Intelligent voice-based intelligent fault diagnosis operation and maintenance system for digital twinning of overcharge equipment

Abstract

The invention discloses an intelligent voice-based overcharge equipment digital twinborn intelligent fault diagnosis operation and maintenance system, which comprises a voice query module, a first vector module, a diagnosis request module, a fault diagnosis module, a second query vector for defining a fault diagnosis request at a current time point as a pre-constructed fault prediction model, and a voice processing module, outputting a multi-task fault label of the overcharge equipment in rated time; the operation and maintenance strategy module is used for defining an operation and maintenance strategy of the overcharge equipment according to the multi-task fault tag; according to the method, the digital twinborn model is adopted to simulate the physical characteristics and the operation state of the equipment, so that the system can still generate high-quality training data under the condition that real data is insufficient or missing, and in combination with the time sequence prediction model, the historical data of the sensor and the operation data generated by digital twinborn simulation can be utilized in the training process, so that the training efficiency is improved. And the fault prediction capability of the model under a complex working condition is improved.

Description

Technical Field

[0001] The present invention relates to a fault diagnosis and maintenance system, and specifically to a supercharging equipment digital twin intelligent fault diagnosis and maintenance system based on intelligent voice. Background Art

[0002] With the popularization of supercharging equipment, intelligent operation and maintenance and accurate fault prediction have become key links in improving equipment reliability and operational efficiency. Traditional fault diagnosis methods mainly rely on real-time data collected by sensors, combined with historical operation and maintenance experience for manual or rule-driven analysis. The patent document with patent publication number CN116184293A discloses a fault diagnosis method and alarm system based on a digital twin lithium battery system. Through immersive XR technology, it is mapped into a simulation platform for simulation of various environments, and the corresponding data is collected to generate sub-reports. The real-time data and environmental information change data are mapped to predict and diagnose the upcoming faults based on the change data information, and the faults are fed back to the real system for prediction and alarm. This allows for adequate preparation in advance and can effectively eliminate unnecessary fault phenomena. However, it does not take into account the prediction of impending faults based on timing characteristics, resulting in a lack of sufficient data samples for the training model, affecting the prediction ability. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a digital twin intelligent fault diagnosis and operation and maintenance system for supercharging equipment based on intelligent voice, which solves the technical problems raised in the background technology through the combination of digital twin model and timing model.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: The supercharging equipment digital twin intelligent fault diagnosis and maintenance system based on intelligent voice includes: Voice query module, used to obtain the user's voice query request; A first vector module, used to obtain a first query vector of a pre-built digital twin model according to the voice query request; a diagnosis request module, configured to obtain a fault diagnosis request of the supercharging device at a current time point according to the first query vector; A fault diagnosis module, used to define a fault diagnosis request at a current time point as a second query vector of a pre-built fault prediction model, and output a multi-task fault label of the supercharging device within a rated time; The fault prediction model is formed by iteratively training the time series model in the fault sample training set of the supercharging device; the multi-task fault label includes: fault category and estimated fault time; The operation and maintenance strategy module is used to define the operation and maintenance strategy of the supercharging device according to the multi-task fault label.

[0005] In some of the embodiments, obtaining a first query vector of a pre-built digital twin model according to the voice query request includes: S2-1. Construct a text query instruction according to the voice query request; S2-2, constructing a text query vector according to the text query instruction; S2-3. Define the text query vector as the first query vector of the digital twin model.

[0006] In some of the embodiments, constructing a text query instruction according to the voice query request includes: S2-1-1: cutting the voice query request into a plurality of windowed audio frames; S2-1-2: extracting audio features from the audio frame to obtain a plurality of audio features; S2-1-3: Calculate the matching degree between the plurality of audio features and the phoneme template in the acoustic model, determine the best matching phoneme for each audio feature, and output the phoneme sequence corresponding to the plurality of audio features; S2-1-4: Decode the output phoneme sequence to generate a word sequence corresponding to the phoneme sequence; S2-1-5: Define the word sequence as the text query instruction.

