An industrial internet-based device state intelligent prediction and maintenance method

Abstract

The application discloses a kind of equipment state intelligent prediction and maintenance method based on industrial internet, including the operating data of target equipment and power information as activity information are collected by multimodal sensor array;Based on the activity information, the blind source separation of activity information guidance is carried out to sound signal, and the sound component of target equipment is extracted;The noise reduction and feature enhancement are carried out to the separated sound component, and the anti-noise acoustics feature is extracted;Based on anti-noise acoustics feature, acoustic anomaly score is calculated, and vibration, temperature, current harmonic anomaly score is calculated synchronously, and parallel auxiliary judgment is activated when acoustic anomaly score falls into fuzzy interval;According to real-time signal-to-noise ratio, the weight of each modal feature is dynamically adjusted, and fusion anomaly score is generated;Based on confidence score, perform grading response strategy;Degree of health is evaluated, and maintenance suggestion is generated, and the application realizes high reliability state prediction and maintenance of power adapter, converter and receiver and other equipment in strong noise environment.

Description

Technical Field

[0001] This invention relates to the field of industrial internet equipment condition monitoring technology, specifically to a method for intelligent prediction and maintenance of equipment condition based on industrial internet. Background Technology

[0002] With the development of the Industrial Internet, converters and receivers, as core components for interconnecting and transmitting signals among industrial field devices, directly impact the stability of the entire system due to their operational reliability. To enhance the state awareness of these devices, existing technologies have begun to integrate multi-dimensional sensor information. For example, in the field of smart adapters, some products have integrated multi-dimensional sensing modules that can collect key parameters such as port voltage, current, and temperature in real time. The aim is to build a unified diagnostic framework by simultaneously collecting multi-dimensional sensor data such as vibration, temperature, and current to achieve higher-precision equipment fault detection.

[0003] Industrial plants contain numerous interference sources, such as broadband electromagnetic noise and structural environmental noise generated by frequency converters, large motors, and welding equipment, making acoustic monitoring equipment susceptible to interference from other sound sources and complex acoustic environments. More challenging is the fact that when multiple similar devices operate simultaneously in the same area, the interference sounds generated by neighboring devices are highly similar in timbre to the target device's sound. This makes it difficult for traditional blind source separation algorithms to accurately extract the target sound component from the mixed signal due to a lack of prior information. Furthermore, most existing multimodal fusion methods employ fixed weights or simple splicing strategies, lacking an adaptive mechanism to dynamically adjust the weights of each mode based on real-time noise intensity. Once the acoustic channel is severely contaminated, the reliability of the overall diagnostic framework drops sharply. Therefore, this paper proposes an intelligent equipment status prediction and maintenance method based on the Industrial Internet to address these issues. Summary of the Invention

[0004] The purpose of this invention is to provide a method for intelligent prediction and maintenance of equipment status based on the Industrial Internet, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for intelligent prediction and maintenance of equipment status based on the Industrial Internet includes the following steps:

[0007] Step 1: Collect the operating data of the target device by deploying a multimodal sensor array at the target device. The multimodal sensor array includes a sound sensor, a vibration sensor, a temperature sensor, and a current sensor, and simultaneously collects the power information of the target device as activity information.

[0008] Step 2: Based on the activity information, perform blind source separation guided by activity information on the collected sound signals to separate the target device sound component and the environmental noise component from the mixed sound signal;

[0009] Step 3: Perform noise reduction and feature enhancement processing on the separated target device sound components, and extract noise immunity features;

[0010] Step 4: Calculate the acoustic anomaly score based on the noise immunity characteristics, and simultaneously calculate the vibration anomaly score, temperature anomaly score, and current harmonic anomaly score based on the data collected by the vibration sensor, temperature sensor, and current sensor, respectively. The temperature anomaly score and current harmonic anomaly score are obtained by comparing them with their respective preset thresholds.

[0011] Step 5: Estimate the signal-to-noise ratio of the current environment in real time, and dynamically adjust the weight of each modal feature in the fusion decision based on the signal-to-noise ratio to generate a fusion anomaly score;

[0012] Step 6: Calculate the confidence score of the current prediction based on the fusion anomaly score, and execute the hierarchical response strategy according to the confidence score;

[0013] Step 7: Based on the output of the graded response strategy, perform a health assessment and predict the remaining service life of the target equipment, and generate maintenance recommendations.

