An experimental instrument intelligent diagnosis method based on deep learning

Abstract

This invention discloses a deep learning-based intelligent diagnostic method and system for experimental instruments, belonging to the field of intelligent diagnostic technology. The method includes: collecting operational data of the experimental instrument; calculating the workload coefficient based on the power change gradient and frequency distribution characteristics in the operational data; classifying the workload coefficient into load levels, establishing a hierarchical dynamic benchmark library, and obtaining a state benchmark vector; calculating the deviation matrix between real-time monitoring data and the state benchmark vector; inputting the deviation matrix into a convolutional long short-term memory network for feature learning; training a fault identification model through backpropagation; selecting a state benchmark vector from the hierarchical dynamic benchmark library based on the current workload coefficient; inputting the real-time deviation features into the fault identification model; and outputting a fault classification probability distribution. This invention improves the intelligence level and decision reliability of experimental instruments in operation, maintenance, and fault early warning.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology, and in particular to an intelligent diagnostic method for experimental instruments based on deep learning. Background Technology

[0002] Traditional diagnostic methods for experimental instruments mainly rely on periodic manual inspections or rule-based judgments based on fixed thresholds, which are ill-suited to the nonlinear and multi-dimensional dynamic changes in instrument operating conditions. In recent years, with the rapid development of IoT technology and big data acquisition systems, the large amounts of multi-source time-series data generated during the operation of experimental instruments, such as temperature, current, vibration, and power, have provided a rich data foundation for introducing intelligent algorithms for fault detection. Meanwhile, deep learning, especially convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), demonstrates excellent performance in temporal feature extraction and multi-classification tasks, making it an important research direction in the field of experimental instrument condition diagnosis.

[0003] Current technologies still have many shortcomings. First, most models do not fully consider the performance characteristics of instruments under different workloads, resulting in a lack of specificity in the model training process and difficulty in generalizing to various operating environments. Second, some solutions do not establish a hierarchical benchmark system that matches the working state, failing to provide representative reference vectors and leading to low accuracy in real-time deviation assessment. Furthermore, traditional state recognition methods based on static thresholds are prone to false alarms and false negatives in practical applications, making it difficult to meet the requirements of high-precision fault detection in experimental environments. Especially in complex tasks or long-term operation scenarios, how to dynamically perceive the workload and construct a refined state benchmark by combining historical evolution data has become a bottleneck in current technological development. Therefore, there is an urgent need for a deep learning-based intelligent diagnostic method for experimental instruments to improve the accuracy and adaptability of diagnosis. Summary of the Invention

[0004] In view of the problems that existing intelligent diagnostic technologies for experimental instruments have, such as inability to accurately model state characteristics under workload, weak generalization ability of diagnostic models, and low confidence in fault identification, this invention is proposed.

[0005] Therefore, the problem to be solved by this invention is how to construct a fault identification model with adaptability to operating conditions in the context of complex and ever-changing operating states and dynamic changes in workload of experimental instruments, so as to achieve accurate diagnosis and intelligent classification of the operating state of experimental instruments, thereby improving the accuracy, real-time performance and generalization ability of diagnosis.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide an intelligent diagnostic method for experimental instruments based on deep learning, comprising: Collect the operating data of the experimental instrument, and calculate the workload coefficient based on the power change gradient and frequency distribution characteristics in the operating data; The workload coefficient is divided into load levels, a hierarchical dynamic benchmark library is established, and the state benchmark vector is obtained. Calculate the deviation matrix between the real-time monitoring data and the state reference vector, input the deviation matrix into a convolutional long short-term memory network for feature learning, and generate a fault identification model through backpropagation training. Based on the current workload coefficient, a state benchmark vector is selected from the hierarchical dynamic benchmark library, and the real-time deviation features are input into the fault identification model to output the fault classification probability distribution.

[0007] As a preferred embodiment of the deep learning-based intelligent diagnostic method for experimental instruments described in this invention, the method includes: selecting a state reference vector from the hierarchical dynamic reference library based on the current workload coefficient, inputting real-time deviation features into the fault identification model, and outputting a fault classification probability distribution, including: Obtain the workload coefficient calculated in real time, determine the current load level based on the numerical range of the workload coefficient, and select a state reference vector from the hierarchical dynamic reference library; Based on the state reference vector, a real-time deviation feature vector is calculated, wherein the real-time deviation feature vector includes temperature deviation, vibration deviation, and current deviation. The real-time deviation feature vector is input into the fault identification model for inference calculation, and the fault classification probability distribution is output. Based on the fault classification probability distribution, fault types are generated, and the confidence scores output by the fault identification model are calculated. Based on the confidence level value, a confidence level assessment mechanism is established, and a dynamic adjustment strategy for the benchmark vector is designed, and a feedback correction mechanism is implemented.