[0007] In some of the embodiments, audio features are extracted from the audio frame to obtain several audio features, including: S2-1-2-1, performing short-time Fourier transform on the audio frame to generate a frequency spectrum representation of the audio frame; S2-1-2-2, performing Mel frequency filtering on the spectrum representation, mapping it to a Mel scale, and generating a Mel spectrum; S2-1-2-3, performing logarithmic transformation on the Mel spectrum to generate a high-sensitivity spectrum feature; S2-1-2-4, performing cepstrum transformation on the highly sensitive spectrum feature, and extracting the audio feature from the spectrum feature; In some of the embodiments, the pre-construction step of the digital twin model includes: A-1. Use ETPA modeling software to create a basic model of the supercharging equipment; A-2. Obtain the basic hardware parameters of the supercharging device; A-3. Deploy the basic hardware parameters of the supercharging device into the basic model to construct an initialization device model of the supercharging device; A-4. Obtain the real-time operating parameters of the supercharging equipment; A-5. Connect the real-time operating parameters to the initialized device model to build a digital twin model of the supercharging device.

[0008] In some embodiments, obtaining, according to the first query vector, a fault diagnosis request of the supercharging device at a current time point includes: S3-1. Inputting the first query vector into the pre-built digital twin model; S3-2. According to the first query vector, query the digital twin parameters of the supercharging device corresponding to the voice query request in the digital twin model; S3-3. Define the digital twin parameters corresponding to the voice query request as the fault diagnosis request of the supercharging device at the current time point.

[0009] In some embodiments, according to the first query vector, querying the digital twin parameters of the supercharging device corresponding to the voice query request in the digital twin model includes: S3-2-1. Obtain state vectors of several sub-devices in the digital twin model; S3-2-2, calculating the cosine similarity between the first query vector and each sub-device state vector, and generating a cosine similarity set; The calculation expression of the cosine similarity is: ; in, represents cosine similarity, A represents the first query vector, B represents the sub-device state vector, is the dot product between the first query vector and the sub-device state vector, and Respectively represent the modulus lengths of the first query vector and the sub-device state vector; S3-2-3. In the cosine similarity set, select the sub-device state vector corresponding to the maximum similarity as the digital twin parameter of the voice query request.

[0010] In some of the embodiments, it is characterized in that a fault diagnosis request at a current time point is defined as a second query vector of a pre-built fault prediction model, and a multi-task fault label of a supercharging device within a rated time is output, including: S4-1, inputting the second query vector into a pre-built fault prediction model; S4-2, based on the trained optimal model parameters, performing forward propagation on the second query vector, and passing it layer by layer to the output layer of the fault prediction model; S4-3. Map the second query vector to a multi-task fault label at the output layer of the fault prediction model.

[0011] In some of the embodiments, the pre-construction step of the fault prediction model includes: B1. Construct a fault sample training set; the input features of the fault sample training set are the digital twin parameters at the previous time point, and the target label is the multi-task fault label rated at the later time point; B2. Select the time series model as the initial training model; B3. Define the weighted loss function of the initial training model; The expression of the weighted loss function is: ; ; ; in, represents the fault classification loss, represents the real fault category in the fault sample training set, represents the probability of the fault category predicted by the timing model, represents the estimated loss during failure time, represents the actual fault occurrence time in the fault sample training set, represents the fault occurrence time predicted by the timing model, and N represents the total number of fault samples; represents the weight coefficient of fault classification loss, represents the weight coefficient of the estimated failure time loss, Represents the weighted loss of the iterative process of the time series model.

[0012] B4. Use the fault sample training set to iterate the initial training model until convergence to form the fault prediction model.

[0013] In some embodiments, the initial training model is iterated using the fault sample training set until convergence to form the fault prediction model, including: B4-1. Receive the initial batch of fault samples in the fault sample training set; B4-2. Perform forward propagation on the initial batch of fault samples and output the multi-task fault prediction labels of the initial batch; B4-3. Calculate the weighted loss between the initial batch of multi-task fault prediction labels and the actual multi-task fault labels in the fault sample training set according to the weighted loss function; B4-4. Calculate the gradient of weighted loss with respect to model parameters; B4-5. Use the gradient descent method to update the model parameters and generate updated model parameters; B4-6. Calculate the weighted loss of the next batch of fault samples based on the updated model parameters; B4-7. Loop through B4-4 to B4-6 until the weighted loss is minimized.

[0014] B4-8. Export the model when the weighted loss is minimized as the fault prediction model.