[0014] Preferably, the blind source separation guided by activity information in step two specifically includes: determining the current operating state of the target device through the power information, wherein the operating state includes standby state, light load state, and full load state; constructing a state-voiceprint mapping dictionary, using the current operating state as a priori label, and embedding state condition constraints in the blind source separation network; utilizing the periodicity of the target device's power frequency harmonics, dividing the mixed sound signal into multiple units periodically, and independently calculating an adaptive separation threshold for each unit; and extracting the target device sound component matching the current operating state from the mixed sound signal according to the state condition constraints and the adaptive separation threshold.

[0015] Specifically, let the mixed audio signal be... This includes the target device signal. With environmental noise ,but ;

[0016] Let the power frequency of the target device be... Its doubling period is , For harmonic order;

[0017] Periodically Divided into There are units, each unit having a length of . :

[0018]

[0019] For the Each unit has an adaptive separation threshold. for:

[0020] in, This is the mean value of the signal in this unit. The standard deviation of the signal in this unit is... The preset threshold coefficient is usually taken as... ;

[0021] With event information As a state condition constraint, conditional probability is introduced into the separable network. Extract the target device sound components that match the current operating status. .

[0022] Preferably, the noise reduction and feature enhancement processing in step three specifically includes:

[0023] Step 3.1: Perform deep noise reduction processing on the separated target device sound components using a stacked noise reduction autoencoder. The stacked noise reduction autoencoder is pre-trained with clean device sound samples with added Gaussian noise and mask noise.

[0024] Step 3.2: Perform multi-scale time-frequency joint analysis on the denoised audio signal, extract multi-granular features in the time domain through a multi-scale deep residual network, and analyze the frequency domain features in the Mel speech spectrogram through a multi-scale attention interleaved network;

[0025] Step 3.3: Extract frequency features using dynamic frequency bending technology, and adaptively adjust the feature extraction granularity according to the signal noise level.

[0026] Preferably, the acoustic anomaly score in step four is calculated in the following manner:

[0027] Step 4.1: Train the autoencoder model and Gaussian mixture model using the noise immunity characteristics of the target device under normal operating conditions;

[0028] Step 4.2: Input the currently acquired noise resistance features into the trained autoencoder model and Gaussian mixture model respectively, and calculate the reconstruction error and log-likelihood probability;

[0029] Step 4.3: Combine the reconstruction error and the log-likelihood probability to generate a normalized acoustic anomaly score. .

[0030] Preferably, the vibration anomaly score, temperature anomaly score, and current harmonic anomaly score are activated as parallel auxiliary judgments when the acoustic anomaly score falls within a preset fuzzy range; the preset fuzzy range ,in This is the normal threshold. To clarify the abnormal threshold;

[0031] The vibration anomaly score is calculated by extracting time-domain and frequency-domain features from the triaxial vibration signals collected by the vibration sensor and then using a vibration autoencoder model. The time-domain features include peak value, root mean square, kurtosis, and peak-to-peak value, and the frequency-domain features include the dominant frequency energy distribution and harmonic components.

[0032] The temperature anomaly score is obtained by extracting the absolute temperature value, temperature change rate, and temperature fluctuation variance from the temperature signal collected by the temperature sensor and comparing it with a preset temperature threshold.

[0033] The current harmonic anomaly score is obtained by performing a fast Fourier transform on the input current signal collected by the current sensor, extracting the total harmonic distortion rate and the amplitude ratio of each harmonic, and then comparing it with a preset harmonic baseline value.

[0034] Preferably, step five, which involves dynamically adjusting the weights of each modal feature based on the signal-to-noise ratio, specifically includes:

[0035] When the signal-to-noise ratio is higher than the first preset threshold, acoustic modal features have the highest weight in the fusion decision.

[0036] When the signal-to-noise ratio is lower than the second preset threshold, the weight of acoustic modal features is reduced, while the weight of current harmonic mode features and temperature mode features is increased.

[0037] The weights of each modal feature are continuously adjusted as the signal-to-noise ratio changes.