[0008] As a preferred embodiment of the deep learning-based intelligent diagnostic method for experimental instruments described in this invention, the method generates fault types based on the fault classification probability distribution, including: When the maximum fault classification probability is less than the preset anomaly judgment threshold, the experimental instrument is marked as abnormal and the undefined fault diagnosis process is triggered; if the maximum fault classification probability is less than the first-level judgment threshold and the temperature deviation or vibration deviation exceeds the characteristic fluctuation range, the fault type of the experimental instrument is a first-level fault; if the maximum fault classification probability is greater than the first-level judgment threshold but less than the second-level judgment threshold, the fault type of the experimental instrument is a second-level fault; if the maximum fault classification probability is greater than the second-level judgment threshold, the fault type of the experimental instrument is a third-level fault. When the maximum fault classification probability value is greater than or equal to the preset anomaly judgment threshold, the experimental instrument is marked as normal and the benchmark reference data is updated.

[0009] As a preferred embodiment of the deep learning-based intelligent diagnostic method for experimental instruments described in this invention, the method includes: calculating the deviation matrix between real-time monitoring data and the state reference vector; inputting the deviation matrix into a convolutional long short-term memory network for feature learning; and training a fault identification model through backpropagation, comprising: Acquire real-time monitoring data from the experimental instruments and extract the temperature reference value, vibration reference value, and current reference value of the state reference vector; Calculate the deviation matrix between the real-time monitoring data and the state reference vector, wherein the rows of the deviation matrix represent the time step and the columns represent the parameter dimensions; The Z-score standardization method is used to preprocess the values ​​of the deviation matrix to obtain the standardized deviation matrix; Construct a convolutional long short-term memory network, wherein the convolutional long short-term memory network includes convolutional layers, long short-term memory layers, and fully connected layers; The standardized bias matrix is ​​input into the convolutional layer for feature extraction to obtain the convolutional feature map; The convolutional feature map is input into the long short-term memory layer, and the long-term dependencies in the convolutional feature map are learned through a gating mechanism to output a temporal feature vector; The temporal feature vector is input into the fully connected layer, and the fault classification prediction result is obtained by mapping through the activation function. A training dataset is constructed based on historical fault sample data, and the cross-entropy loss function is used to calculate the prediction error between the fault classification prediction result and the true label. The gradient values ​​of the prediction error with respect to the parameters of each layer of the convolutional short-term memory network are calculated using the backpropagation algorithm. Based on the gradient values, an adaptive momentum estimation optimizer is used to update the weight parameters and bias parameters of the convolutional short-term memory network to generate a fault identification model.

[0010] As a preferred embodiment of the deep learning-based intelligent diagnostic method for experimental instruments described in this invention, the following steps are taken: The workload coefficient is divided into load levels, a hierarchical dynamic benchmark library is established, and a state benchmark vector is obtained, including: Based on the rated power and historical operating data of the experimental instruments, the power range is divided into three power intervals; The workload factor is matched with the power range to establish a mapping relationship between the workload factor and the load level; Based on the matching results, a hierarchical dynamic benchmark library is established, wherein the hierarchical dynamic benchmark library includes a light load benchmark layer, a medium load benchmark layer, and a heavy load benchmark layer; During the operation of the experimental instrument, state parameter data corresponding to the working load coefficient are collected synchronously, wherein the state parameter data includes temperature data, vibration data and current data; Based on historical normal operation data, the corresponding state reference vector is initialized for the load level, and the temperature reference value, vibration reference value and current reference value of the state reference vector are respectively set to a numerical range. Statistical analysis is performed on the state parameter data to calculate the mean, standard deviation, and confidence interval of each parameter under different load levels, and the state reference vector for each load level is initialized. A moving average algorithm is used to calculate the moving average of the state baseline vector to establish a real-time data buffer. Based on the calculation results, the state reference vector corresponding to the load level is dynamically updated, and the numerical range is adjusted synchronously.

[0011] As a preferred embodiment of the deep learning-based intelligent diagnostic method for experimental instruments described in this invention, the method includes: matching the workload coefficient with the power range, comprising: If the workload factor is in the first power range, then the load level is light load. If the workload factor is in the second power range, then the load level is medium load. If the workload factor is in the third power range, then the load level is heavy load.

[0012] As a preferred embodiment of the deep learning-based intelligent diagnostic method for experimental instruments described in this invention, the method includes: collecting operational data of the experimental instrument and calculating the workload coefficient based on the power change gradient and frequency distribution characteristics in the operational data, including: The data acquisition module collects the operating data of the experimental instrument, including power data, voltage data, current data, rotational speed data, and load data. The power data is subjected to time series analysis to calculate the power change gradient at adjacent time points and form a power gradient sequence in chronological order. The power gradient sequence is subjected to frequency domain transformation, and the frequency distribution characteristics are obtained by fast Fourier transform. Based on the power change gradient and the frequency distribution characteristics, the workload coefficient is calculated using a weighted fusion algorithm.

