Signal-to-text generation type pre-training large model

Abstract

The invention relates to the technical field of predictive maintenance of mechanical equipment, in particular to a signal-to-text generation type pre-training large model. Comprising the following steps: constructing a signal unit, constructing a signal operator, constructing a random parameter, constructing a signal function, outputting text description of mathematical characteristics of the signal function, outputting signal data of the signal function, outputting text description of physical characteristics of the signal function, and establishing a'signal data-text description 'question and answer pair with a thinking chain. Making a pre-training data set with a thinking chain; training a model; finely tuning the model; testing the model; the signal data can be converted into the text information, an equipment operation and maintenance manager is helped to make an operation and maintenance decision, equipment faults are avoided, and the production efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of predictive maintenance of mechanical equipment, and particularly to a signal-to-text generative pre-trained large model. Background Art

[0002] As a key component of prognostics and health management, fault diagnosis plays a crucial role in ensuring the smooth operation and effective maintenance of complex systems. By detecting faults in the system and accurately identifying their types and severity, timely and precise fault diagnosis is essential for reducing downtime, lowering maintenance costs, and improving the safety and reliability of the system. As modern systems become increasingly complex and highly interconnected, the identification and resolution of faults become more difficult. Although traditional fault diagnosis methods, such as manual inspection, empirical analysis, and physical modeling, are effective in some cases, they are often time-consuming and error-prone. The advancement of machine learning, especially deep learning, has brought revolutionary changes to fault diagnosis, driving the shift towards data-driven methods. However, these machine learning-based models are highly dependent on high-quality features in terms of performance, and these features are often difficult to obtain and face challenges in processing large-scale, high-dimensional data, especially when this data exhibits strong non-linearity.

[0003] Large language models, such as OpenAI's GPT series, have demonstrated their advanced reasoning and generation capabilities in dealing with complex language tasks. Although these models are mainly designed for language processing, their underlying technologies and architectures also provide the possibility for processing and analyzing non-traditional text data. This makes them potentially applicable to the analysis of data beyond pure text data, including signal data from sensors.

[0004] Combining the requirements of fault diagnosis and the data processing capabilities of large language models, it is very necessary to develop a fault diagnosis system that combines these technologies. For this purpose, we propose a generative pre-trained large model based on signal-to-text, which is specifically designed to process signal data. By converting signal data into intuitive text descriptions, this model aims to provide a more understandable and operable interface, thereby assisting technicians in identifying and responding to system feedback more quickly. This method improves the efficiency and intuitiveness of fault diagnosis through improved user interaction and direct fault explanation. Summary of the Invention

[0005] The purpose of the present invention is to solve the deficiencies existing in the prior art, and to propose a signal-to-text generative pre-trained large model, which can assist experts in analyzing the operating state of equipment and formulating certain equipment maintenance plans, providing basic support for the operating state monitoring, prediction, and maintenance of mechanical equipment.

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

[0007] Design a signal-to-text generative pre-trained large model, including the following steps:

[0008] First step: Construct signal units. Signal units refer to the basic signal types that make up a signal function, including harmonic signals, impulse decay signals, wavelet signals, non-stationary random signals, and stationary random signals;

[0009] Second step: Construct signal operators. Signal operators refer to the ways of performing operations between signal units, including addition, subtraction, multiplication, division, convolution, power operation, and integration;

[0010] Third step: Construct random parameters. Random parameters refer to the variable parameters in signal units and signal operators. The variable parameters in signal units include: amplitude, frequency, and phase in harmonic signals; impact amplitude, frequency, and decay time constant in impulse decay signals; scale stretching parameter and time translation parameter in wavelet signals; The variable parameters in signal operators include: adjusting the number or proportion of addition, subtraction, multiplication, division, convolution, power operation, and integration operations in the signal operator, and the number or proportion coefficient of the signal operator;

[0011] Fourth step: Construct a signal function. A signal function refers to a signal mathematical expression composed of signal units, signal operators, and random parameters, which contains all the quantization parameters required to describe a signal;