[0038] Preferably, the factors for calculating the confidence score in step six include: the current signal-to-noise ratio, the blind source separation quality assessment result, and the degree of consistency between the anomaly scores of each modality;

[0039] The hierarchical response strategy includes:

[0040] When the confidence score is higher than the first confidence threshold, the fusion anomaly score is directly output as the prediction result;

[0041] When the confidence score is between the first confidence threshold and the second confidence threshold, a multi-model integration voting mechanism is activated. The multi-model includes a first prediction model based on the original signal, a second prediction model based on the signal after blind source separation, a third prediction model based on the signal after noise reduction, and a fourth prediction model based on non-audio modes. The real-time confidence scores of each model are used as weights for weighted voting.

[0042] When the confidence score is between the second confidence threshold and the third confidence threshold, the acoustic modal channel is suspended, and downgraded prediction is performed only based on the vibration mode, temperature mode and current harmonic mode, and manual review is triggered.

[0043] When the confidence score is lower than the third confidence threshold, automatic prediction is paused, the original collected data is recorded, and on-site inspection is triggered.

[0044] Preferably, the health assessment and remaining lifespan prediction in step seven specifically include:

[0045] Step 7.1: Establish a multimodal anomaly scoring time-series database for the entire lifecycle of the target device;

[0046] Step 7.2: Model the time series database using a long short-term memory network, and output the health index and the remaining lifespan prediction range;

[0047] Step 7.3: Based on the combination pattern of the acoustic anomaly score, vibration anomaly score, temperature anomaly score, and current harmonic anomaly score, infer the root cause type of the fault;

[0048] Step 7.4: Generate corresponding maintenance recommendations based on the described root cause type of the fault;

[0049] Preferably, the inference rule for the root cause type of the fault includes:

[0050] When both the acoustic anomaly score and the vibration anomaly score are higher than their respective thresholds, the root cause of the fault is inferred to be loose transformer core or inductive resonance.

[0051] When both the acoustic anomaly score and the current harmonic anomaly score are higher than their respective thresholds, the root cause of the fault is inferred to be aging of the switching transistor or an open circuit of the rectifier diode.

[0052] When both the acoustic anomaly score and the temperature change rate anomaly score exceed their respective thresholds, the root cause of the fault is inferred to be an increase in the equivalent series resistance of the electrolytic capacitor or an overload of the equipment.

[0053] Preferably, the method further includes step eight: using the manual review results and the actual fault type, repair actions, and post-repair voiceprint changes as labeled samples for back-injection training dataset, and incrementally learning and updating each model in the method.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] 1. In this invention, blind source separation guided by activity information is used, and the real-time working status of the equipment is embedded in the separation network as a priori label. The periodicity of power frequency harmonics and adaptive thresholds are used to accurately extract the target sound components. The separated sound components are deeply purified by stacked noise reduction autoencoders and multi-scale time-frequency joint analysis. Acoustic anomaly scoring and multi-modal parallel auxiliary judgment of vibration, temperature and current harmonics are used. When acoustic features are ambiguous, independent criteria of other physical dimensions are introduced for cross-validation, which effectively eliminates misjudgments caused by environmental noise.

[0056] 2. In this invention, dynamic modal weight adjustment driven by signal-to-noise ratio adaptively reduces acoustic modal weights and tilts towards noise-resistant modes such as current harmonics and temperature based on real-time signal-to-noise ratio, ensuring reliable fusion anomaly scores can still be obtained under strong noise coverage; a confidence-driven four-level graded response strategy is used to perform multi-model integration voting, downgraded prediction, or trigger manual review when prediction is unreliable, avoiding unplanned downtime losses due to false alarms; and LSTM time-series modeling and elastic weight consolidation incremental learning mechanism are used to realize health index calculation, remaining service life range prediction, and continuous model optimization, ensuring highly reliable predictive maintenance of target equipment even in extreme noise environments. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the process flow of the present invention;

[0058] Figure 2 This is a schematic diagram of the determination process structure of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Example 1

[0061] Please see Figure 1 In this embodiment, an AC-DC power adapter with an input of 220V / 50Hz, an output of 24V / 5A, and a rated power of 120W, located in an industrial plant, is used as the target device. The hardware deployment scheme of this embodiment is as follows:

[0062] A multimodal sensor array is deployed at the target device, including: two high-sensitivity MEMS microphones with a signal-to-noise ratio >65dB and a frequency response range of 20Hz-20kHz. One microphone serves as a near-end microphone, deployed 10-20cm from the target device to collect mixed sound signals; the other serves as a far-end microphone, deployed 1-2 meters from the target device to collect ambient reference noise signals. A triaxial accelerometer with a range of ±16g and a bandwidth ≥5kHz is magnetically attached to the surface of the target device's casing to collect triaxial vibration signals. A patch-type NTC temperature sensor with an accuracy of ±0.5℃ and a response time <3s is fixed to the hot spot area (near the internal transformer) of the target device's casing with thermally conductive adhesive to collect temperature signals. A single openable current transformer (accuracy class 0.5) is non-invasively attached to the target device's input power line to collect input current signals. Simultaneously, a power monitoring module synchronously collects the real-time power information of the target device as activity information.

[0063] The acquisition parameters of each sensor are set as follows: the sampling rate of the sound sensor is 44.1kHz, the sampling rate of the vibration sensor is 10kHz, the sampling rate of the current sensor is 10kHz, the sampling rate of the temperature sensor is 1Hz, and the sampling rate of the power monitoring module is 100Hz. The acquisition data of all sensors are synchronized by a clock and then transmitted to the edge AI computing box via industrial Ethernet. The edge AI computing box has a pre-set data preprocessing program that performs preliminary processing such as filtering and DC bias removal on the raw acquisition data before storing it in the local time series database.

[0064] Example 2

[0065] Please see Figures 1 to 2 The specific implementation method of blind source separation guided by activity information in step two is described in detail in this embodiment.

[0066] The acquired mixed sound signal includes the target device sound component and the ambient noise component. The ambient noise comes from broadband noise generated by nearby inverters, motors and other equipment in the factory. Since the target device is an AC-DC power adapter, its internal switching transistors typically operate in the range of 50kHz-200kHz. However, the generated sound signal is mainly concentrated in the power frequency harmonics and its modulation sidebands, exhibiting obvious periodicity and regularity.

[0067] First, the real-time power information of the target device is obtained through the power monitoring module. Based on the power value, the current working state is divided into standby state (power < 5W), light load state (5W ≤ power < 60W), and full load state (power ≥ 60W). In this embodiment, when the power monitoring module measures the current power to be 95W, it is determined that the target device is in a full load state.

[0068] With event information As a priori labels, state condition constraints are embedded in the blind source separation network. Prior to this, a state-speaker mapping dictionary has been constructed through offline acquisition: pure sound samples of the target device in standby, light load, and full load states are collected respectively, and their respective frequency domain features are extracted as reference templates.

[0069] In practice, the separation should be carried out according to the following steps:

[0070] First, utilizing the power frequency of the target device. Take its frequency doubling period The 10th harmonic cycle is used as the segmentation unit length, and the mixed sound signal is divided into multiple units in 2ms increments.

[0071] Second, for each unit, calculate its mean and standard deviation according to the formula. Calculate the separation threshold of this unit, where In this embodiment, we use 2.0.

[0072] Third, using the voiceprint template under full load as a condition constraint, frequency components that match the full load state are extracted first.

[0073] Fourth, the target device audio components separated from each unit are spliced ​​together to obtain the complete separated target device audio signal. .

[0074] In this embodiment, with an ambient noise level of approximately 85 dB in the industrial environment, the signal-to-noise ratio of the target device's sound signal is improved by approximately 12 dB compared to the directly acquired mixed signal after adopting the activity information-guided blind source separation method described above, providing high-quality input for subsequent feature extraction.

[0075] Example 3

[0076] Please see Figures 1-2 The specific implementation of the noise reduction and feature enhancement processing described in step three is explained in detail in this embodiment.

[0077] First, deep noise reduction processing is performed on the separated target device sound components. This embodiment uses a three-layer stacked noise reduction autoencoder with the following network structure: the input layer has a dimension of 512, corresponding to 512-dimensional MFCC features; the first hidden layer has a dimension of 256; the second hidden layer has a dimension of 128; and the third hidden layer has a dimension of 64. The decoder and autoencoder have a symmetrical structure. This autoencoder has been pre-trained in a laboratory environment with clean device sound samples. During training, Gaussian noise and random mask noise at signal-to-noise ratios of 5dB, 0dB, and -5dB are added to the clean samples to enhance the model's generalization ability.