[0013] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein the computer program instructions, when executed by the processor, implement the steps of the deep learning-based intelligent diagnostic method for experimental instruments as described in the first aspect of the present invention.

[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of the deep learning-based intelligent diagnostic method for experimental instruments as described in the first aspect of the present invention.

[0015] Compared with existing technologies, the advantages of this invention are as follows: By collecting operational data from experimental instruments and calculating the workload coefficient based on power change gradient and frequency distribution characteristics, a quantitative assessment of the load intensity of the current operating state of the equipment is achieved. This effectively reflects the operational complexity and fluctuation characteristics of the experimental instruments, improves the adaptability of the fault diagnosis model to multiple operating conditions, and avoids the misjudgment problem caused by neglecting load differences in traditional methods. By classifying the workload coefficient into load levels and establishing a hierarchical dynamic benchmark library based on this classification, the state benchmark vector under the corresponding level is obtained, realizing dynamic modeling of multi-dimensional features such as temperature, vibration, and current under different load levels. This effectively solves the defect that the traditional static threshold method cannot cover all operating conditions and improves the benchmark reference. Accuracy and timeliness are improved. Furthermore, by calculating the deviation matrix between real-time monitoring data and the state baseline vector, and inputting it into a convolutional long short-term memory network model for feature extraction and sequence modeling, deep learning of fault modes and automatic extraction of nonlinear temporal features are achieved. This effectively enhances the model's recognition accuracy and robustness under complex operating conditions. Simultaneously, a fault identification model is constructed by combining backpropagation algorithm and adaptive optimization strategy, providing a high-quality foundation for subsequent inference. Based on the current workload coefficient, a matching state baseline vector is selected, and combined with the real-time deviation feature vector, it is input into the fault identification model. The model outputs a fault classification probability distribution and generates a confidence score evaluation. This not only improves the transparency and interpretability of fault judgment but also dynamically adjusts the baseline library content through a confidence-driven feedback correction mechanism. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of the intelligent diagnostic method for experimental instruments based on deep learning, as described in Example 1. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0019] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0020] As mentioned in the background section, existing technologies still have many shortcomings. First, most models do not fully consider the performance characteristics of instruments under different workloads, resulting in a lack of specificity in the model training process and difficulty in generalizing to various operating environments. Second, some solutions do not establish a hierarchical benchmark system that matches the working state, failing to provide representative reference vectors and leading to low accuracy in real-time deviation assessment. Furthermore, traditional state recognition methods based on static thresholds are prone to false alarms and false negatives in practical applications, making it difficult to meet the requirements of high-precision fault detection in experimental environments. Especially in complex tasks or long-term operation scenarios, how to dynamically perceive the workload and combine it with historical evolution data to construct a refined state benchmark has become a bottleneck in current technological development. Therefore, there is an urgent need for a deep learning-based intelligent diagnostic method for experimental instruments to improve the accuracy and adaptability of diagnosis.

[0021] Figure 1 This is a flowchart illustrating a laboratory intelligent solution preparation control method based on the Internet of Things (IoT) according to an embodiment of the present invention. Figure 1 As shown, a laboratory intelligent liquid preparation control method based on the Internet of Things includes: S1: Collect the operating data of the experimental instrument and calculate the workload coefficient based on the power change gradient and frequency distribution characteristics in the operating data.

[0022] Specifically, the data acquisition module collects the operating data of the experimental instrument, which includes power data, voltage data, current data, rotational speed data, and load data; the power data is subjected to time series analysis to calculate the power change gradient at adjacent time points and form a power gradient sequence in chronological order.

[0023] Preferably, the formula for calculating the power change gradient is: ; in, The gradient value, The power value at time t+1 Let be the power value at time t. For time intervals.

[0024] It should be noted that the data acquisition module continuously monitors the operating status of the experimental instrument at a preset sampling frequency; the preset sampling frequency is 100Hz to 1000Hz; the time interval is determined by the preset sampling frequency; and the power gradient sequence reflects the dynamic power change characteristics of the experimental instrument.

[0025] Furthermore, the power gradient sequence is transformed in the frequency domain, and the frequency distribution characteristics are obtained through fast Fourier transform.

[0026] It should be noted that the frequency distribution characteristics include the dominant frequency component, the frequency peak distribution, and the spectral energy density; the dominant frequency component and the spectral energy density in the frequency distribution characteristics are used as feature vectors; the feature vectors are used to characterize the operating state characteristics of the experimental instrument.

