule is:

[0180]

[0181] where is the learning rate, t is the number of training steps.

[0182] For the task of classifying the operating states of oil equipment, the present invention finally uses a time series encoder with optimized parameters to obtain a high-dimensional embedding as the input to the classifier, and finally obtains the equipment operating state:

[0183]

[0184] where, represents the input spectral fingerprint X the high-dimensional feature vector extracted by the feature encoding network r , C is the total number of state categories. Taking three types of key equipment, namely beam pumping units, centrifugal pumps, and drilling motors, as examples, their common operating state classifications are as follows.

[0185] Beam Pumping Unit, Normal state: The energy distribution in each frequency band is balanced, and the amplitude of the characteristic frequency is stable; Crankpin Wear: Sideband modulation phenomenon appears in the 630 - 850 Hz frequency band; Gear Tooth Breakage: High-order harmonics appear at the meshing frequency; Horsehead Misalignment: The energy at the 1 / 3 multiple frequency increases abnormally.

[0186] Centrifugal Pump, Normal state: The amplitude of the impeller passing frequency (BPF) conforms to the design threshold; Cavitation: Continuous broadband noise appears in the high-frequency band (>5 kHz); Bearing Inner Race Defect: Impact peak appears at the characteristic frequency (BPFI); Impeller Imbalance: The amplitude of the 1X rotation frequency exceeds the baseline by 200%.

[0187] Drilling Motor, Normal state: The amplitude of the electromagnetic characteristic frequency (such as slot harmonics) is stable; Broken Rotor Bar: Generates a characteristic frequency component of (1±2s)f_s; Stator Winding Short: The 3rd harmonic shows abnormal enhancement

[0188] Shaft Misalignment: Generates 2X rotation frequency and combined frequency components. The cosine similarity calculation in the formula effectively eliminates the deviation caused by the difference in characteristic amplitudes, making the classification result only depend on the direction similarity in the feature space.

[0189] Figure 2 in it is Prototype contrast learning loss function, is the intra-prototype contrast loss function, is the inter-prototype contrast loss function, is the overall loss function, is the image sequence contrast loss function, represents the contrast loss between the image representation and all spectral sequence representations in the batch, while represents the contrast loss between the spectral sequence and the image representation.

[0190] Experimental verification:

[0191] Table 1 Comparison of prediction effects

[0192]

[0193] As shown in Table 1, based on the fusion of dynamic spectrum fingerprint technology and lightweight architecture optimization, the proposed algorithm achieves a breakthrough in the accuracy of multi-device fault diagnosis while maintaining low memory occupancy. Compared with the RAM occupancy of up to 70.5 - 91.3 MB of the traditional Transformer model, this solution reduces the memory requirement to 18.7 - 38.5 MB (a decrease of 73.4% - 58.3%) through the feature view compression mechanism. At the same time, it achieves an accuracy of 0.9395 and an R2 determination coefficient of 0.9942 on pumping unit equipment, a 13.5% improvement compared to the sub-optimal Transformer model in performance. Especially in the scenario of drilling motors with limited computing resources, the proposed method realizes a significant 24.51% improvement in fault diagnosis accuracy relative to the baseline model at the cost of only a 15% increase in RAM, verifying the efficiency and robustness advantages of the algorithm in complex industrial environments.

[0194] Example 2

[0195] This example provides a system, including:

[0196] A data acquisition module, configured to:

[0197] A computer-readable storage medium storing a plurality of instructions adapted to be loaded and executed by a processor of a terminal device to perform the non-invasive device operating state monitoring method based on electromagnetic spectrum.

[0198] A terminal device, comprising a processor and a computer-readable storage medium, the processor being configured to implement each instruction; the computer-readable storage medium being configured to store a plurality of instructions adapted to be loaded and executed by the processor to perform the non-invasive device operating state monitoring method based on electromagnetic spectrum.

[0199] The above are all preferred embodiments of the present invention, and the protection scope of the present invention is not limited thereby. Therefore, all equivalent changes made according to the structure, shape, and principle of the present invention shall be covered by the protection scope of the present invention.

Claims

1. A non-invasive device operating state monitoring method based on the electromagnetic spectrum, characterized in that, Including: Dynamic multi-dimensional spectrum fingerprint extraction; Prototype generation based on the extracted multi-dimensional spectrum fingerprints; High-dimensional feature generation through a time series encoder; Intra-prototype contrast learning and inter-prototype contrast learning are respectively performed based on the generated prototypes; Construct a prototype contrast loss function based on contrast learning; Perform image modeling based on the extracted multi-dimensional spectrum fingerprints, and construct an image contrast loss function based on the image modeling; Perform function mixing based on a loss function mixing strategy to obtain an overall loss function; Optimize the parameters of the time series encoder based on the overall loss function; Discriminate the device state based on the high-dimensional features generated by the time series encoder; Output the discrimination result; The prototype generation based on the extracted multi-dimensional spectrum fingerprints includes randomly generating two different augmented views for each spectrum fingerprint using each augmentation method in a data augmentation library containing G types of augmentation types, obtaining two sets of augmented view sets; Then input the enhanced view into the temporal encoder to obtain its high-dimensional hidden representation , : ， wherein are learnable parameters of the time encoder, The prototype calculation formula is: ， Among them , G represents the number of enhancement types, represents a non-linear projection function.

