Transmitter system based on multi-sensor fusion and edge AI and self-diagnosis method

Abstract

The invention discloses a transmitter system based on multi-sensor fusion and edge AI and a self-diagnosis method, and belongs to the technical field of industrial process control instruments, and an industrial intelligent transmitter system comprises a multi-sensor unit, a signal acquisition unit, a microcontroller, an AI coprocessor and a dual-mode communication unit. The multiple sensing units are used for synchronously collecting original data; the signal acquisition unit is used for converting original data into digital signals; the microcontroller is used for processing the digital signal and carrying out environment adaptive compensation to obtain compensated sensor data; the AI coprocessor is used for performing feature extraction and fault probability reasoning on the compensated sensor data to realize real-time self-diagnosis; and the dual-mode communication unit is used for transmitting the compensated process control signal and diagnosis information in different channels. The system has the functions of environment adaptive compensation, real-time self-diagnosis and the like, and the problems of parameter drift and fault response delay in an industrial environment are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial process control instrumentation technology, specifically relating to a transmitter system and self-diagnostic method based on multi-sensor fusion and edge AI. Background Technology

[0002] In industrial process monitoring fields such as petrochemicals and power generation, transmitters are the core for acquiring key process parameters (such as pressure, temperature, and flow rate). Traditional transmitters primarily function as single-parameter signal conversion and transmission, and suffer from the following significant problems: (1) Environmental sensitivity: The industrial environment is complex and changeable (temperature fluctuations, vibrations, electromagnetic interference, etc.), which can easily cause sensor measurement values ​​to drift (parameter drift), affecting measurement accuracy and reliability.

[0003] (2) Delayed fault diagnosis: Fault diagnosis of traditional transmitters often relies on the background system or periodic maintenance, resulting in a large response delay. It is difficult to detect and locate sudden faults or performance degradation in a timely manner, which poses a safety hazard.

[0004] (3) Information silos: The information from a single sensor is limited and it is difficult to fully reflect the status of the equipment or process; moreover, process control signals and equipment status diagnostic information are usually transmitted in a mixed manner, which is inefficient and may also interfere with each other.

[0005] (4) Low level of intelligence: It lacks local data processing and decision-making capabilities, relies on the upper-level system for complex analysis and compensation, has poor real-time performance, and increases network load.

[0006] While existing technologies incorporate intelligent elements, they still fall short in effectively addressing the dynamic coupling effects of the environment, achieving high-real-time localized self-diagnosis, and optimizing information transmission architecture. Therefore, a smart transmitter system that integrates multiple sensors, enables adaptive environmental compensation, and provides real-time fault diagnosis is needed. Summary of the Invention

[0007] This invention provides a transmitter system and self-diagnostic method based on multi-sensor fusion and edge AI, which has functions such as environmental adaptive compensation and real-time self-diagnosis, and significantly improves the problems of parameter drift and fault response delay in industrial environments.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a transmitter system based on multi-sensor fusion and edge AI, comprising: The multi-sensor unit includes a main sensor and an auxiliary sensor for synchronously acquiring raw data, which includes the main process parameters of the industrial process acquired by the main sensor and the environmental parameters acquired by the auxiliary sensor. The signal acquisition unit, connected to the multi-sensor unit, is used to filter and amplify the acquired raw data and convert it into a digital signal. A microcontroller, connected to the signal acquisition unit, is used to process the digital signal and run a multi-sensor fusion algorithm for environmental adaptive compensation to obtain compensated sensor data. The AI ​​coprocessor is connected to the microcontroller and is used to run a pre-set neural network model to extract features and infer fault probability from the compensated sensor data, thereby achieving real-time self-diagnosis. A dual-mode communication unit, connected to the microcontroller and AI coprocessor, is used to transmit diagnostic results and compensated sensor data through separate channels.