[0015] The present invention provides a supercharging equipment digital twin intelligent fault diagnosis and operation and maintenance system based on intelligent voice, which has the following beneficial effects: The present invention adopts a digital twin model to simulate the physical characteristics and operating status of the equipment, so that the system can still generate high-quality training data when the real data is insufficient or missing. It is combined with time series prediction models (such as LSTM, GRU). During the training process, both the sensor historical data and the operating data generated by digital twin simulation can be used to improve the model's fault prediction ability under complex working conditions.

[0016] Furthermore, through semantic matching and vector calculation, the user's voice query can be accurately mapped to the digital twin model, obtaining high-precision equipment operating parameters and improving the accuracy of fault diagnosis. Through the trained time series prediction model, the health status of the equipment in a specific time period in the future can be predicted, including the fault type, occurrence time, and impact range, so as to formulate the optimal operation and maintenance strategy in advance. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a structural block diagram of the digital twin intelligent fault diagnosis and operation and maintenance system of supercharging equipment based on intelligent voice of the present invention; Figure 2 It is a flow chart of the intelligent fault diagnosis and operation and maintenance system of the digital twin of supercharging equipment based on intelligent voice of the present invention; Figure 3 This is a modeling flow chart of the digital twin model described in the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] First, the prior art and related concepts involved in the embodiments of the present invention are described: ETAP: An analysis and simulation software widely used in the design, analysis and operation of power systems, supporting comprehensive modeling and simulation from power generation to transmission and distribution systems.

[0020] Example 1: Please refer to Figures 1 to 3 The present invention provides a supercharging equipment digital twin intelligent fault diagnosis and operation and maintenance system based on intelligent voice, including: Voice query module, used to obtain the user's voice query request; A first vector module, used to obtain a first query vector of a pre-built digital twin model according to the voice query request; a diagnosis request module, configured to obtain a fault diagnosis request of the supercharging device at a current time point according to the first query vector; A fault diagnosis module, used to define a fault diagnosis request at a current time point as a second query vector of a pre-built fault prediction model, and output a multi-task fault label of the supercharging device within a rated time; The fault prediction model is formed by iteratively training the time series model in the fault sample training set of the supercharging device; the multi-task fault label includes: fault category and estimated fault time; The operation and maintenance strategy module is used to define the operation and maintenance strategy of the supercharging device according to the multi-task fault label.

[0021] In this embodiment, intelligent fault diagnosis and early warning of supercharging equipment are achieved through the combination of intelligent voice interaction, digital twin modeling and time series prediction models. First, based on the user's voice request, the system parses and generates a query vector that matches the digital twin model to accurately map the current operating status of the equipment. Subsequently, the real-time data of the equipment is deeply analyzed through the fault prediction model, and the time series prediction capability trained with historical fault samples is combined to infer future fault types and occurrence times. Finally, the system defines the operation and maintenance strategy based on the prediction results; Among them, the definition of the operation and maintenance strategy can be a targeted selection from a pre-built operation and maintenance strategy library. For example, when the system detects that the battery management system (BMS) of a supercharging device is operating abnormally at the current time point, and infers through the fault prediction model that it may have a battery temperature overheating failure in the next 12 hours, in this case, the system will match the best operation and maintenance plan for battery overheating from the pre-built operation and maintenance strategy library based on the prediction results, such as reducing the charging power, increasing the cooling fan power, or notifying the operation and maintenance personnel to conduct on-site inspections, so as to intervene in advance and prevent serious damage to the equipment.

[0022] The step of obtaining the first query vector by the first vector module includes: S2-1. Construct a text query instruction according to the voice query request; S2-2, constructing a text query vector according to the text query instruction; In this embodiment, the text query instruction can be converted into a text query vector by using a word embedding text conversion method; the word embedding text conversion method is based on deep learning technology (such as Word2Vec, GloVe) to perform text conversion, and its vocabulary is converted into a low-dimensional vector, and the vector of each word captures the semantic relationship and similarity. For example, Word2Vec and GloVe will represent "king" and "emperor" as very similar vectors, reflecting their semantic relationship.

[0023] S2-3. Define the text query vector as the first query vector of the digital twin model.