[0078] During inference, the separated audio signals are segmented into frames to extract MFCC feature vectors. Input the data into the SDAE network to obtain the reconstructed output. Calculate the reconstruction error When the reconstruction error is greater than the preset threshold, it indicates that the frame signal is heavily polluted by noise. The reconstruction output of the autoencoder network is used to replace the original input to achieve deep noise reduction.

[0079] After denoising, a multi-scale time-frequency joint analysis is performed on the denoised audio signal. Specifically, the denoised signal is divided into frames, and a short-time Fourier transform is performed on each frame after applying a Hanning window to obtain the spectrogram. The spectrogram is mapped to a Mel spectrogram using 128 Mel triangular filters. The dimensions are 128×128.

[0080] The Melan language spectrogram is simultaneously input into two branches: the temporal branch employs a multi-scale deep residual network with convolutional kernel scales of 3×3, 5×5, and 7×7 to extract temporal features at different time granularities; the frequency domain branch employs a multi-scale attention-interleaved network, where the attention mechanism weights along the frequency dimension to enhance the contribution of fault-related frequency bands. The output features of the two branches are fused through a concatenation operation to obtain the final noise-resistant feature vector. .

[0081] Example 4

[0082] Please see Figures 1-2 The specific implementation method of acoustic anomaly scoring calculation and parallel auxiliary judgment described in step four is explained in detail in this embodiment.

[0083] First, using 1000 sets of noise immunity feature samples collected under normal operating conditions of the target device as the training set, an autoencoder model and a Gaussian mixture model are trained. The autoencoder model adopts a four-layer structure: 1024-dimensional input layer → 256-dimensional encoding layer → 64-dimensional encoding layer → 256-dimensional decoding layer → 1024-dimensional decoding layer. The Gaussian mixture model has a certain number of Gaussian components. In this embodiment, we use 8.

[0084] During inference, the noise resistance features currently collected and extracted will be used. Input the autoencoder and calculate the reconstruction error. Simultaneously input a Gaussian mixture model and calculate the log-likelihood probability. After normalizing both, take According to the formula Generate acoustic anomaly score .

[0085] In this embodiment, a normal threshold is set. Define the abnormal threshold ,like If it is judged as normal; If it is directly judged as abnormal; If it falls into the fuzzy range, the parallel auxiliary judgment module is activated.

[0086] Assuming the current measured acoustic anomaly score If the signal falls into the fuzzy range, the system will automatically activate three auxiliary judgment channels: vibration, temperature, and current harmonics.

[0087] In the vibration channel, based on the vibration signals acquired by the triaxial accelerometer, the synthesized vibration amplitude is calculated, and the peak value, RMS, kurtosis, peak-to-peak value, and frequency domain features of the time domain are extracted. The vibration feature vector is then input into a pre-trained vibration autoencoder model to calculate the reconstruction error, which is then mapped to a vibration anomaly score using the sigmoid function. In this example, if the actual measurement Above the vibration anomaly threshold The vibration was determined to be abnormal.

[0088] In the temperature channel, the rate of temperature change is calculated based on the temperature data collected by the NTC temperature sensor. Set a safe temperature threshold Safe temperature rise rate threshold If the current temperature Temperature change rate If none of them exceed the threshold, then the temperature anomaly score is calculated. The temperature was determined to be normal.

[0089] In the current harmonic path, a Fast Fourier Transform is performed on the input current signal acquired by the current transformer to calculate the Total Harmonic Distortion (THD). The baseline during normal operation of the equipment is assumed. Limit value Current measured According to the formula Below the current harmonic anomaly threshold The current harmonics were determined to be normal.

[0090] Based on the above parallel auxiliary judgment results: That is, the vibration is abnormal, the temperature is normal, and the current harmonics are normal. According to the fault root cause inference rule, when the acoustic abnormality score and the vibration abnormality score are both higher than their respective thresholds, the fault root cause is inferred to be the loose transformer core or inductor resonance. This inference result is consistent with the actual inspection result - after subsequent disassembly and inspection, it was confirmed that the fixing glue of the transformer core inside the adapter had aged and cracked, which caused the core to loosen and produce abnormal vibration and noise.

[0091] Example 5

[0092] Please see Figures 1-2 The specific implementation of the signal-to-noise ratio driven dynamic weight adjustment and hierarchical response strategy described in steps five and six is ​​detailed in this embodiment.