[0027] Furthermore, based on the power change gradient and frequency distribution characteristics, the workload factor is calculated using a weighted fusion algorithm.

[0028] For example, during the operation of the experimental instrument, the data acquisition module continuously monitors parameters such as power, voltage, and current at a sampling frequency of 500Hz. Assuming the power value at time t is 1200W and at time t+1 it is 1250W, with a time interval of 0.002 seconds (corresponding to 500Hz), the power change gradient is calculated as (1250-1200) / 0.002 = 25000W / s. By analyzing the power gradient sequence using Fast Fourier Transform, it is found that the dominant frequency component is 15Hz, and the spectral energy density is concentrated in the 10-20Hz range. Thus, a feature vector [15, 0.85] is generated and weighted to obtain a workload coefficient of 0.62.

[0029] S2: Divide the workload coefficient into load levels, establish a hierarchical dynamic benchmark library, and obtain the state benchmark vector.

[0030] Specifically, based on the rated power and historical operating data of the experimental instruments, the power range is divided into three power intervals; the working load factor is matched with the power intervals to establish a mapping relationship between the working load factor and the load level.

[0031] It should be noted that the division of power ranges is determined based on the power characteristic curve of the experimental instrument and operating experience data; the first power range corresponds to 0-40% of the rated power range; the second power range corresponds to 40%-70% of the rated power range; and the third power range corresponds to 70%-100% of the rated power range.

[0032] Furthermore, if the workload factor is in the first power range, the load level is light load; if the workload factor is in the second power range, the load level is medium load; and if the workload factor is in the third power range, the load level is heavy load.

[0033] Preferably, load level identification codes are assigned to light loads, medium loads, and heavy loads respectively; the load level identification codes adopt a digital encoding form, where light loads correspond to identification code 1, medium loads correspond to identification code 2, and heavy loads correspond to identification code 3; the load level identification codes are linked to the working load coefficient for storage; the load level classification results are verified through historical fault data, and the fault occurrence frequency and fault type distribution under different load levels are statistically analyzed; based on the verification results, the power range division parameters are optimized and adjusted to ensure the accuracy and effectiveness of the load level classification.

[0034] Specifically, based on the matching results, a hierarchical dynamic benchmark library is established, which includes a light load benchmark layer, a medium load benchmark layer, and a heavy load benchmark layer.

[0035] It should be noted that each baseline layer stores the state baseline vector corresponding to the load level; the hierarchical dynamic baseline library is organized in a tree structure and automatically indexes the corresponding baseline layer according to the numerical range of the workload coefficient.

[0036] Furthermore, during the operation of the experimental instrument, state parameter data corresponding to the working load coefficient are collected synchronously, including temperature data, vibration data, and current data. Based on historical normal operation data, the state reference vector corresponding to the load level is initialized, and the temperature reference value, vibration reference value, and current reference value of the state reference vector are set with numerical range intervals respectively.

[0037] It should be noted that temperature data is acquired via a temperature sensor at a sampling frequency of 10Hz, vibration data is acquired via a vibration sensor at a sampling frequency of 1000Hz, and current data is acquired via a current sensor at a sampling frequency of 100Hz; the range of the temperature reference value is... The numerical range of the vibration reference value is The numerical range of the current reference value is The numerical range is determined based on statistical analysis under normal operating conditions and includes the mean plus or minus two standard deviations.

[0038] Preferably, the state reference vector is represented as follows: Where i is the load level identifier code, This is the temperature reference value for the corresponding load level. This is the vibration reference value for the corresponding load level. The current reference value is the value corresponding to the load level; the state reference vector corresponding to the light load is... The state reference vector corresponding to the medium load is: The state reference vector corresponding to the heavy load is .

[0039] Furthermore, statistical analysis is performed on the state parameter data to calculate the mean, standard deviation, and confidence interval of each parameter under different load levels, and the state baseline vector for each load level is initialized. It should be noted that the mean is used as the benchmark center value; the standard deviation is used to determine the fluctuation range of the benchmark; the confidence interval determines the upper and lower boundaries of the benchmark value; and the benchmark vector for light load conditions is denoted as... The reference vector for the medium-load state is denoted as The heavy load state reference vector is denoted as... Where T, V, and I are the temperature reference value, vibration reference value, and current reference value, respectively.

[0040] Specifically, statistical analysis is performed on the state parameter data to calculate the mean, standard deviation, and confidence interval of each parameter under different load levels, and the state baseline vector for each load level is initialized.

[0041] Preferably, the sliding window parameter is set for the moving average algorithm, and the length of the sliding window is determined according to the operating cycle characteristics of the experimental instrument; the length of the sliding window is set to N sampling points, where the value of N ranges from 50 to 200 sampling points; the sliding window performs sliding sampling of the state parameter data in chronological order, with each sliding step being 1 sampling point.