[0024] In this embodiment, through the combination of speech analysis, text vectorization and semantic matching technology, accurate conversion from user natural language input to digital twin system query is achieved. The final generated text query vector can seamlessly connect to the digital twin model to match the query request with the device operating status.

[0025] Furthermore, in step S2-1, the specific steps of constructing the text query instruction include: S2-1-1: cutting the voice query request into a plurality of windowed audio frames; The processing of audio signals is usually based on short-term analysis, so the speech signal is divided into several smaller time periods (i.e., audio frames). Each frame is about 20-40 milliseconds, which makes it easy to capture the instantaneous characteristics of the audio signal. In this embodiment, the speech signal is cut into multiple audio frames of fixed length by a sliding window method, thereby focusing on short-term audio features and avoiding long-term dependence of the audio signal.

[0026] S2-1-2: extracting audio features from the audio frame to obtain a plurality of audio features; The audio signal of each audio frame is converted into a digital feature representation through audio feature extraction. Common audio feature extraction methods include Mel-frequency cepstral coefficients (MFCC).

[0027] S2-1-3: Calculate the matching degree between the plurality of audio features and the phoneme template in the acoustic model, determine the best matching phoneme for each audio feature, and output the phoneme sequence corresponding to the plurality of audio features; Specifically, the features of each audio frame are matched against a template in the acoustic model, which is usually based on a hidden Markov model (HMM), which can identify the phonemes (the smallest unit of speech) corresponding to the audio features.

[0028] In this embodiment, the acoustic model is trained using a large amount of annotated speech data to establish a mapping relationship between phonemes and audio features.

[0029] Then the matching degree between the audio features and the phoneme template in the acoustic model is calculated, and the phoneme (such as "b", "t", "a") that best matches each audio feature is selected.

[0030] In HMM, the matching degree of audio features is usually performed by calculating the forward algorithm. Each phoneme model can be regarded as a part of HMM, and the parameters of the HMM model include state transition probability, observation probability and initial state probability.

[0031] Assume that each phoneme has multiple states and the audio features is the observed signal. For each audio frame, the audio feature , the HMM model calculates its probability in each state.

[0032] The calculation expression of the matching degree is: ; in, Indicates the model parameters Under this condition, the audio characteristics are observed probability; is the forward probability, which indicates the probability that the model is in state s at time t; is the audio feature observed in state s The probability of, N is the number of states of the HMM model.

[0033] Finally, the acoustic model is used to identify the best matching phoneme corresponding to the audio feature, and the audio feature of each audio frame is matched to one or more phonemes.

[0034] S2-1-4: Decode the output phoneme sequence to generate a word sequence corresponding to the phoneme sequence; This step converts the phoneme sequence into an actual word sequence through the decoder. Specifically, the decoder decodes the phoneme sequence into words or phrases through a language model and a probabilistic algorithm, taking into account phonemes such as context and grammar to generate a reasonable word sequence. Among them, the language model adjusts the decoding results of the phoneme sequence according to word frequency and context information, excludes meaningless or inappropriate combinations, and ensures that the decoded text conforms to the actual context. The speech model can adopt the n-gram model; the n-gram model is one of the most common and basic language models, based on statistical learning methods, and predicts the next word by calculating the frequency and co-occurrence relationship between words.

[0035] That is, the phoneme sequence decoded by the HMM acoustic model will be used as the input of the decoder; The decoder calculates the most likely corresponding word for each phoneme based on the context of the phoneme sequence and the language model; The decoder uses a language model to constrain the decoding of the phoneme sequence, ensuring that the output word sequence is grammatically sound and fits the context.

[0036] S2-1-5: defining the word sequence as the text query instruction; The decoded word sequence is used as a structured text query instruction. The text query instruction will be used in subsequent digital twin model query, fault diagnosis and other steps, such as "check charging pile No. 5".

[0037] Furthermore, in step S2-1-2, the specific steps of extracting audio features include: S2-1-2-1, performing short-time Fourier transform on the audio frame to generate a frequency spectrum representation of the audio frame; Through short-time Fourier transform, the time domain signal of the audio frame can be converted into a frequency domain representation, so as to analyze the signal frequency components of the audio frame.