[0009] Furthermore, the multi-sensor fusion algorithm is as follows:

[0010] In the formula, To compensate for the main sensor data, The data consists of raw data from the main sensor, where α is the temperature compensation coefficient, β is the vibration influence factor, t is the upper limit of integration, τ is the integration variable, T is the temperature, and α(τ) is the time-domain signal function of the vibration acceleration. This is a reference temperature.

[0011] Furthermore, the temperature compensation coefficient is obtained by establishing a mapping relationship between temperature and the temperature compensation coefficient using a cubic polynomial, which is: ,in, c 0、 c 1. c 2. c 3 is a constant coefficient.

[0012] Furthermore, the main sensor is a pressure sensor or a flow sensor; the auxiliary sensors include at least two of the following: a temperature sensor, a vibration sensor, and a humidity sensor.

[0013] Furthermore, the signal acquisition unit includes a signal conditioning circuit and a 24-bit Σ-Δ analog-to-digital converter; the signal conditioning circuit includes filtering and amplification circuits for preliminary filtering, noise reduction, and amplification of the original multi-source data signal; the 24-bit Σ-Δ analog-to-digital converter is used to convert the analog signal preprocessed by the signal conditioning circuit into a digital signal and transmit it to the microcontroller.

[0014] Furthermore, the neural network model is an INT8 quantized ResNet-8 model, which includes an input layer, a convolutional layer, a max pooling layer, four residual blocks, a global average pooling layer, a fully connected layer, and an output layer connected in sequence. The input layer receives a 128-dimensional × 3-channel feature vector, with the three channels corresponding to pressure, temperature, and vibration data, respectively; the number of channels in the four residual blocks are 16, 16, 32, and 32, respectively; the output layer outputs the probability distributions of three types of operating conditions: normal operating conditions, sensor drift, and circuit or mechanical failure.

[0015] Furthermore, the dual-mode communication unit includes both wired and wireless communication modes; the wired communication outputs a standard 4-20mA analog signal and is compatible with the HART protocol; the wireless communication outputs LoRa WAN.

[0016] Secondly, this invention provides a self-diagnostic method for an industrial intelligent transmitter, comprising the following steps: S1: Simultaneously acquire raw data from multiple sensors, perform signal conditioning and analog-to-digital conversion to obtain digital signals; S2: Denoise and filter the digital signal, and run a multi-sensor fusion algorithm for adaptive environmental compensation; S3: Extract the time-domain and frequency-domain features of the compensated data to generate multi-channel feature vectors; S4: Input the multi-channel feature vector into the trained ResNet-8 model and output the fault probability; S5: Execute local alarms or upload data based on the fault probability value.

[0017] Furthermore, the training process of the ResNet-8 model includes the following steps: SA1: Collect synchronous three-channel time-series data of pressure, temperature and vibration from multiple industrial transmitters during operation, and construct training sets and validation sets containing known fault modes respectively; SA2: Perform data augmentation and channel normalization on the time-series data in the training and validation sets, and extract time-domain and frequency-domain features to generate corresponding multi-channel feature vector samples. SA3: Initialize the ResNet-8 model parameters. The weights of the convolutional layers are initialized using the He normal distribution. The scaling factor γ of the batch normalization layer is initialized to 1 and the offset factor β is initialized to 0. SA4: The ResNet-8 model is trained using the cross-entropy loss function and the Adam optimization algorithm. A gradient pruning threshold is set, and the adaptive moment estimation algorithm is used for optimization during training. An early stopping mechanism is implemented based on the validation set loss to save the optimal model.

[0018] Furthermore, the trained ResNet-8 model undergoes three levels of optimization before deployment: INT8 quantization, layer fusion, and channel pruning.

[0019] Compared with the prior art, the present invention has at least the following beneficial technical effects: (1) This invention integrates multiple sensing units to achieve collaborative sensing of multi-source data, thus eliminating the problem of information silos in traditional transmitters, i.e., the information from a single sensor is limited and it is difficult to fully reflect the status of the equipment or process.