2. The non-invasive device operation status monitoring method based on electromagnetic spectrum according to claim 1, wherein, The dynamic multi-dimensional spectrum fingerprint extraction includes a multi-band parallel sampling circuit of a radio frequency sensor to synchronously collect the original electromagnetic radiation signals in the frequency band of 0.1 MHz - 3 GHz at a sampling rate of 200 MHz , covering the fundamental wave, harmonic wave and modulation sideband characteristics of the device; among them, based on the adaptive dynamic range compression method, the dynamic range of the signal amplitude is detected in real time at the embedded terminal, and the dynamic gain G is calculated as follows: ， wherein is the gain adjustment amount, is the peak value of the signal amplitude within the current sampling window, and the preset target amplitude range is , automatically adjust the gain of the front-end amplifier to avoid data saturation caused by strong interference signals or loss of weak features.

3. The non-invasive device operation status monitoring method based on electromagnetic spectrum according to claim 2, wherein The dynamic multi-dimensional spectrum fingerprint extraction further includes, aiming at the problem of electromagnetic signal aliasing in industrial sites, pre-setting a two-stage anti-aliasing filter at the acquisition end. The first stage uses an analog-domain tunable band-pass filter to dynamically adjust the center frequency by real-time monitoring of the fundamental frequency characteristics of the target device dynamically calculate the quality factor Q value. The second stage performs time-frequency domain energy threshold segmentation on the aliasing residual components through wavelet packet transform in the digital domain to eliminate vibration coupling noise from the original electromagnetic signal, expressed as: ， Among them, J is the wavelet packet decomposition level, j is the current decomposition level, k is the band index, is the wavelet packet coefficient, is the time domain form of the bd6 wavelet basis function, is the output signal of the first - stage filter.

4. A method for monitoring the operating state of a non-invasive device based on electromagnetic spectrum according to claim 3, characterized in that, Including: The dynamic multi-dimensional spectrum fingerprint extraction further includes spatio-temporal alignment of multi-modal signals. Among them, the rising edge of the electromagnetic signal detected by the RF sensor is used as the global trigger source, and the acquisition windows of the infrared thermal imager and the microphone are synchronously started through the FPGA hardware trigger pin to achieve microsecond-level time alignment of cross-modal data. Taking the electromagnetic trigger timestamp as the reference, the time window of the multi-modal signal is intercepted: ， where is the temperature signal, is the acoustic vibration signal, is the time window width.

5. A method for monitoring the operating state of a non-invasive device based on electromagnetic spectrum according to claim 4, characterized in that, Including: The prototype-based in-prototype contrast learning and inter-prototype contrast learning are respectively performed, including achieving uniform distribution of different spectral fingerprint prototypes in the representation space through in-prototype contrast learning, where the penalty intensity of negative samples is controlled by a threshold, and the distance metric function is used to calculate the j th and k th spectral enhancement views and the distance therebetween, and then the softmax function is used to map the distance to a value, and the calculation formula is: 。 6. The non-invasive device operation status monitoring method based on electromagnetic spectrum according to claim 5, wherein Including: The construction of the prototype contrast loss function based on contrast learning includes intra-prototype contrast learning and inter-prototype contrast learning, using a variety of data augmentation means to construct a strongly generalized representation and improve the classification performance of downstream tasks. The overall prototype-based contrast loss is defined as: ， wherein α is a hyperparameter.

7. A method for monitoring the operating state of a non-invasive device based on the electromagnetic spectrum according to claim 6, characterized in that, Including: The image contrast loss function constructed based on image modeling includes obtaining representations through an image encoder: , and extracting corresponding representations through a time series encoder ; By screening information applicable to sequence-image contrast learning, more appropriate time series representations and were obtained from and image representations . The spectral sequence-image contrast loss function is expressed as: ， Among them, represents the contrast loss between the i th image representation and all spectral sequence representations in the batch, while represents the contrast between the i th spectral sequence and the image representation.

8. A method for monitoring the operating state of a non-invasive device based on electromagnetic spectrum according to claim 7, characterized in that, Including: The overall loss function is obtained by mixing functions based on a loss function mixing strategy, including adopting a hybrid loss strategy based on dynamic adaptive weights to adjust the weight ratio according to the performance of different loss functions during training to obtain the overall loss function , and updating the parameters of the overall loss function time series encoder . The updated parameter gradient for use is as follows: ， Among them represents the encoder parameter, and the symbol represents the partial derivative, which is used to describe the rate of change of a certain variable in a multivariable function.

9. A non-invasive device operating state monitoring system based on the electromagnetic spectrum, characterized in that, Including: A data acquisition module, configured to perform dynamic multi-dimensional spectrum fingerprint extraction; A prototype module, configured to perform prototype generation based on the extracted multi-dimensional spectrum fingerprints; perform intra-prototype contrast learning and inter-prototype contrast learning respectively based on the generated prototypes; Construct a prototype contrast loss function based on contrast learning; An image module, configured to perform image modeling based on the extracted multi-dimensional spectrum fingerprints, and construct an image contrast loss function based on the image modeling; A mixing module, configured to perform function mixing based on a loss function mixing strategy to obtain an overall loss function; A discrimination module, configured to discriminate the device state based on the overall loss function; output the discrimination result; The prototype generation based on the extracted multi-dimensional spectrum fingerprints includes randomly generating two different augmented views for each spectrum fingerprint using each augmentation method in a data augmentation library containing G types of augmentation types, obtaining two sets of augmented view sets; Then input the enhanced view into the temporal encoder to obtain its high-dimensional hidden representation , : ， where are the learnable parameters of the time encoder, The prototype calculation formula is as follows: ， Among them , G represents the number of enhancement types, represents a non-linear projection function.

Images