[0042] Furthermore, a moving average algorithm is used to calculate the moving average of the state baseline vector to establish a real-time data buffer.

[0043] It should be noted that the formula for calculating the moving average value of the temperature reference value is: The formula for calculating the sliding average value of the vibration reference value is: The formula for calculating the moving average value of the current reference value is: The moving average algorithm automatically updates its calculation results whenever new state parameter data is received.

[0044] Preferably, a real-time data buffer is established to store the latest collected temperature data, vibration data, and current data. The real-time data buffer adopts a circular buffer structure with a buffer capacity of 1000 data points. When the buffer is full, the oldest data is automatically overwritten. The real-time data buffer allocates data to the corresponding load level buffer queue according to the classification results of the workload coefficient.

[0045] Furthermore, based on the calculation results, the state baseline vector corresponding to the load level is dynamically updated, and the numerical range is adjusted synchronously.

[0046] It should be noted that if the workload factor corresponds to a light load, then the state reference vector should be updated. If the workload factor corresponds to a medium load, then update the state reference vector. If the workload factor corresponds to a heavy load, update the state reference vector. The numerical range adjustment adopts an adaptive algorithm. If the moving average exceeds the current numerical range, the numerical range interval is expanded according to a preset ratio. The numerical range adjustment formula is: new numerical range = original numerical range × (1 + adjustment coefficient), where the adjustment coefficient is dynamically determined according to the degree of deviation and the value range is 0.05-0.2.

[0047] For example, based on the load level range divided according to the rated power of 2000W, 0.62 falls into the second power range (40%-70%), and is determined to be a medium load (identifier code 2). The system retrieves the state reference vector [45±5℃, 2.1±0.3mm / s, 8.5±1.2A] from the load reference layer in the hierarchical dynamic reference library index. The real-time collected temperature, vibration, and current data are calculated using a sliding window (N=100) to obtain 48℃, 2.5mm / s, and 9.1A, with a deviation matrix of [3,0.4,0.6] from the reference value.

[0048] S3: Calculate the deviation matrix between the real-time monitoring data and the state baseline vector, input the deviation matrix into the convolutional long short-term memory network for feature learning, and generate a fault identification model through backpropagation training.

[0049] Specifically, real-time monitoring data of the experimental instruments is acquired, and the temperature reference value, vibration reference value, and current reference value of the state reference vector are extracted. The deviation matrix between the real-time monitoring data and the state reference vector is calculated, where the number of rows in the deviation matrix corresponds to the time step and the number of columns corresponds to the parameter dimension. The Z-score standardization method is used to preprocess the values ​​of the deviation matrix to obtain the standardized deviation matrix.

[0050] Preferably, the temperature deviation between the real-time temperature data and the temperature reference value is calculated, the vibration deviation between the real-time vibration data and the vibration reference value is calculated, and the current deviation between the real-time current data and the current reference value is calculated.

[0051] Furthermore, a convolutional long short-term memory network is constructed, comprising convolutional layers, long short-term memory layers, and fully connected layers. The standardized bias matrix is ​​input into the convolutional layers for feature extraction to obtain convolutional feature maps. The convolutional feature maps are then input into the long short-term memory layers, and a gating mechanism is used to learn the long-term dependencies in the convolutional feature maps to output temporal feature vectors. Finally, the temporal feature vectors are input into the fully connected layers, and the fault classification prediction results are obtained through activation function mapping.

[0052] Preferably, the convolutional layer uses a one-dimensional convolutional kernel with a kernel size of 3×1 and a stride of 1, used to extract local features of the standardized deviation matrix; the long short-term memory layer adopts a bidirectional structure, including forward long short-term memory units and backward long short-term memory units, and the number of hidden layer neurons is set to 128; the parameters of the bidirectional long short-term memory structure are configured, including forget gate weights, input gate weights, and output gate weights; the forward long short-term memory unit processes the standardized deviation matrix from the beginning to the end of the time series; the backward long short-term memory unit processes the standardized deviation matrix from the end to the beginning of the time series; the output feature vectors of the bidirectional processing are concatenated and fused to form bidirectional temporal features.

[0053] Furthermore, a training dataset is constructed based on historical fault sample data, and the cross-entropy loss function is used to calculate the prediction error between the fault classification prediction result and the true label. The gradient value of the prediction error with respect to the parameters of each layer of the convolutional short-term memory network is calculated through the backpropagation algorithm. Based on the gradient value, the weight parameters and bias parameters of the convolutional short-term memory network are updated using an adaptive momentum estimation optimizer to generate a fault identification model.