[0012] Fifth step: Output a text description of the mathematical characteristics of the signal function. The text description refers to the textual representation of the mathematical characteristics of the signal data corresponding to the signal function, including quantitative and qualitative descriptions of the mathematical characteristics of the composite signal;

[0013] Sixth step: Output the signal data of the signal function. Signal data refers to the signal time domain, frequency domain, and time-frequency domain data obtained by signal sampling and processing of the signal function, including signal waveforms, signal spectra, and signal time-frequency spectra. During the data sampling process, the sampling frequency and the highest frequency of the signal need to satisfy the sampling theorem;

[0014] Seventh step: Output a text description of the physical characteristics of the signal function. Convert the physical characteristics of the equipment, including structure, material, process, and working conditions, into text descriptions; for equipment containing rolling bearings, the structural characteristics include the inner race diameter, outer race diameter, ball diameter, and number of balls of the rolling bearing; for equipment containing gear pairs, the structural parameters include the number of teeth of the gear and the number of gear stages; for equipment containing impellers, the structural parameters include the number of blades; the working condition characteristics include rotational frequency, pressure, and medium;

[0015] Step 8: Establish "signal data - text description" Q&A pairs with a chain of thought. The "signal data - text description" Q&A pairs refer to "questions", "chains of thought", and "answers" that contain signal data and text descriptions. A specific signal function contains more than one Q&A pair, which is used to interpret and describe the mathematical features contained in the signal data from multiple perspectives and at multiple levels;

[0016] Step 9: Create a pre-training dataset with a chain of thought. Creating a pre-training dataset means creating a dataset for training a signal-to-text generative large model, which contains text description Q&A pairs of signal data and the output "questions", "chains of thought", and "answers";

[0017] Step 10: Model training. Model training refers to using the pre-training dataset with a chain of thought to train a signal-to-text generative large model so that it can extract and generate a chain of thought from the signal data and transform specific features into text descriptions. During the training process, first, use word embedding techniques such as BERT to convert the words in the text into embedding vectors. Then, select a Transformer architecture with multiple self-attention layers and a feed-forward neural network to process the embedding vectors, and add positional encoding to each input word embedding to preserve the sequence information of the text data;

[0018] Step 11: Model fine-tuning and model testing. Fine-tuning the model means that after obtaining the pre-trained model, use the LoRA method to unfreeze and fine-tune the pre-trained model to improve its performance on the new signal dataset. Fine-tuning aims to make the model better adapt to specific signal categories and text generation requirements, as well as solve the problems of overfitting and underfitting. It includes adjusting some parameters of the model, mainly by changing the learning rate, number of training epochs, LoRA rank, and loss amount for retraining;

[0019] Step 12: Model deployment. Model deployment involves integrating the trained model into a server or cloud platform so that the model can process real-time signal data and automatically generate relevant text descriptions;

[0020] Step 13: Using the model means that after the model is deployed, actually apply the model to automatically convert signal data into text descriptions.

[0021] Preferably, in the first step, the harmonic signal is expressed as:

[0022]

[0023] where is the initial phase, A is the amplitude of the signal, f is the frequency, and t is the time;

[0024] The impulse decay signal is expressed as:

[0025]

[0026] where is the initial phase, A is the impact amplitude of the signal, f is the frequency, t is the time, and α is the decay time constant;

[0027] The wavelet signal is expressed as:

[0028]

[0029] where a is the scale expansion parameter, b is the time translation parameter, and ψ(t) is the wavelet prototype, called the basic wavelet;

[0030] A random signal refers to a time series whose value at each moment is a random variable. Although the value of a random signal cannot be determined a priori, these values generally follow a certain statistical law. In other words, a random signal can be described by the statistical characteristics of its probability distribution. According to whether the k-th moment is related to time, random signals can be divided into non-stationary random signals and stationary random signals. The n-th order stationary random signal is defined as:

[0031] μ(t 1 , …, t k ) = μ(t 1 + τ, …, t k + τ) (16)

[0032] where μ(t) represents the mean value of the stationary random signal.