[0038] S2-1-2-2, performing Mel frequency filtering on the spectrum representation, mapping it to a Mel scale, and generating a Mel spectrum; The Mel scale is designed based on the frequency response characteristics of human hearing, and particularly enhances the sensitivity of the low-frequency part because the human ear is more sensitive to low-frequency sounds.

[0039] S2-1-2-3, performing logarithmic transformation on the Mel spectrum to generate a high-sensitivity spectrum feature; Through logarithmic transformation, the information in the Mel spectrum is compressed, making low-intensity signals (small signals) more sensitive, thereby improving the detection capability of low-intensity signals.

[0040] S2-1-2-4, performing cepstrum transformation on the highly sensitive spectrum feature, and extracting the audio feature from the spectrum feature; The cepstrum transform mainly removes the slow-changing components in the spectral features, making the fast-changing sound components more prominent and enhancing the short-term characteristics of speech. The features after cepstrum transform can be understood as the most representative features of the highly sensitive spectral features, which can effectively represent the key content of the audio frame.

[0041] In this embodiment, multi-level audio feature extraction and transformation processing ensures that the voice signal can be parsed and represented in the best way before being converted into a text query instruction. By extracting the main components of high-dimensional spectral features, removing redundant information and enhancing the short-term features of speech, audio features suitable for speech recognition and semantic analysis can be extracted.

[0042] Exemplarily, the pre-construction steps of the digital twin model described in this embodiment include: A-1. Use ETPA modeling software to create a basic model of the supercharging equipment; The basic model defines the main structure and functional components of the supercharging equipment, including batteries, charging piles, battery management system (BMS), current regulation module, etc. It is a preliminary model that does not contain real-time operation data. Its purpose is to provide the basic composition and electrical behavior framework of the equipment.

[0043] A-2. Obtain the basic hardware parameters of the supercharging device; Among them, the basic hardware parameters include basic parameters: battery type, capacity, voltage, current, power and other information. It also includes: the electrical network architecture of the supercharging equipment: battery management system, current regulation module, transformer, charging pile, load control and other modules.

[0044] A-3. Deploy the basic hardware parameters of the supercharging device into the basic model to construct an initialization device model of the supercharging device; A-4. Obtain the real-time operating parameters of the supercharging equipment; Collect the real-time operating parameters generated during the operation of the supercharging device in order to dynamically simulate the device. The real-time operating parameters include, but are not limited to: real-time battery voltage, current, power, device temperature, load status, operating mode and other information.

[0045] A-5. Connect the real-time operating parameters to the initialized device model to build a digital twin model of the supercharging device.

[0046] The real-time operating parameters are input into the initialized equipment model so that the model can reflect the current status of the equipment in real time. Through data access, the digital twin model can now not only reflect the static structure of the equipment, but also simulate the dynamic operating behavior of the equipment.

[0047] In this embodiment, the digital twin of the supercharging device is constructed throughout its life cycle by integrating ETPA modeling software with dynamic data. By dynamically accessing the initialized device model, the digital twin model not only has a static structure, but also can reflect the operating status of the supercharging device in real time.

[0048] The step of the diagnosis request module obtaining a fault diagnosis request of the supercharging device at the current time point includes: S3-1. Inputting the first query vector into the pre-built digital twin model; S3-2. According to the first query vector, query the digital twin parameters of the supercharging device corresponding to the voice query request in the digital twin model; S3-3. Define the digital twin parameters corresponding to the voice query request as the fault diagnosis request of the supercharging device at the current time point.

[0049] In this embodiment, the first query vector is used to query in conjunction with the digital twin model to achieve accurate mapping of voice requests to device operating status. By defining key operating parameters as fault diagnosis requests at the current time point, a complete status portrait of the device at a specific time is constructed.

[0050] Furthermore, in step S3-2, the specific steps of querying the digital twin parameters include: S3-2-1. Obtain state vectors of several sub-devices in the digital twin model; Specifically, each sub-device (such as battery, current regulation module, temperature sensor, etc.) has a corresponding vector representation, which represents the state of the sub-device at the current moment. The vector representation is generated in a similar way to word embedding, representing the semantics of the device parameters.