[0054] It should be noted that normal state samples consist of data whose deviation values ​​in the standardized deviation matrix are within the normal range, while fault state samples consist of data whose deviation values ​​in the standardized deviation matrix exceed the normal range. Each training sample contains an input feature vector and a corresponding fault type label.

[0055] For example, the standardized deviation matrix is ​​input into a convolutional long short-term memory network. The convolutional layer uses a 3×1 kernel to extract local features, and the bidirectional LSTM layer (128 neurons) outputs temporal features. Finally, the fully connected layer generates a fault probability distribution: [normal 0.15, bearing wear 0.65, coil overheating 0.2]. Because the maximum probability of 0.65 exceeds the first-level threshold of 0.6 but is lower than the second-level threshold of 0.8, an orange warning is triggered, limiting the equipment power to 1500W and generating a maintenance suggestion. At this time, the confidence level is 0.5 (0.65-0.2=0.45<0.6), and the system initiates a multi-level verification mechanism. Finding that the vibration deviation continues to increase, it is upgraded to a second-level fault judgment and an expert review is requested. When the ambient temperature rises sharply by 10℃, the temperature compensation algorithm relaxes the confidence threshold to 0.48 to avoid misjudgment. Meanwhile, if the sliding average value exceeds the original range (48℃ > 50℃ upper limit), the system expands the temperature reference range to [45±5.5℃] by an adjustment factor of 0.1, and dynamically updates the heavy load state reference vector to ensure adaptive capability.

[0056] S4: Based on the current workload coefficient, select the state benchmark vector from the hierarchical dynamic benchmark library, input the real-time deviation features into the fault identification model, and output the fault classification probability distribution.

[0057] Specifically, the workload coefficient is obtained in real time, the current load level is determined based on the numerical range of the workload coefficient, and a state reference vector is selected from the hierarchical dynamic reference library; based on the state reference vector, a real-time deviation feature vector is calculated, which includes temperature deviation, vibration deviation and current deviation.

[0058] It should be noted that if the current load is light, the light load reference vector V_light is selected; if the current load is medium, the medium load reference vector V_medium is selected; and if the current load is heavy, the heavy load reference vector V_heavy is selected. The reference vectors include the temperature reference value, vibration reference value, and current reference value for the corresponding load level.

[0059] Furthermore, the real-time deviation feature vector is input into the fault identification model for inference calculation, and the fault classification probability distribution is output. Based on the fault classification probability distribution, the fault type is generated, and the confidence value output by the fault identification model is calculated.

[0060] Preferably, when the maximum fault classification probability is less than the preset anomaly judgment threshold, the experimental instrument is marked as abnormal and an undefined fault diagnosis process is triggered; if the maximum fault classification probability is less than the first-level judgment threshold and the temperature deviation or vibration deviation exceeds the characteristic fluctuation range, the fault type of the experimental instrument is a first-level fault, and a yellow warning signal is triggered, the equipment operating load is reduced to a safe level, and a high-frequency monitoring mode is started to record fault characteristic data; if the maximum fault classification probability is greater than the first-level judgment threshold but less than the second-level judgment threshold, the fault type of the experimental instrument is a second-level fault, and an orange warning signal is triggered, the equipment operating parameters are restricted to a safe range, a maintenance suggestion report is generated, and priority maintenance is arranged; if the maximum fault classification probability is greater than the second-level judgment threshold, the fault type of the experimental instrument is a third-level fault, and a red alarm is immediately triggered and an emergency shutdown is performed, the fault site data is locked, a fault analysis report is generated, and maintenance personnel are notified; when the maximum fault classification probability is greater than or equal to the preset anomaly judgment threshold, the experimental instrument is marked as normal and the benchmark reference data is updated.

[0061] It should be noted that the preset anomaly judgment threshold is a probability lower limit determined based on the statistical distribution characteristics of the equipment's historical normal operation data; the first-level judgment threshold is set based on the cluster center distance of typical primary fault modes in the feature space; and the second-level judgment threshold is determined comprehensively based on the equipment's safe operation boundary conditions and the requirements for identifying severe fault modes.

[0062] Preferably, the confidence score is calculated using the difference between the highest probability value and the second highest probability value; the confidence score ranges from 0 to 1, with a higher value indicating a more reliable recognition result. When the confidence score is below 0.3, the baseline vector adjustment mechanism is triggered.

[0063] Specifically, a confidence assessment mechanism is established based on the confidence level value, and a dynamic adjustment strategy for the benchmark vector is designed, with a feedback correction mechanism implemented.