[0033] Preferably, in the second step, the convolution formula of the signal unit f(x) and the signal unit h(x) is as follows:

[0034]

[0035] The power operation f(x) n of the signal unit f(x) is as follows:

[0036]

[0037] The integral μ(x) of the signal unit f(x) is as follows:

[0038]

[0039] Preferably, in the fourth step, the typical function formula of the signal function:

[0040] Multiple harmonic signals:

[0041]

[0042] where is the initial phase, A iA is the amplitude of the signal, f is the frequency, t is the time, and N is the total number of harmonics;

[0043] Multiple random simple harmonic signals:

[0044]

[0045] where is the initial phase, B i is the amplitude of the signal, f i is the frequency, t is the time, and N is the total number of harmonics;

[0046] Multiple multiple - frequency simple harmonic plus random simple harmonic signals:

[0047] y(t) = y 1 +y 2 (22)

[0048] Periodic impact signal:

[0049]

[0050] where A 0 is the amplitude, u(t) is the unit step function, f n is the natural frequency of the system damping, f d is the fault characteristic frequency, k belongs to integers, k 1 belongs to 0 or the fraction 1 / 2, for outer - ring and inner - ring faults k 1 = 0, for rolling - element faults k 1 = 1 / 2;

[0051] Amplitude - modulated and frequency - modulated signal:

[0052]

[0053] where ψ i , α i and θ i are phases, A i and B i are the amplitudes of the signal,

[0054] f m and f ch are the fault characteristic frequency and the meshing frequency, t is the time, and N is the highest order of amplitude modulation and frequency modulation.

[0055] Preferably, in the fifth step, first, qualitatively describe the signal name and composition characteristics according to the signal time-domain waveform, that is, signal units and signal operators; then quantitatively describe the time-domain characteristics such as the period, amplitude, and index of the signal time-domain waveform, and then quantitatively describe the fundamental frequency and its amplitude, frequency-domain index, whether there are harmonics, sidebands, carrier frequency, etc. of the signal; finally, combine the time-domain waveform and the frequency-domain characteristics to qualitatively describe the corresponding relationship between the time-domain waveform characteristics and the spectral characteristics in general.

[0056] A signal-to-text generation pre-trained large model proposed by the present invention has the beneficial effects that it can effectively perform fault diagnosis and predictive maintenance of industrial mechanical equipment, thereby reducing unexpected shutdowns and production costs, and can actively provide predictive maintenance solutions and decisions for factories.

[0057] Of course, it is not necessary for any product implementing the present invention to achieve all the above-mentioned advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0059] Figure 1 It is a schematic structural diagram of the flowchart of a signal-to-text generation pre-trained large model proposed by the present invention;

[0060] Figure 2 It is a schematic structural diagram of the main standard framework and steps of a signal-to-text generation pre-trained large model proposed by the present invention;

[0061] Figure 3 It is a schematic structural diagram of the main idea for establishing a question-and-answer pair dataset of "signal data - text description" with a chain of thought;

[0062] Figure 4 It is a schematic structural diagram of the time-domain waveform in the specific implementation manner;

[0063] Figure 5 It is a schematic structural diagram of the spectrogram in the specific implementation manner. SPECIFIC EMBODIMENTS

[0064] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0065] Refer to Figures 1-5 , a signal-to-text generative pre-trained large model, comprising the following steps:

[0066] The first step is to construct signal units. Signal units refer to the basic signal types that make up a signal function, including harmonic signals, impulse decay signals, wavelet signals, non-stationary random signals, and stationary random signals;

[0067] Among them, the harmonic signal is expressed as:

[0068]

[0069] Where is the initial phase, A is the amplitude of the signal, f is the frequency, t is the time;

[0070] The impulse decay signal is expressed as:

[0071]

[0072] Where is the initial phase, A is the impulse amplitude of the signal, f is the frequency, t is the time, and α is the decay time constant;

[0073] The wavelet signal is expressed as:

[0074]

[0075] Where a is the scale expansion parameter, b is the time translation parameter, and ψ(t) is the wavelet prototype, called the basic wavelet;

[0076] A random signal refers to a time series whose value at each moment is a random variable. Although the value of a random signal cannot be determined a priori, these values generally follow a certain statistical law. In other words, a random signal can be described by the statistical characteristics of its probability distribution. According to whether the k-th moment is related to time, random signals can be divided into non-stationary random signals and stationary random signals. The n-th order stationary random signal is defined as:

[0077] μ(t 1 , …, t k ) = μ(t 1 + τ, …, t k + τ) (28)

[0078] Where μ(t) represents the mean value of the stationary random signal.