[0093] First, the signal-to-noise ratio of the current environment is estimated in real time to compare the signal power collected by the near-end microphone and the far-end microphone. This is combined with the power estimates of the target signal and noise components after blind source separation, according to the formula... In this embodiment, the calculation assumes the current measured... It is a mildly noisy environment.

[0094] The fusion weights of each modal feature are adaptively adjusted based on the signal-to-noise ratio. In this embodiment, the weight parameters for the four modes (acoustic, vibration, temperature, and current harmonics) are set as follows: , ; , ; , ; , After Softmax normalization, under the condition of SNR=2dB, the modal weights are approximately: acoustic 0.35, vibration 0.28, temperature 0.17, and current harmonics 0.20.

[0095] According to the formula Calculate the fusion anomaly score, assuming the current anomaly scores for each channel are: , , , ,but .

[0096] The confidence score is further calculated. In this embodiment, the weighting coefficient is the signal-to-noise ratio. Separation quality Modal consistency ,set up , The current SNR is 2dB. Calculate... Blind source separation quality assessment Modal consistency Then the overall confidence level .

[0097] In this embodiment, a confidence threshold is set. (high), (middle), ,current , between and In between, the second-level response strategy is implemented—the multi-model integrated voting mechanism is activated.

[0098] The specific implementation of multi-model ensemble voting is as follows: Four prediction models are run in parallel—Model A is based on the original noisy signal, Model B is based on the signal after blind source separation, Model C is based on the signal after decoder denoising, and Model D is based only on non-audio modalities. The output results of each model are weighted and voted on using their respective real-time confidence scores. In this example, because the confidence score is at a moderate level, the downgrade prediction strategy has not yet been triggered, and the system outputs a fusion anomaly score. The confidence level was marked as "medium," and a manual review work order was triggered simultaneously.

[0099] In contrast, when the SNR drops to -5dB, the acoustic mode weight automatically drops to 0.12, while the current harmonic mode weight rises to 0.35. The contribution of the acoustic channel in the fusion decision is actively weakened, ensuring that the fusion anomaly score mainly comes from electrical and temperature signals with stronger noise resistance, effectively avoiding misjudgments caused by strong noise.

[0100] Example 6

[0101] Please see Figures 1-2 The specific implementation methods of health assessment, remaining lifespan prediction and incremental learning update described in steps seven and eight are detailed in this embodiment.

[0102] First, a multimodal anomaly scoring time-series database for the entire lifecycle of the target device is established. In this embodiment, the adapter is monitored for a cumulative period of 6 months, and data such as fusion anomaly scores, anomaly scores of each channel, and device operating conditions are recorded daily to form a time-series database, with a sampling interval of 1 hour.

[0103] Based on the aforementioned time-series database, an LSTM network was used for modeling. The network structure was as follows: input dimension 4 (anomaly scores for four modalities), hidden layer dimension 64, number of layers 2, and output dimension 1 (health index or remaining lifespan RUL value). During training, 80% of the historical data was used as the training set, and 20% as the validation set.

[0104] According to the formula Calculate the health index, where In this embodiment, we take 1.2. This is the moving average of the fusion anomaly scores over the past 24 hours. If the current... ,but This indicates that the device's health has dropped to 58%, placing it in a "needs attention" state.

[0105] The LSTM network predicts the RUL range for the next 30 days based on historical trends. Calculations show that... sky, Heaven, then The adapter is expected to fail within 30-60 days (95% confidence interval).

[0106] Based on the above prediction results, the system automatically generates maintenance recommendations: Given that the root cause of the fault is inferred to be "loose transformer core" (inference from Example 4), and the remaining lifespan is predicted to be approximately 30-60 days, it is recommended to arrange planned maintenance within the next two weeks to re-fix the transformer core or replace the adapter to avoid unplanned downtime.

[0107] Furthermore, after manual verification and on-site repair, the actual fault type (loose transformer core), repair actions (re-applying adhesive to fix the magnetic core), and voiceprint changes over a week following the repair are used as labeled samples for back-injection training. This embodiment employs an elastic weighted consolidation method for incremental learning, with regularization coefficients... We started with 1000 samples. An incremental learning update was triggered every 100 newly labeled samples, with a full fine-tuning performed monthly. After three months of continuous iteration, the model's anomaly detection accuracy improved from the initial 93% to 97%.