[0064] Furthermore, when the confidence level is ≥0.8, the prediction result is deemed reliable, the current fault diagnosis conclusion will be directly adopted, the equipment status database will be updated, and the corresponding fault level handling plan will be executed. When the confidence level is 0.6 ≤ confidence level <0.8, the prediction result is deemed to need verification, and a multi-level verification mechanism will be initiated. If the confidence level shows an upward trend for three consecutive sampling periods, the current prediction result will be adopted and the confidence level rating will be increased. If the deviation of key characteristic parameters exceeds the safety threshold, the fault level judgment will be upgraded. If the auxiliary sensor data is consistent with the main prediction result, the current diagnosis conclusion will be confirmed. If there is a contradiction in characteristic parameters, the expert system review process will be triggered. When the confidence level is <0.6, the prediction result is deemed unreliable, the current abnormal data will be temporarily stored, the enhanced monitoring mode will be initiated, historical operating data will be reviewed and compared, and a reassessment will be conducted after the confidence level recovers to the reliable range.

[0065] Furthermore, for confidence level correction under special operating conditions: if the equipment is in the start-up / shutdown transition phase, the confidence level threshold is automatically relaxed by 20%; if the ambient temperature exceeds the rated range, the temperature compensation algorithm is activated to adjust the confidence level calculation; for older equipment, the confidence level is weighted and corrected based on the running time.

[0066] In summary, this invention, by collecting operational data from experimental instruments and calculating the workload coefficient based on power change gradient and frequency distribution characteristics, achieves a quantitative assessment of the load intensity of the equipment's current operating state. This effectively reflects the operational complexity and fluctuation characteristics of the experimental instruments, improves the adaptability of the fault diagnosis model to multiple operating conditions, and avoids the misjudgment problem caused by neglecting load differences in traditional methods. By classifying the workload coefficient into load levels and establishing a hierarchical dynamic benchmark library based on this classification, the invention obtains the state benchmark vector under the corresponding level, enabling dynamic modeling of multi-dimensional features such as temperature, vibration, and current under different load levels. This effectively solves the deficiency of traditional static threshold methods in not being able to cover all operating conditions, and improves the accuracy and timeliness of benchmark references. Furthermore, by calculating the deviation matrix between real-time monitoring data and the state baseline vector, and inputting it into a convolutional long short-term memory network model for feature extraction and sequence modeling, deep learning of fault modes and automatic extraction of nonlinear temporal features are achieved, effectively enhancing the model's recognition accuracy and robustness under complex operating conditions. Simultaneously, a fault identification model is constructed by combining backpropagation algorithm and adaptive optimization strategy, providing a high-quality foundation for subsequent inference. Based on the current workload coefficient, a matching state baseline vector is selected, and combined with the real-time deviation feature vector, it is input into the fault identification model to output the fault classification probability distribution and generate a confidence evaluation. This not only improves the transparency and interpretability of fault judgment but also dynamically adjusts the baseline library content through a confidence-driven feedback correction mechanism.

[0067] This embodiment also provides a computer device applicable to the intelligent diagnostic method for experimental instruments based on deep learning, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent diagnostic method for experimental instruments based on deep learning as proposed in the above embodiment.

[0068] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0069] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based intelligent diagnosis method for experimental instruments, characterized in that: include, Collect the operating data of the experimental instrument, and calculate the workload coefficient based on the power change gradient and frequency distribution characteristics in the operating data; The workload coefficient is classified into load levels to establish a hierarchical dynamic benchmark library and obtain the state benchmark vector; the state benchmark vector is represented as follows: Where i is the load level identifier code, This is the temperature reference value for the corresponding load level. This is the vibration reference value for the corresponding load level. This is the current reference value for the corresponding load level; A training dataset is constructed based on historical fault sample data. The deviation matrix between the monitoring data in the training dataset and the state baseline vector is calculated. The deviation matrix is ​​input into a convolutional long short-term memory network for feature learning. A fault identification model is generated through backpropagation training. Based on the current workload coefficient, a state benchmark vector is selected from the hierarchical dynamic benchmark library, the real-time deviation features are input into the fault identification model, and the fault classification probability distribution is output. Obtain the workload coefficient calculated in real time, determine the current load level based on the numerical range of the workload coefficient, and select a state reference vector from the hierarchical dynamic reference library; Based on the state reference vector, a real-time deviation feature vector is calculated, wherein the real-time deviation feature vector includes temperature deviation, vibration deviation, and current deviation. The real-time deviation feature vector is input into the fault identification model for inference calculation, and the fault classification probability distribution is output. Based on the fault classification probability distribution, fault types are generated, and the confidence scores output by the fault identification model are calculated. Based on the aforementioned confidence level values, a confidence level assessment mechanism is established, and a dynamic adjustment strategy for the benchmark vector is designed, along with a feedback correction mechanism. The data acquisition module collects the operating data of the experimental instrument, including power data, voltage data, current data, rotational speed data, and load data. The power data is subjected to time series analysis to calculate the power change gradient at adjacent time points and form a power gradient sequence in chronological order. The power gradient sequence is subjected to frequency domain transformation, and the frequency distribution characteristics are obtained by fast Fourier transform. Based on the power change gradient and the frequency distribution characteristics, the workload coefficient is calculated using a weighted fusion algorithm.