[0079] The second step is to construct signal operators. Signal operators refer to the ways of performing operations between signal units, including addition, subtraction, multiplication, division, convolution, power operation, and integration;

[0080] Among them, the convolution formula of the signal unit f(x) and the signal unit h(x) is as follows:

[0081]

[0082] The power operation f(x) of the signal unit f(x) n is as follows:

[0083]

[0084] The integral μ(x) of the signal unit f(x) is as follows:

[0085]

[0086] Step 3: Construct random parameters. Random parameters refer to the variable parameters in signal units and signal operators. The variable parameters in signal units include: amplitude, frequency, and phase in simple harmonic signals; impact amplitude, frequency, and decay time constant in impact decay signals; scale stretching parameter and time translation parameter in wavelet signals. The variable parameters in signal operators include: adjusting the number or proportion in addition, subtraction, multiplication, division, convolution, power operation, and integral operations in signal operators, and the number or proportional coefficient of signal operators;

[0087] Step 4: Construct a signal function. A signal function refers to a signal mathematical expression composed of signal units, signal operators, and random parameters, which contains all the quantization parameters required to describe a signal. The following gives the function formulas of several typical signals:

[0088] Multiple frequency - doubled simple harmonic signals:

[0089]

[0090] where is the initial phase, A i is the amplitude of the signal, f is the frequency, t is the time, and N is the total number of harmonics;

[0091] Multiple random simple harmonic signals:

[0092]

[0093] where is the initial phase, B i is the amplitude of the signal, f i is the frequency, t is the time, and N is the total number of harmonics;

[0094] Multiple frequency - doubled simple harmonic plus random simple harmonic signals:

[0095] y(t) = y 1 +y 2 (34) Periodic impact signal:

[0096]

[0098] where A 0 is the amplitude, u(t) is the unit step function, f n is the natural frequency of the system damping, f d is the fault characteristic frequency, k belongs to integers, k 1 belongs to 0 or the fraction 1 / 2, for outer and inner race faults k 1 = 0, for rolling element faults k 1 = 1 / 2;

[0099] Amplitude - modulated and frequency - modulated signal:

[0100]

[0101] where ψ i , α i and θ i are phases, A i and B i are the amplitudes of the signal,

[0102] f m and f ch are the fault characteristic frequency and the meshing frequency, t is time, and N is the highest order of amplitude modulation and frequency modulation.

[0103] Step 5, Output the text description of the mathematical characteristics of the signal function. The text description refers to the textual representation of the mathematical characteristics of the signal data corresponding to the signal function, including the quantitative and qualitative descriptions of the mathematical characteristics of the composite signal; the specific steps and description content are as Figure 2 shown. Specifically, first, qualitatively describe the signal name and composition characteristics according to the signal time - domain waveform, that is, the signal unit and the signal operator; then quantitatively describe the time - domain characteristics such as the period, amplitude, and index of the signal time - domain waveform, and then quantitatively describe the frequency - domain characteristics such as the fundamental frequency and its amplitude, frequency - domain index, whether there are harmonics, sidebands, carrier frequencies, etc. of the signal; finally, combine the time - domain waveform and the frequency - domain characteristics to qualitatively describe the corresponding relationship between the time - domain waveform characteristics and the spectrum characteristics in general.

[0104] For simple harmonic signals, 5 qualitative text description templates are set as shown in the following table.

[0105]

[0106]

[0107]

[0108]

[0109] Judge which type of the above text descriptions the actual signal belongs to according to its characteristics.