[0108] Contents not described in detail in this specification are existing technologies known to those skilled in the art. Standard parts used in this invention can be purchased commercially, and irregularly shaped parts can be custom-made according to the description and drawings. The specific connection methods for each part all employ conventional methods such as bolts, rivets, and welding, which are already mature technologies. The machinery, parts, and equipment all use conventional models from the prior art, and the circuit connections also employ conventional connection methods from the prior art, which will not be detailed here.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent prediction and maintenance of equipment status based on the Industrial Internet, characterized in that, Includes the following steps: Step 1: Collect the operating data of the target device by deploying a multimodal sensor array at the target device. The multimodal sensor array includes a sound sensor, a vibration sensor, a temperature sensor, and a current sensor. Simultaneously collect the power information of the target device as activity information. Step 2: Based on the activity information, perform blind source separation guided by activity information on the collected sound signals to separate the target device sound component and the environmental noise component from the mixed sound signal; Step 3: Perform noise reduction and feature enhancement processing on the separated target device sound components, and extract noise immunity features; Step 4: Calculate the acoustic anomaly score based on the noise immunity characteristics, and simultaneously calculate the vibration anomaly score, temperature anomaly score, and current harmonic anomaly score based on the data collected by the vibration sensor, temperature sensor, and current sensor, respectively. The temperature anomaly score and current harmonic anomaly score are obtained by comparing them with their respective preset thresholds. Step 5: Estimate the signal-to-noise ratio of the current environment in real time, and dynamically adjust the weight of each modal feature in the fusion decision based on the signal-to-noise ratio to generate a fusion anomaly score; Step 6: Calculate the confidence score of the current prediction based on the fusion anomaly score, and execute the hierarchical response strategy according to the confidence score; Step 7: Based on the output of the graded response strategy, perform a health assessment and predict the remaining service life of the target equipment, and generate maintenance recommendations.

2. The method for intelligent prediction and maintenance of equipment status based on the Industrial Internet according to claim 1, characterized in that, Step two, the activity information-guided blind source separation, specifically includes: Step 2.1: Determine the current operating state of the target device using the power information. The operating state includes standby state, light load state, and full load state. Step 2.2: Construct a state-voiceprint mapping dictionary, using the current working state as a priori label, and embed state condition constraints in the blind source separation network; Step 2.3: Utilizing the periodicity of the target device's power frequency harmonics, the mixed audio signal is periodically divided into multiple units, and each unit independently calculates an adaptive separation threshold; Step 2.4: Based on the state condition constraints and the adaptive separation threshold, extract the target device sound component that matches the current working state from the mixed sound signal.

3. The method for intelligent prediction and maintenance of equipment status based on the Industrial Internet according to claim 1, characterized in that, The noise reduction and feature enhancement processing described in step three specifically includes: Step 3.1: Perform deep noise reduction processing on the separated target device sound components using a stacked noise reduction autoencoder. The stacked noise reduction autoencoder is pre-trained with clean device sound samples with added Gaussian noise and mask noise. Step 3.2: Perform multi-scale time-frequency joint analysis on the denoised audio signal, extract multi-granular features in the time domain through a multi-scale deep residual network, and analyze the frequency domain features in the Mel speech spectrogram through a multi-scale attention interleaved network; Step 3.3: Extract frequency features using dynamic frequency bending technology, and adaptively adjust the feature extraction granularity according to the signal noise level.

4. The method for intelligent prediction and maintenance of equipment status based on the Industrial Internet according to claim 1, characterized in that, The acoustic anomaly score mentioned in step four is calculated in the following way: Step 4.1: Train the autoencoder model and Gaussian mixture model using the noise immunity characteristics of the target device under normal operating conditions; Step 4.2: Input the currently acquired noise resistance features into the trained autoencoder model and Gaussian mixture model respectively, and calculate the reconstruction error and log-likelihood probability; Step 4.3: Combine the reconstruction error and the log-likelihood probability to generate a normalized acoustic anomaly score. .