2. The intelligent diagnostic method for experimental instruments based on deep learning as described in claim 1, characterized in that: Based on the fault classification probability distribution, fault types are generated, including: When the maximum fault classification probability is less than the preset anomaly judgment threshold, the experimental instrument is marked as abnormal and the undefined fault diagnosis process is triggered; if the maximum fault classification probability is less than the first-level judgment threshold and the temperature deviation or vibration deviation exceeds the characteristic fluctuation range, the fault type of the experimental instrument is a first-level fault; if the maximum fault classification probability is greater than the first-level judgment threshold but less than the second-level judgment threshold, the fault type of the experimental instrument is a second-level fault; if the maximum fault classification probability is greater than the second-level judgment threshold, the fault type of the experimental instrument is a third-level fault. When the maximum fault classification probability value is greater than or equal to the preset anomaly judgment threshold, the experimental instrument is marked as normal and the benchmark reference data is updated.

3. The intelligent diagnostic method for experimental instruments based on deep learning as described in claim 2, characterized in that: A training dataset is constructed based on historical fault sample data. The deviation matrix between the monitoring data and the state baseline vector in the training dataset is calculated. This deviation matrix is ​​then input into a convolutional long short-term memory network for feature learning. A fault identification model is generated through backpropagation training, including: Acquire centralized monitoring data of the experimental instrument's training dataset, and extract the temperature reference value, vibration reference value, and current reference value of the state reference vector; Calculate the deviation matrix between the monitoring data and the state baseline vector, wherein the rows of the deviation matrix represent the time step and the columns represent the parameter dimensions; The Z-score standardization method is used to preprocess the values ​​of the deviation matrix to obtain the standardized deviation matrix; Construct a convolutional long short-term memory network, wherein the convolutional long short-term memory network includes convolutional layers, long short-term memory layers, and fully connected layers; The standardized bias matrix is ​​input into the convolutional layer for feature extraction to obtain the convolutional feature map; The convolutional feature map is input into the long short-term memory layer, and the long-term dependencies in the convolutional feature map are learned through a gating mechanism to output a temporal feature vector; The temporal feature vector is input into the fully connected layer, and the fault classification prediction result is obtained by mapping through the activation function. A training dataset is constructed based on historical fault sample data, and the cross-entropy loss function is used to calculate the prediction error between the fault classification prediction result and the true label. The gradient values ​​of the prediction error with respect to the parameters of each layer of the convolutional short-term memory network are calculated using the backpropagation algorithm. Based on the gradient values, an adaptive momentum estimation optimizer is used to update the weight parameters and bias parameters of the convolutional short-term memory network to generate a fault identification model.

4. The intelligent diagnostic method for experimental instruments based on deep learning as described in claim 3, characterized in that: The workload coefficient is divided into load levels, a hierarchical dynamic benchmark library is established, and a state benchmark vector is obtained, including: Based on the rated power and historical operating data of the experimental instruments, the power range is divided into three power intervals; The workload factor is matched with the power range to establish a mapping relationship between the workload factor and the load level; Based on the matching results, a hierarchical dynamic benchmark library is established, wherein the hierarchical dynamic benchmark library includes a light load benchmark layer, a medium load benchmark layer, and a heavy load benchmark layer; During the operation of the experimental instrument, state parameter data corresponding to the working load coefficient are collected synchronously, wherein the state parameter data includes temperature data, vibration data and current data; Based on historical normal operation data, the corresponding state reference vector is initialized for the load level, and the temperature reference value, vibration reference value and current reference value of the state reference vector are respectively set to a numerical range. Statistical analysis is performed on the state parameter data to calculate the mean, standard deviation, and confidence interval of each parameter under different load levels, and the state reference vector for each load level is initialized. A moving average algorithm is used to calculate the moving average of the state baseline vector to establish a real-time data buffer. Based on the calculation results, the state reference vector corresponding to the load level is dynamically updated, and the numerical range is adjusted synchronously.

5. The intelligent diagnostic method for experimental instruments based on deep learning as described in claim 4, characterized in that: Matching the workload factor with the power range includes: If the workload factor is in the first power range, then the load level is light load. If the workload factor is in the second power range, then the load level is medium load. If the workload factor is in the third power range, then the load level is heavy load.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep learning-based intelligent diagnostic method for experimental instruments as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep learning-based intelligent diagnostic method for experimental instruments as described in any one of claims 1 to 5.

Images