[0020] Furthermore, this invention, through a multi-sensor fusion algorithm, can achieve environmental adaptive compensation, significantly improving the problems of measurement drift, low measurement accuracy, and low reliability caused by the complex and variable industrial environment (temperature fluctuations, vibration, etc.) of traditional transmitters.

[0021] (2) The present invention integrates an AI coprocessor, which performs real-time fault diagnosis by running a pre-set neural network model and uploads fault information, which significantly improves the problem that fault diagnosis of traditional transmitters often depends on the background system or periodic maintenance.

[0022] (3) The present invention adopts a dual-mode communication unit to realize both wired and wireless transmission. The dual-mode communication unit transmits through separate channels, separating the compensated sensor data with high real-time requirements from the relatively non-real-time diagnostic / status information, avoiding mutual interference, ensuring the priority and real-time performance of process control, and ensuring the effective delivery of diagnostic information.

[0023] The self-diagnostic method provided by this invention first performs environmental adaptive compensation on the data, significantly reducing the impact of environmental factors such as temperature and vibration on the accuracy of the data, making the compensated data more realistically reflect the actual operating conditions of the equipment. Based on this, the time-domain and frequency-domain features of the compensated data are extracted and multi-channel feature vectors are generated to achieve a comprehensive representation of key fault-related information. Subsequently, the multi-channel feature vectors are input into a trained ResNet-8 model to output the fault probability. Utilizing the model's feature learning and pattern recognition capabilities, the corresponding fault occurrence probability is output, thereby achieving accurate identification of the equipment's health status and overcoming the shortcomings of traditional fault diagnosis methods based on empirical rules in terms of generalization ability and adaptability. Furthermore, a hierarchical response, either local alarm or data upload, is executed based on the fault probability value, enabling high-probability faults to be quickly alerted and handled promptly. This solves the problems of response delay, high false positive rate, and lack of precise guidance in handling measures in traditional diagnostic methods, enhancing the safety and maintenance efficiency of equipment operation. Attached Figure Description

[0024] Figure 1 This is a system architecture block diagram of the intelligent transmitter system proposed in this invention; Figure 2 This is a flowchart of the self-diagnosis method proposed in this invention.

[0025] Figure 3 This is a flowchart of the ResNet-8 model implementation and training process.

[0026] Figure 4 This is a diagram of the improved ResNet-8 one-dimensional convolutional neural network architecture.

[0027] Figure 5 This is a flowchart of the model training process based on supervised learning using the cross-entropy loss function.

[0028] Figure 6 This is a flowchart for edge deployment optimization. Detailed Implementation

[0029] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0030] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0031] It should be noted that when an element is referred to as being "set on" another element, it can be directly on the other element or may be interposed with another element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or may be interposed with another element. The terms "upper," "lower," "front," "rear," "left," "right," "top," "bottom," "inner," and "outer," etc., used herein to indicate orientation or positional relationships are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0033] This invention relates to a transmitter system based on multi-sensor fusion and edge AI, which has functions such as environmental adaptive compensation and real-time self-diagnosis, effectively solving problems such as parameter drift and fault response delay in industrial environments.

[0034] Combined with appendix Figure 1An industrial intelligent transmitter system based on multi-sensor fusion and edge AI includes a multi-sensor unit, a signal acquisition unit, a microcontroller, an AI coprocessor, and a dual-mode communication unit.

[0035] The multi-sensor unit includes at least one main sensor and two auxiliary sensors. The main sensor can be, but is not limited to, a pressure sensor, a flow sensor, etc., while the auxiliary sensors include, but are not limited to, temperature sensors, vibration sensors, and humidity sensors. The main sensor is used to synchronously collect the main process parameters (such as pressure and flow rate) of the target industrial process, while the auxiliary sensors are used to collect relevant environmental parameters (such as ambient temperature, vibration amplitude, and humidity).

[0036] Multiple sensing units are integrated into the transmitter body or deployed nearby to achieve collaborative sensing of multi-source data.