[0110] Step 6: Output the signal data of the signal function. The signal data refers to the signal time domain, frequency domain, and time-frequency domain data obtained by signal sampling and processing of the signal function, including signal waveform, signal spectrum, and signal time-frequency spectrum. During the data sampling process, the sampling frequency and the highest frequency of the signal need to satisfy the sampling theorem, and the sampling time cannot be less than 3 periods of the minimum frequency;

[0111] Step 7: Output the text description of the physical characteristics of the signal function, and convert the physical characteristics of the equipment including structure, material, process, and working condition into text descriptions; for equipment containing rolling bearings, the structural characteristics include the inner raceway diameter, outer raceway diameter, ball diameter, and number of balls of the rolling bearing; for equipment containing gear pairs, the structural parameters include the number of teeth of the gear and the number of gear stages; for equipment containing impellers, the structural parameters include the number of blades; the working condition characteristics include rotational frequency, pressure, and medium. For the signals of the equipment or components to be diagnosed, it is necessary to know and provide their signal physical characteristics, and then input them into the large model. The physical characteristics include rotational frequency and the fault characteristic frequencies of components calculated from the design parameters of the components.

[0112] Step 8: Establish "signal data - text description" Q&A pairs with a chain of thought. The "signal data - text description" Q&A pairs refer to the "questions", "chains of thought", and "answers" that contain signal data and text descriptions. A specific signal function contains more than one Q&A pair, which is used to interpret and describe the mathematical characteristics contained in the signal data from multiple angles and multiple levels; describe the mathematical characteristics of the signal in the order from the time domain to the frequency domain. The time-domain mathematical characteristics mainly include time-domain statistical indicators (root mean square, mean, kurtosis, linear kurtosis, margin, minimum value, maximum value, peak-to-peak value, mean, skewness, root mean square amplitude, absolute average value, variance, waveform index, and peak value) and time-domain waveform characteristics (fundamental frequency period, amplitude modulation period, periodic impact); the frequency-domain mathematical characteristics mainly include frequency-domain statistical indicators (root mean square, kurtosis, linear kurtosis, center frequency, standard deviation, mean, frequency variance, frequency standard deviation, frequency-domain energy) and spectrum characteristics (amplitude of the 1 - N1 times frequency of the fundamental frequency, the N2 frequency values with the largest amplitudes, the center frequency and sideband interval of amplitude modulation, amplitude of the fractional frequency of the fundamental frequency, amplitude of the 1 - N3 times frequency of non-fundamental frequencies, and the N4 center frequency values with the largest amplitudes). At the same time, introduce the idea of the chain of thought, that is, the text description of the physical characteristics of the known signal and the reasoning steps for answering questions in detail. During the reasoning process, it is necessary to consider the physical characteristics and mathematical characteristics of the signal, as well as the logic and steps for analyzing and solving problems, so as to effectively improve the application of the large model in equipment fault diagnosis. The specific idea is as Figure 3 shown.

[0113] Figure 4 The following is the signal time-domain waveform diagram of an embodiment of the present invention, Figure 5 and the following is the spectrogram of the embodiment. Specifically, for the simple harmonic wave signal data of the embodiment, the following three standard issues are as follows:

[0114] What kind of signal is this? Please describe the mathematical characteristics in the time-domain waveform diagram of the signal and the mathematical characteristics in the spectrogram? (In this case, time-domain and frequency-domain statistical indicators are not considered)

[0115] Chain of thought: First, analyze and observe the periodicity, waveform and its symmetry of the time-domain waveform to determine whether it has periodicity, whether it is a simple harmonic wave or multi-harmonic wave signal or a random signal. Then, measure and calculate the period duration and amplitude of the signal in the time domain. Secondly, analyze and identify the fundamental frequency and the amplitudes of its harmonics through the Fourier transform results. Comprehensively explain or summarize the reasons and corresponding relationships for these phenomena based on the characteristics between the time domain and the frequency domain.