[0051] S3-2-2, calculating the cosine similarity between the first query vector and each sub-device state vector, and generating a cosine similarity set; The calculation expression of the cosine similarity is: ; in, represents the cosine similarity, A represents the first query vector, that is, the vector representation of the text query vector converted from the user's voice query request, B represents the sub-device state vector, for example, the vector representation of the battery voltage, current, temperature, etc., is the dot product between the first query vector and the sub-device state vector, indicating the similarity between the two. and Respectively represent the modulus of the first query vector and the sub-device state vector, that is, the size of the vector, which is obtained by summing the squares of each component of the vector (such as voltage, current, etc.) and then taking the square root. The components of the first query vector can be numerical representations of words or phrases, and the sub-device state vector can be component representations of the device in different dimensions, such as battery voltage, battery current, and battery temperature.

[0052] Specifically, the dot product of the first query vector and the sub-device state vector measures the projection product of the two vectors in the same direction. A large dot product value means that the directions of the two vectors are relatively consistent, that is, the semantics are similar; and the modulus of each vector indicates the size of the vector. Cosine similarity avoids the impact caused by the different sizes of the vectors by dividing the dot product by the modulus of the two vectors.

[0053] S3-2-3. In the cosine similarity set, select the sub-device state vector corresponding to the maximum similarity as the digital twin parameter of the voice query request.

[0054] In this embodiment, through semantic vector matching and cosine similarity calculation, the mapping of user query intention and the real-time status of the supercharging device is realized, and then the device parameters that best match the query request can be identified and defined as the current digital twin query result.

[0055] Exemplarily, the step of outputting a multi-task fault label by the fault prediction model includes: S4-1, inputting the second query vector into a pre-built fault prediction model; S4-2, based on the trained optimal model parameters, performing forward propagation on the second query vector, and passing it layer by layer to the output layer of the fault prediction model; S4-3. Map the second query vector to a multi-task fault label at the output layer of the fault prediction model.

[0056] In this embodiment, the query vector is forward propagated using the trained optimal model parameters to extract deep temporal features layer by layer, and a final multi-task fault label is generated at the output layer, including the fault category and the estimated fault occurrence time.

[0057] Embodiment 2: The technical solution of Embodiment 2 is different from that of Embodiment 1 in that a pre-construction step of the fault prediction model in Embodiment 1 is disclosed, and the pre-construction step includes: B1. Construct a fault sample training set. The input features of the fault sample training set are the digital twin parameters at the previous time point, that is, the operating state vector of the equipment (such as battery voltage, current, temperature and other time series data). The target label is the multi-task fault label rated at the later time. B2. Select the time series model as the initial training model; Among them, the timing model can be selected from time series prediction models such as long short-term memory network (LSTM), gated recurrent unit (GRU), etc. to adapt to the dynamic changes of the equipment operation status.

[0058] B3. Define the weighted loss function of the initial training model; The expression of the weighted loss function is: ; ; ; in, represents the fault classification loss, represents the real fault category in the fault sample training set, represents the probability of the fault category predicted by the timing model, represents the estimated loss during failure time, represents the actual fault occurrence time in the fault sample training set, represents the fault occurrence time predicted by the timing model, and N represents the total number of fault samples; represents the weight coefficient of fault classification loss, represents the weight coefficient of the estimated failure time loss, Represents the weighted loss of the iterative process of the time series model.

[0059] B4. Use the fault sample training set to iterate the initial training model until convergence to form the fault prediction model.

[0060] In this embodiment, by constructing a fault sample training set and adopting a time series deep learning model for training, and iterative training based on gradient descent optimization until the weighted loss function converges, a fault prediction model that can predict the fault type and fault occurrence time is formed, providing reliable data support for intelligent operation and maintenance.

[0061] The method of iterating the initial training model using the fault sample training set until convergence to form the fault prediction model includes: B4-1. Receive the initial batch of fault samples in the fault sample training set; B4-2. Perform forward propagation on the initial batch of fault samples and output the multi-task fault prediction labels of the initial batch; B4-3. Calculate the weighted loss between the initial batch of multi-task fault prediction labels and the actual multi-task fault labels in the fault sample training set according to the weighted loss function; B4-4. Calculate the gradient of weighted loss with respect to model parameters; B4-5. Use the gradient descent method to update the model parameters and generate updated model parameters; B4-6. Calculate the weighted loss of the next batch of fault samples based on the updated model parameters; B4-7. Loop through B4-4 to B4-6 until the weighted loss is minimized.