[0037] The signal acquisition unit includes a signal conditioning circuit and a 24-bit Σ-Δ ADC (Analog-to-Digital Converter). The signal conditioning circuit includes filtering and amplification circuits, which are used to perform preliminary filtering, noise reduction, and amplification on the raw data signals from the multiple sensing units. The 24-bit high-precision Σ-Δ ADC is used to acquire the sensor data preprocessed by the signal conditioning circuit and convert it into a digital signal for transmission to the microcontroller.

[0038] The microcontroller is a single-chip microcontroller or integrated circuit development board with data processing and communication functions, such as an STM32 microcontroller, Arduino, or 51 microcontroller. It is used to control the entire transmitter system, including data acquisition and conversion of the control signal acquisition unit, operation of the AI ​​coprocessor, and data transmission and communication of the dual-mode communication unit. At the same time, the microcontroller processes the digital signal sampled by the 24-bit Σ-Δ ADC and converted from analog to digital (such as software filtering, calibration, synchronization, quantization, etc.), and runs a multi-sensor fusion algorithm for environmental adaptive compensation.

[0039] AI coprocessors are hardware devices optimized for accelerating AI tasks, such as the K210 and Atlas 200. They are used to run neural network models, such as ResNet-8 models quantized with INT8, to extract features from sensor output signals and infer fault probabilities, thereby enabling self-diagnostic functions.

[0040] The dual-mode communication unit includes both wired and wireless communication modes. Wired communication outputs a standard 4-20mA analog signal and is compatible with the Highway Addressable Remote Transmitter (HART) protocol. Wireless communication uses LoRa WAN. The dual-mode communication unit is used to output sensor data processed by the microcontroller, as well as diagnostic results.

[0041] The output of the multi-sensor unit is connected to the input of the signal acquisition unit, which in turn is connected to the input of the microcontroller. The microcontroller is bidirectionally connected to the AI ​​coprocessor, and its output is connected to the input unit of the dual-mode communication unit. The multi-sensor unit acquires raw multi-source data; the signal acquisition unit performs preliminary filtering and amplification of the raw multi-source data, and performs analog-to-digital conversion; the microcontroller processes the converted data (software filtering, calibration, synchronization, quantization, etc.) and runs a multi-sensor fusion algorithm for environmental adaptive compensation; the AI ​​coprocessor runs a pre-built neural network model for self-diagnosis; and the dual-mode communication unit transmits data and fault information.

[0042] Specifically, the multi-sensor fusion algorithm is as follows:

[0043] In the formula, α is the temperature compensation coefficient, β is the vibration influence factor, t is the upper limit of integration (representing the time end of integration), τ is the integration variable, α(τ) is the time-domain signal function of vibration acceleration, and P raw Main sensor raw data, P corr To compensate for the main sensor data, T ref For reference temperature; Furthermore, The temperature compensation term, α, is obtained through polynomial fitting, with the specific steps as follows: Step 1: Measure the output value of the main sensor in the range of 0~100% at different temperature points. For example, measure the output of the main sensor in the range of 0%, 25%, 50%, 75%, and 100% at temperature points of -20 ℃, 0 ℃, 20 ℃, 40 ℃, and 80 ℃ respectively. Step 2: Collect output data at different temperature points and ranges, and calculate the temperature compensation term using the formula. The temperature compensation coefficient α at different points is obtained; Step 3: Establish the mapping relationship between temperature and temperature compensation coefficient α using a cubic polynomial: We obtain α at different temperatures, where T represents temperature. c 0、 c 1. c 2. c 3 is a constant coefficient.

[0044] Furthermore, α is the vibration compensation term, β is the vibration influence factor, and α(τ) is the vibration acceleration time-domain signal. β represents the cumulative vibration intensity, reflecting the cumulative effect of vibration on the sensor. β takes a smaller value at low frequencies (<100Hz), such as 0.05; a moderate value at high frequencies (>1 kHz), such as 0.1; and a larger value in the resonance region, such as 0.15. This achieves micro-compensation at low frequencies, medium compensation at high frequencies, and strong compensation in the resonance region.