[0116] Answer: This is a multi-harmonic periodic signal, that is, a non-simple harmonic periodic signal. In the time-domain waveform diagram of this signal, the period of the signal is 0.01 seconds and the amplitude is 4.5 mm / sec. In the spectrogram of this signal, the frequency of the fundamental frequency (1st harmonic) is 100 Hz and the amplitude is 4.022 mm / sec; the frequency of the 2nd harmonic is 200 Hz and the amplitude is 0.689 mm / sec. The frequency of the 3rd harmonic is 300 Hz and the amplitude is 0.345 mm / sec. In the time-domain waveform diagram of this signal, the time-domain waveform is asymmetric up and down, that is, it is a non-simple harmonic wave, but it is periodic. Therefore, in the spectrogram, the 100 Hz and its 2nd and 3rd harmonics can be seen.

[0117] Step 9, Make a pre-training dataset with a chain of thought. Making a pre-training dataset means creating a dataset for training a signal-to-text generative large model, which contains signal data and text description Q&A pairs of "questions", "chains of thought" and "answers" for output;

[0118] For the aforementioned signal units, signal operators, and signal functions, use a signal generator to randomly change parameters to generate a pre-training dataset D{signal data (X x ), question 1 (X q1 ), chain of thought 1 (X c1 ), answer 1 (X a1 ); question 2 (X q2 ), chain of thought 2 (X c2 ), answer 2 (X a2 ); question 3 (X q3 ), chain of thought 3 (X c3 ), answer 3 (X a3 ); question 4 (X q4) Chain of Thought 4(X c1 ) Answer 4(X a4 ) ; … ; Question K(X qK ) Chain of Thought K(X cK ) Answer K(X aK )}. Then, clean the training dataset, mainly removing content unrelated to the questions and incorrect content where the data does not match the questions, chains of thought, and answers.

[0119] Step 10: Model Training. Model training refers to using a pre-trained dataset containing chains of thought to train a signal-to-text generative large model so that it can extract and generate chains of thought from signal data and transform specific features into text descriptions. During training, first, use word embedding techniques such as BERT to convert words in the text into embedding vectors. Then, select a Transformer architecture with multiple self-attention layers and feed-forward neural networks to process the embedding vectors, and add position encoding to each input word embedding to preserve the sequence information of the text data. To improve the model's processing ability and reduce overfitting, implement multi-head attention and regularization techniques such as Dropout and L2 regularization in the model.

[0120] The goal of the training setting is to minimize the difference between the model output (chain of thought and answer) and the actual data. For this purpose, use the cross-entropy loss function and combine label smoothing techniques to measure and optimize this difference, and adopt the Teacher Forcing strategy to guide the model to generate the correct chain of thought and answer. Optimize the parameters by using the AdamW optimizer, and adjust the model's parameters through multiple iterations to minimize the training error, thereby optimizing the model performance.

[0121] Step 11: Model Fine-tuning and Model Testing. After obtaining the pre-trained model, fine-tune the pre-trained model using the LoRA method to improve its performance on the new signal dataset. Fine-tuning aims to make the model better adapt to specific signal categories and text generation requirements, as well as solve the problems of overfitting and underfitting. It includes adjusting some parameters of the model, mainly by changing the learning rate, number of training epochs, LoRA rank, and loss amount for retraining;

[0122] Specifically, perform low-rank decomposition on the pre-trained weight matrix W o ∈R m×n :

[0123] W = W o + ΔW (37)

[0124] Where ΔW = BA, B and A are two trained parameter matrices, the dimension of B is m×r, the dimension of A is r×n, r = min(m,n) is the rank of the matrix, and W is the new weight matrix. Therefore, it is not necessary to train all parameters, and only the weights in A and B are updated by gradient. W o and the injection matrix ΔW are both multiplied by the input X, and the coordinates are merged to obtain a feature representation adapted to the new task. Therefore, this new forward process is calculated as:

[0125] F = WX = W o X + ΔWX = W o X + BAX (38)