[0062] B4-8. Export the model when the weighted loss is minimized as the fault prediction model.

[0063] In this embodiment, efficient convergence and optimal performance of the supercharging equipment fault prediction model are achieved through batch gradient optimization training of the deep learning model. First, the fault sample training data is gradually received, and multi-task fault prediction labels are generated during the forward propagation process, including fault category and fault occurrence time. Subsequently, the prediction error is calculated based on the weighted loss function, and the model parameters are updated through gradient calculation and optimization to gradually adjust them to the optimal state. During the entire training process, iterative optimization is continuously performed, and the loss of the next batch of samples is cyclically calculated, and the model weights are continuously adjusted until the loss function converges, ensuring that the fault prediction model has high accuracy and generalization capabilities. Finally, the optimal fault prediction model is derived when the weighted loss is minimized, so that it can be used to predict potential faults in the operation and maintenance of supercharging equipment, providing decision support for the intelligent operation and maintenance of supercharging equipment.

[0064] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.) means.

[0065] The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., DVD ), or semiconductor media. The semiconductor media may be a solid state drive.

[0066] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not performed. Another point, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interface, indirect coupling or communication connection of the device or unit, which can be electrical, mechanical or other forms.

[0067] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. The digital twin intelligent fault diagnosis and operation and maintenance system for supercharging equipment based on intelligent voice is characterized by: include: Voice query module, used to obtain the user's voice query request; A first vector module, used to obtain a first query vector of a pre-built digital twin model according to the voice query request; a diagnosis request module, configured to obtain a fault diagnosis request of the supercharging device at a current time point according to the first query vector; A fault diagnosis module, used to define a fault diagnosis request at a current time point as a second query vector of a pre-built fault prediction model, and output a multi-task fault label of the supercharging device within a rated time; The fault prediction model is formed by iteratively training the time series model in the fault sample training set of the supercharging device; the multi-task fault label includes: fault category and estimated fault time; The operation and maintenance strategy module is used to define the operation and maintenance strategy of the supercharging device according to the multi-task fault label.

2. The intelligent voice-based supercharging equipment digital twin intelligent fault diagnosis and operation and maintenance system according to claim 1 is characterized in that: According to the voice query request, a first query vector of the pre-built digital twin model is obtained, including: S2-1. Construct a text query instruction according to the voice query request; S2-2, constructing a text query vector according to the text query instruction; S2-3. Define the text query vector as the first query vector of the digital twin model.

3. The intelligent voice-based supercharging equipment digital twin intelligent fault diagnosis and operation and maintenance system according to claim 2 is characterized in that: According to the voice query request, construct a text query instruction, including: S2-1-1, cutting the voice query request into a plurality of windowed audio frames; S2-1-2, extracting audio features from the audio frame to obtain a plurality of audio features; S2-1-3, calculating the matching degree between the plurality of audio features and the phoneme template in the acoustic model, determining the best matching phoneme for each audio feature, and outputting the phoneme sequence corresponding to the plurality of audio features; S2-1-4, decoding the output phoneme sequence to generate a word sequence corresponding to the phoneme sequence; S2-1-5. Define the word sequence as the text query instruction.

4. The intelligent voice-based digital twin intelligent fault diagnosis and operation and maintenance system for supercharging equipment according to claim 3 is characterized in that: Extracting audio features from the audio frame to obtain several audio features, including: S2-1-2-1, performing short-time Fourier transform on the audio frame to generate a frequency spectrum representation of the audio frame; S2-1-2-2, performing Mel frequency filtering on the spectrum representation, mapping it to a Mel scale, and generating a Mel spectrum; S2-1-2-3, performing logarithmic transformation on the Mel spectrum to generate a high-sensitivity spectrum feature; S2-1-2-4. Perform cepstrum transformation on the highly sensitive spectrum feature, and extract the audio feature from the spectrum feature.

5. The intelligent voice-based supercharging equipment digital twin intelligent fault diagnosis and operation and maintenance system according to claim 1 is characterized in that: The pre-construction steps of the digital twin model include: A-1. Use ETPA modeling software to create a basic model of the supercharging equipment; A-2. Obtain the basic hardware parameters of the supercharging device; A-3. Deploy the basic hardware parameters of the supercharging device into the basic model to construct an initialization device model of the supercharging device; A-4. Obtain the real-time operating parameters of the supercharging equipment; A-5. Connect the real-time operating parameters to the initialized device model to build a digital twin model of the supercharging device.