5. The method for intelligent prediction and maintenance of equipment status based on the Industrial Internet according to claim 4, characterized in that, The vibration anomaly score, temperature anomaly score, and current harmonic anomaly score are activated as parallel auxiliary judgments when the acoustic anomaly score falls within a preset fuzzy range; the preset fuzzy range ,in This is the normal threshold. To clarify the abnormal threshold; The vibration anomaly score is calculated by extracting time-domain and frequency-domain features from the triaxial vibration signals collected by the vibration sensor and then calculating them through a vibration autoencoder model. The time-domain features include peak value, root mean square, kurtosis, and peak-to-peak value, and the frequency-domain features include the dominant frequency energy distribution and harmonic components. The temperature anomaly score is obtained by extracting the absolute temperature value, temperature change rate, and temperature fluctuation variance from the temperature signal collected by the temperature sensor and comparing it with a preset temperature threshold. The current harmonic anomaly score is obtained by performing a fast Fourier transform on the input current signal collected by the current sensor, extracting the total harmonic distortion rate and the amplitude ratio of each harmonic, and then comparing it with a preset harmonic baseline value.

6. The method for intelligent prediction and maintenance of equipment status based on the Industrial Internet according to claim 1, characterized in that, Step five, which describes dynamically adjusting the weights of each modal feature based on the signal-to-noise ratio, specifically includes: When the signal-to-noise ratio is higher than the first preset threshold, acoustic modal features have the highest weight in the fusion decision. When the signal-to-noise ratio is lower than the second preset threshold, the weight of acoustic modal features is reduced, while the weight of current harmonic mode features and temperature mode features is increased. The weights of each modal feature are continuously adjusted as the signal-to-noise ratio changes.

7. The method for intelligent prediction and maintenance of equipment status based on the Industrial Internet according to claim 1, characterized in that, The factors for calculating the confidence score in step six include: the current signal-to-noise ratio, the blind source separation quality assessment results, and the degree of consistency between the anomaly scores of each modality; The hierarchical response strategy includes: When the confidence score is higher than the first confidence threshold, the fusion anomaly score is directly output as the prediction result; When the confidence score is between the first confidence threshold and the second confidence threshold, a multi-model integration voting mechanism is activated. The multi-model includes a first prediction model based on the original signal, a second prediction model based on the signal after blind source separation, a third prediction model based on the signal after noise reduction, and a fourth prediction model based on non-audio modes. The real-time confidence scores of each model are used as weights for weighted voting. When the confidence score is between the second confidence threshold and the third confidence threshold, the acoustic modal channel is suspended, and downgraded prediction is performed only based on the vibration mode, temperature mode and current harmonic mode, and manual review is triggered. When the confidence score is lower than the third confidence threshold, automatic prediction is paused, the original collected data is recorded, and on-site inspection is triggered.

8. The method for intelligent prediction and maintenance of equipment status based on the Industrial Internet according to claim 1, characterized in that, Step seven, the health assessment and remaining lifespan prediction, specifically includes: Step 7.1: Establish a multimodal anomaly scoring time-series database for the entire lifecycle of the target device; Step 7.2: Model the time series database using a long short-term memory network, and output the health index and the remaining lifespan prediction range; Step 7.3: Based on the combination pattern of the acoustic anomaly score, vibration anomaly score, temperature anomaly score, and current harmonic anomaly score, infer the root cause type of the fault; Step 7.4: Generate corresponding maintenance recommendations based on the root cause type of the fault.

9. The method for intelligent prediction and maintenance of equipment status based on the Industrial Internet according to claim 8, characterized in that, The inference rules for the root cause type of the fault include: When both the acoustic anomaly score and the vibration anomaly score are higher than their respective thresholds, the root cause of the fault is inferred to be loose transformer core or inductive resonance. When both the acoustic anomaly score and the current harmonic anomaly score are higher than their respective thresholds, the root cause of the fault is inferred to be aging of the switching transistor or an open circuit of the rectifier diode. When both the acoustic anomaly score and the temperature change rate anomaly score exceed their respective thresholds, the root cause of the fault is inferred to be an increase in the equivalent series resistance of the electrolytic capacitor or an overload of the equipment.

10. The method for intelligent prediction and maintenance of equipment status based on the Industrial Internet according to claim 1, characterized in that, It also includes step eight: using the manual review results and the actual fault type, repair actions and voiceprint changes after repair as labeled samples to back-inject training dataset, and incrementally learning and updating each model in the method.

Images