[0126] where F is the output feature, X is the input feature, and W o X is the original output without LoRA. A is initialized with random Gaussian, and B is initialized with zero to ensure ΔW = 0 at the beginning. In addition, in practice, ΔWX is scaled by α / r to stabilize the learning of the injection parameters, where α is a hyperparameter. In addition, the probability of the target answer in the text can be expressed as:

[0127]

[0128] where, X x is the input signal, X c is all the thought chains, X a is all the answers, X instruct is the instruction pair (X qi , X ci , X ai ), X instruct,＜i is the context instruction up to the i-th step; X c,＜i is the target thought chain token sequence up to the i-th step; X a,＜i is the target answer token sequence up to the i-th step; the task-specific parameter increment Δθ = Δθ(Θ) is encoded by a smaller set of parameters Θ, |Θ| = |θ o |. The task of finding Δθ becomes the optimization of Θ:

[0129]

[0130] Then, use the real signal data to verify whether the accuracy of the model output result reaches 80% or more. If not, further fine-tuning is performed.

[0131] Step 12, Model Deployment. Model deployment involves integrating the trained model into a server or cloud platform so that the model can process real-time signal data and automatically generate relevant text descriptions;

[0132] Step 13, using the model means that after the model is deployed, the model is actually applied to automatically convert signal data into text descriptions.

[0133] In the description of this specification, the descriptions referring to terms such as "one embodiment", "example", "specific example", etc. mean that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0134] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and utilize the present invention well. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A large pre-trained model for signal-to-text generation, characterized in that: The following steps are involved: The first step is to construct signal units. Signal units refer to the basic signal types that constitute signal functions, including simple harmonic signals, impulse attenuation signals, wavelet signals, non-stationary random signals, and stationary random signals. The second step is to construct signal operators. Signal operators refer to the way of performing operations between signal units, including addition, subtraction, multiplication, division, convolution, power operation and integration. The third step is to construct random parameters. Random parameters refer to variable parameters in signal units and signal operators. Variable parameters in signal units include: amplitude, frequency and phase in simple harmonic signals; impact amplitude, frequency and decay time constant in impact attenuation signals; scale expansion parameters and time shift parameters in wavelet signals; variable parameters in signal operators include: adjusting the number or proportion of addition, subtraction, multiplication, division, convolution, power operation and integration operations in signal operators, and the number or proportion coefficient of signal operators; Step 4: Construct a signal function. A signal function is a mathematical expression of a signal consisting of a signal unit, a signal operator, and random parameters. It contains all the quantitative parameters required to describe a signal. Step 5: Output the text description of the mathematical characteristics of the signal function. The text description refers to the textual representation of the mathematical characteristics of the signal data corresponding to the signal function, including the quantitative and qualitative description of the mathematical characteristics of the composite signal. Step 6: Output the signal data of the signal function. Signal data refers to the signal time domain, frequency domain and time-frequency domain data obtained by sampling and processing the signal function, including signal waveform, signal spectrum and signal time-frequency spectrum. In the data sampling process, the sampling frequency and the highest frequency of the signal need to satisfy the sampling theorem; Step 7: Output the text description of the physical characteristics of the signal function, convert the physical characteristics of the equipment including structure, material, process and working condition into text description; for equipment with rolling bearings, the structural characteristics include the inner raceway diameter, outer raceway diameter, ball diameter and number of balls; for equipment with gear pairs, the structural parameters include the number of gear teeth and the number of gear stages; for equipment with impellers, the structural parameters include the number of blades; working condition characteristics include rotational frequency, pressure and medium; Step 8. Establish a "signal data-text description" question-answer pair with a thinking chain. The "signal data-text description" question-answer pair refers to the "question", "thinking chain" and "answer" containing signal data and text description. A specific signal function contains more than one question-answer pair, which is used to interpret and describe the mathematical features contained in the signal data from multiple angles and multiple levels; Step 9: Create a pre-training dataset with thought chains. Creating a pre-training dataset means creating a dataset for training a large signal-to-text generative model, which contains signal data and the output text description question-answer pairs of "questions", "thought chains" and "answers"; Step 10: Model training. Model training refers to using a pre-trained dataset containing thought chains to train a large signal-to-text generative model, enabling it to extract and generate thought chains from signal data and convert specific features into text descriptions. During the training process, first use word embedding technology such as BERT to convert words in the text into embedding vectors. Then, select the Transformer architecture with multiple self-attention layers and feedforward neural networks to process the embedding vectors, and add position encoding to each input word embedding to retain the sequence information of the text data. Step 11: Model fine-tuning and model testing. Fine-tuning the model means unfreezing and fine-tuning the pre-trained model using the LoRA method after obtaining the pre-trained model to improve its performance on the new signal dataset. Fine-tuning aims to make the model better adapt to specific signal categories and text generation requirements, as well as to solve the problems of overfitting and underfitting. It includes adjusting some parameters of the model, mainly changing the learning rate, training rounds, and LoRA rank and loss amount for retraining; Step 12: Model deployment. Model deployment involves integrating the trained model into a server or cloud platform so that the model can process real-time signal data and automatically generate relevant text descriptions. Step 13: Using the model means actually applying the model to automatically convert signal data into text descriptions after the model is deployed.