[0045] This invention also provides a self-diagnosis method for an industrial intelligent transmitter system, as shown in the appendix. Figure 2 The specific implementation steps are as follows: S1: Acquire sensor data at a sampling rate of 100 Hz. The sensor data includes raw analog pressure, temperature and vibration signals. The raw analog pressure, temperature and vibration signals are filtered, amplified and converted from analog to digital to obtain digital pressure signal, digital temperature signal and digital vibration signal. S2: Perform wavelet denoising and Kalman filtering on the digital pressure signal, digital temperature signal, and digital vibration signal obtained from S1, and run a multi-sensor fusion algorithm to perform adaptive compensation for environmental factors such as temperature and vibration; S3: Extract time-frequency domain features of temperature, vibration, and compensated pressure data to generate 128-dimensional feature vectors for 3 channels (pressure, temperature, and vibration); S4: Using the 128-dimensional feature vectors of the three channels obtained in S3 as input, run the trained neural network model (ResNet-8 model) and output the fault probability. ; S5: If >0.9 triggers a local alarm, if >0.7 Upload diagnostic data.

[0046] Specifically, the implementation and training process of the ResNet-8 model is as follows: Figure 3 As shown, it includes the following steps: SA1: Sample Data Preparation: Acquire synchronous time-series data from multiple sensors (pressure, temperature, and vibration) from industrial transmitters to construct training and validation sets, including the following steps: SA1-1: Constructing the training set: Collect operating data from 20 industrial transmitters (including pressure, temperature, and vibration data in 3 channels), generating 3 (channels) × 120,000 time series segments. The distribution of each channel category is as follows: normal operating conditions: 72,000 time series segments; sensor drift: 24,000 time series segments; circuit or mechanical fault: 24,000 time series segments; process each time series segment into a 128-dimensional vector to obtain a sample.

[0047] SA1-2: Building a validation set: Collect operating data from 5 additional industrial transmitters (including pressure, temperature, and vibration data in 3 channels) to generate 3×30,000 time series segments, including power surge, EMI interference, and unknown fault modes. SA2: Preprocessing of training and validation sets: The training set preprocessing includes the following steps: SA2-1: Read 120,000 time-series segments from a single channel in the training set; SA2-2: Randomly extract 1024 sampling points from a single time segment to obtain 1 sample time segment; SA2-3: Randomly flip the timeline of the sample time segment obtained from SA2-2 with a probability of 30%; SA2-4: Perform channel normalization on the sample time segments processed by SA2-3 to obtain normalized sample time segments. The three-channel normalization formula is: z∈{pressure P, temperature T, vibration V} in This represents the value of channel z at a certain time point in the sample time segment after processing in steps SA2-3. This represents the standardized value after normalization. The z-mean value of the channels in the time segment of the sample after processing in step SA2-3. The standard deviation of channel z in the time segment of the sample after processing in step SA2-3; SA2-5: Extract the time-domain features (including mean, variance, kurtosis, margin factor, etc.) and frequency-domain features (centroid frequency, frequency variance, harmonic component energy ratio, first 5 order spectral peaks and their frequencies, etc.) of the sample time-series segments after SA2-4 normalization, and generate a 128-dimensional feature vector (64-dimensional time domain + 64-dimensional frequency domain). SA2-6: Repeat steps SA2-2 to SA2-5 to complete the preprocessing of 120,000 time series segments and obtain 120,000 128-dimensional feature vectors; simultaneously perform the above processing on 3 channels (pressure, temperature, vibration) to obtain 120,000 128×3 feature vectors, that is, obtain 120,000 training set samples. The validation set preprocessing includes the following steps: SA2-7: Read 30,000 time segments from the verification set; SA2-8: Extract three consecutive time segments with 1024 sampling points each from fixed positions (start point, 1 / 3 position point, 2 / 3 position point) of a single time segment to form a sample time segment; SA2-9: Perform channel normalization on the sample time segments obtained from SA2-8 (same as SA2-4) to obtain normalized sample time segments; SA2-10: Stack the three segments in the sample time sequence segment after SA2-9 normalization along the feature dimension to obtain a new sample time sequence segment; SA2-11: Extract the time-domain features (mean, variance, kurtosis, margin factor, etc.) and frequency-domain features (centroid frequency, frequency variance, harmonic component energy ratio, peak values ​​and frequencies of the first 5 spectral orders, etc.) of the new sample time-series segment obtained from SA2-10, and generate a 128-dimensional feature vector (64-dimensional time domain + 64-dimensional frequency domain). SA2-12: Repeat steps SA2-8 to SA2-11 to complete the preprocessing of 30,000 time series segments and obtain 30,000 128-dimensional feature vectors; simultaneously perform the above processing on 3 channels (pressure, temperature, vibration) to obtain 30,000 128×3 feature vectors, that is, obtain 30,000 validation set samples.