6. The intelligent voice-based digital twin intelligent fault diagnosis and operation and maintenance system for supercharging equipment according to claim 1 is characterized in that: Obtaining a fault diagnosis request of the supercharging device at a current time point according to the first query vector includes: S3-1. Inputting the first query vector into the pre-built digital twin model; S3-2. According to the first query vector, query the digital twin parameters of the supercharging device corresponding to the voice query request in the digital twin model; S3-3. Define the digital twin parameters corresponding to the voice query request as the fault diagnosis request of the supercharging device at the current time point.

7. The intelligent voice-based digital twin intelligent fault diagnosis and operation and maintenance system for supercharging equipment according to claim 6 is characterized in that: According to the first query vector, querying the digital twin parameters of the supercharging device corresponding to the voice query request in the digital twin model includes: S3-2-1. Obtain state vectors of several sub-devices in the digital twin model; S3-2-2, calculating the cosine similarity between the first query vector and each sub-device state vector, and generating a cosine similarity set; The calculation expression of the cosine similarity is: ; in, represents cosine similarity, A represents the first query vector, B represents the sub-device state vector, is the dot product between the first query vector and the sub-device state vector, and Respectively represent the modulus lengths of the first query vector and the sub-device state vector; S3-2-3. In the cosine similarity set, select the sub-device state vector corresponding to the maximum similarity as the digital twin parameter of the voice query request.

8. The intelligent voice-based supercharging equipment digital twin intelligent fault diagnosis and operation and maintenance system according to claim 1 is characterized in that: The fault diagnosis request at the current time point is defined as the second query vector of the pre-built fault prediction model, and the multi-task fault labels of the supercharging equipment within the rated time are output, including: S4-1, inputting the second query vector into a pre-built fault prediction model; S4-2, based on the trained optimal model parameters, performing forward propagation on the second query vector, and passing it layer by layer to the output layer of the fault prediction model; S4-3. Map the second query vector to a multi-task fault label at the output layer of the fault prediction model.

9. The intelligent voice-based supercharging equipment digital twin intelligent fault diagnosis and operation and maintenance system according to claim 1 is characterized in that: The pre-construction steps of the fault prediction model include: B1. Construct a fault sample training set; the input features of the fault sample training set are the digital twin parameters at the previous time point, and the target label is the multi-task fault label rated at the later time point; B2. Select the time series model as the initial training model; B3. Define the weighted loss function of the initial training model; The expression of the weighted loss function is: ； ； ； in, represents the fault classification loss, represents the real fault category in the fault sample training set, represents the probability of the fault category predicted by the timing model, represents the estimated loss during failure time, represents the actual fault occurrence time in the fault sample training set, represents the fault occurrence time predicted by the timing model, and N represents the total number of fault samples; represents the weight coefficient of fault classification loss, represents the weight coefficient of the estimated failure time loss, Represents the weighted loss of the iterative process of the time series model; B4. Use the fault sample training set to iterate the initial training model until convergence to form the fault prediction model.

10. The intelligent voice-based digital twin intelligent fault diagnosis and operation and maintenance system for supercharging equipment according to claim 9 is characterized in that: Iterating the initial training model using the fault sample training set until convergence to form the fault prediction model, including: B4-1. Receive the initial batch of fault samples in the fault sample training set; B4-2. Perform forward propagation on the initial batch of fault samples and output the multi-task fault prediction labels of the initial batch; B4-3. Calculate the weighted loss between the initial batch of multi-task fault prediction labels and the actual multi-task fault labels in the fault sample training set according to the weighted loss function; B4-4. Calculate the gradient of weighted loss with respect to model parameters; B4-5. Use the gradient descent method to update the model parameters and generate updated model parameters; B4-6. Calculate the weighted loss of the next batch of fault samples based on the updated model parameters; B4-7, loop through B4-4 to B4-6 until the weighted loss is minimized; B4-8. Export the model when the weighted loss is minimized as the fault prediction model.

Images