2. The signal-to-text generative pre-trained large model according to claim 1, characterized in that: In the first step, the simple harmonic signal is expressed as: in is the initial phase, A is the amplitude of the signal, f is the frequency, and t is the time; The shock attenuation signal is expressed as: in is the initial phase, A is the impulse amplitude of the signal, f is the frequency, t is the time, and α is the decay time constant; The wavelet signal is represented as: Where a is the scale expansion parameter, b is the time translation parameter, and ψ(t) is the wavelet prototype, called the basic wavelet; A random signal is a time series whose value at each moment is a random variable. Although the value of a random signal cannot be determined a priori, these values ​​generally obey certain statistical laws. In other words, a random signal can be described by the statistical characteristics of a probability distribution. According to whether the k-order moment is related to time, random signals can be divided into non-stationary random signals and stationary random signals. Defined as: μ(t1,…,t k )=μ(t1+τ,…,t k +τ) (4) Where μ(t) represents the mean of the stationary random signal.

3. The signal-to-text generative pre-trained large model according to claim 1, characterized in that: In the second step, the convolution formula of the signal unit f(x) and the signal unit h(x) is as follows: The power operation f(x) of the signal unit f(x) n as follows: The integral μ(x) of the signal unit f(x) is given by:

4. The signal-to-text generative pre-trained large model according to claim 1, characterized in that: In the fourth step, the typical function formula of the signal function is: Multiple frequency harmonic signals: in is the initial phase, A i is the amplitude of the signal, f is the frequency, t is the time, and N is the total number of harmonics; Multiple random simple harmonic signals: in is the initial phase, B i is the amplitude of the signal, f i is the frequency, t is the time, and N is the total number of harmonics; Multiple harmonics plus random harmonics signals: y(t)=y1+y2 (10) Periodic impulse signal: Where A0 is the amplitude, u(t) is the unit step function, and f n is the damped natural frequency of the system, f d is the fault characteristic frequency, k is an integer, k1 is 0 or a fraction 1 / 2, k1=0 for outer and inner ring faults, and k1=1 / 2 for rolling element faults; AM / FM signal: where ψ i , α i and θ i is the phase, A i and B i is the amplitude of the signal, and are the fault characteristic frequency and meshing frequency, t is the time, and N is the highest order of amplitude modulation and frequency modulation.

5. The signal-to-text generative pre-trained large model according to claim 1, characterized in that: In the fifth step, firstly, the signal name and composition characteristics, that is, the signal unit and signal operator, are qualitatively described according to the signal time domain waveform; then, the time domain characteristics such as the period, amplitude, and index of the signal time domain waveform are quantitatively described; then, the signal fundamental frequency and its amplitude, frequency domain index, whether there are harmonics, sidebands, carrier frequency and other frequency domain characteristics are quantitatively described; Finally, the correspondence between the time domain waveform characteristics and the frequency domain characteristics is qualitatively described by combining the time domain waveform and frequency domain characteristics.

Images