[0048] SA3: Model Initialization: An improved ResNet-8 one-dimensional convolutional neural network architecture is adopted, the architecture of which is as follows: Figure 4 As shown, it includes an input layer, a convolutional layer, a max pooling layer, four residual blocks, a global average pooling layer, a fully connected layer, and an output layer connected in sequence. The input layer receives a 128×3 feature vector. The input data is fed into a convolutional layer, which has three branches to process pressure, temperature, and vibration features respectively. The output of the convolutional layer undergoes batch normalization to accelerate training and improve model stability. The batch-normalized data is activated by the ReLU function, introducing non-linear features and enhancing the model's expressive power. The ReLU-activated convolutional layer output is then fed into a max-pooling layer for downsampling, reducing feature scale and computational complexity. Four residual blocks (with 16, 16, 32, and 32 channels respectively) progressively increase the abstraction level of the features. A 1×1 convolution is used to adjust the number of channels in the residual blocks. A global average pooling layer performs global average pooling on the features, compressing the features of each channel into a single value, resulting in a fixed-length feature vector. The globally average pooled feature vector is then fed into a fully connected layer for final feature mapping and classification. The output layer outputs a probability vector P, representing the probability distribution of the input data belonging to each category.

[0049] The model initialization process includes three stages: First, the weights of the convolutional layer are initialized using a He normal distribution, and the bias term is initialized to zero. The scaling factor γ of the batch normalization layer is initialized to 1, and the offset factor β is initialized to 0. Finally, all other modules are initialized using the feature extraction layer weight parameters pre-trained on the industrial fault dataset publicly available by MAFAULDA.

[0050] SA4: Model Training The preprocessed training set samples are input into the improved ResNet-8 one-dimensional convolutional neural network model for training. The training process adopts an end-to-end supervised learning approach, and its workflow is as follows: Figure 5 As shown, the specific workflow is as follows: after model initialization, the data loader is run to load data; the one-dimensional neural network architecture is run to perform forward propagation and processing of data; loss calculation is performed; after optimization by the algorithm optimization component, backpropagation is performed and parameters are updated; validation set evaluation is performed; and the optimal model is saved. The core technical features of the model training phase include: using the cross-entropy loss function to supervise the learning process, the mathematical expression of which is:

[0051] Where L is the loss function, This indicates the weights of the main and auxiliary parameters. =0.6 is the weight of the main parameter (pressure). =0.3 and =0.1 represents the weights of the auxiliary parameters (temperature and vibration), N is the batch size, c is the fault label category (1 for normal operation, 2 for sensor drift, 3 for circuit or mechanical fault), λ=0.001 is the L2 regularization coefficient, and w represents all trainable weight parameters of the model. This is the true class indicator for sample i (1 if it belongs to class c, 0 otherwise). Let i be the predicted probability that observed sample i belongs to category c; During training, the Adaptive Moment Estimation (Adam) algorithm is used for optimization, with parameters set to control momentum intensity. =0.9, controls the smoothness of the adaptive term. =0.999, initial learning rate 0.001, decaying by a factor of 0.5 every 50 epochs. To prevent gradient explosion, a gradient clipping threshold of 1.0 is set. Dropout regularization and L2 weight decay are applied after every 10 iterations.

[0052] The validation phase follows a strict procedure: After freezing all model parameters, input the preprocessed validation set samples; calculate the cross-entropy loss between the probability distribution of the model output and the real fault labels (normal operating conditions, sensor drift, circuit or mechanical faults); and simultaneously record the F1 score evaluation metric. An early stopping mechanism is used to monitor changes in the validation loss; training terminates when the loss function value decreases by less than 0.001 for five consecutive rounds. The model parameter with the highest F1 score on the validation set during the entire training process is saved as the completed model.

[0053] SA5: Model Optimization and Deployment To adapt to the resource-constrained environment of industrial transmitters, a three-level optimization was performed on the trained model. The optimization flowchart is as follows: Figure 6 Specifically, it includes the following steps: SA5-1: INT8 Quantization: Using 1,000 representative samples (dataset samples covering multiple operating conditions, such as samples containing normal operating conditions, sensor drift, circuit or mechanical failures, etc.) as the calibration set, a symmetric quantization scheme is used to convert the 32-bit floating-point (FP32) parameters and activation values ​​in the model into 8-bit integers (INT8). The symmetric quantization formula is as follows:

[0054] Where x is the FP32 value to be quantized, Q(x) is the quantized INT8 integer, and the scaling factor is... Symmetric quantization converts FP32 weights to INT8, reducing the model's storage requirements and computational complexity.

[0055] SA5-2: Layer fusion: Combines the sequence of convolution-batch normalization-ReLU operations into a single computational unit, reducing the number of memory accesses.

[0056] SA5-3: Channel pruning: Based on the L1 norm of the weights, the convolutional channels are sorted and the 30% of channels with the lowest importance are removed. This is combined with a sparse training strategy to maintain model performance.

[0057] After completing SA5-1 to SA5-3, the final model is obtained, and the final model is deployed to the AI ​​coprocessor.

[0058] SA6: Diagnostic Output and Decision-Making Mechanism The final model output is a one-dimensional probability vector. , This represents the probability under normal operating conditions. For sensor drift probability, Given the probability of circuit or mechanical failure, perform diagnostic decisions based on the output probability: when When the value is >0.9: a local alarm is triggered; when When the value is >0.7: Upload the diagnostic data package; when When the value is ≤0.7, the output is normal. The diagnostic data package includes raw sensor data, a 128-dimensional feature vector, the probability distribution of the model output, and the device status (fault label category). This design ensures the integrity and traceability of fault information.

[0059] Example 1 In industrial process monitoring applications, the intelligent transmitter system of this invention is used for pressure transmitters. The main sensor is a MEMS piezoresistive pressure sensor, the auxiliary sensors are a high-precision temperature sensor (such as PT1000) and a vibration acceleration sensor (such as ADXL345), the 24-bit Σ-Δ ADC is an ADS124S08, the microcontroller is an STM32H743VIT6 (with FPU and hardware delta function accelerator), the AI ​​coprocessor is a Kendryte K210, the LoRa module uses an SX1276 chip, and the HART modem is an AD5700.

[0060] Specifically, the main sensor (pressure sensor) collects the primary process variable, pressure, while auxiliary sensors (temperature sensor and vibration sensor) collect temperature and vibration data respectively for compensation. The pressure data is filtered and amplified by a signal conditioning circuit, then sampled by an ADC and converted from analog to digital before being transmitted to a microcontroller for processing (calibration, synchronization, quantization, etc.). Simultaneously, the microcontroller runs a multi-sensor fusion algorithm for adaptive temperature and vibration compensation. The compensated pressure data is then transmitted to a HART modem for processing, coupled to a 4-20mA analog signal, and output via a wired connection. Meanwhile, the AI ​​coprocessor K210 receives the pre-processed sensor data from the microcontroller via an SPI interface, extracts time-domain and frequency-domain features, runs a pre-built ResNet-8 model, and outputs the fault probability. ,like If the value is greater than 0.9, a local alarm will be triggered. If the value is greater than 0.7, diagnostic data is uploaded; otherwise, normal output is performed. In this embodiment, the main sensor data is output via a wired connection using a 4-20mA standard analog signal compatible with the HART protocol, while the diagnostic signal is transmitted wirelessly via LoRaWAN, achieving dual-channel separation of process signals and diagnostic data transmission.

[0061] The term "constituting of" in describing a combination should include the identified elements, components, parts, or steps, as well as other elements, components, parts, or steps that do not substantially affect the essential novel features of the combination. The use of the terms "comprising" or "including" to describe combinations of elements, components, parts, or steps herein also contemplates embodiments that are essentially composed of such elements, components, parts, or steps. The use of the term "may" herein is intended to indicate that any described attribute included by "may" is optional.

[0062] Multiple elements, components, parts, or steps can be provided by a single integrated element, component, part, or step. Alternatively, a single integrated element, component, part, or step can be divided into multiple separate elements, components, parts, or steps. The use of "a" or "an" to describe an element, component, part, or step does not imply the exclusion of other elements, components, parts, or steps.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A transmitter system based on multi-sensor fusion and edge AI, characterized in that, include: The multi-sensor unit includes a main sensor and an auxiliary sensor for synchronously acquiring raw data, which includes the main process parameters of the industrial process acquired by the main sensor and the environmental parameters acquired by the auxiliary sensor. The signal acquisition unit, connected to the multi-sensor unit, is used to filter and amplify the acquired raw data and convert it into a digital signal. A microcontroller, connected to the signal acquisition unit, is used to process the digital signal and run a multi-sensor fusion algorithm for environmental adaptive compensation to obtain compensated sensor data. The AI ​​coprocessor is connected to the microcontroller and is used to run a pre-set neural network model to extract features and infer fault probability from the compensated sensor data, thereby achieving real-time self-diagnosis. The dual-mode communication unit is connected to the microcontroller and AI coprocessor and is used to transmit diagnostic results and compensated sensor data through separate channels.

2. The transmitter system based on multi-sensor fusion and edge AI according to claim 1, characterized in that, The multi-sensor fusion algorithm is as follows: In the formula, To compensate for the main sensor data, The data consists of raw data from the main sensor, where α is the temperature compensation coefficient, β is the vibration influence factor, t is the upper limit of integration, τ is the integration variable, T is the temperature, and α(τ) is the time-domain signal function of the vibration acceleration. This is a reference temperature.

3. A transmitter system based on multi-sensor fusion and edge AI according to claim 2, characterized in that, The temperature compensation coefficient is obtained by establishing a mapping relationship between temperature and temperature compensation coefficient using a cubic polynomial. The cubic polynomial is: ,in, c 0、 c 1. c 2. c 3 is a constant coefficient.

4. The transmitter system based on multi-sensor fusion and edge AI according to claim 1, characterized in that, The main sensor is a pressure sensor or a flow sensor; the auxiliary sensor includes at least two of the following: a temperature sensor, a vibration sensor, and a humidity sensor.

5. A transmitter system based on multi-sensor fusion and edge AI according to claim 1, characterized in that, The signal acquisition unit includes a signal conditioning circuit and a 24-bit Σ-Δ analog-to-digital converter; the signal conditioning circuit includes filtering and amplification circuits for preliminary filtering, noise reduction and amplification of the original multi-source data signal; the 24-bit Σ-Δ analog-to-digital converter is used to convert the analog signal preprocessed by the signal conditioning circuit into a digital signal and transmit it to the microcontroller.

6. A transmitter system based on multi-sensor fusion and edge AI according to claim 1, characterized in that, The neural network model is an INT8 quantized ResNet-8 model, which includes an input layer, a convolutional layer, a max pooling layer, four residual blocks, a global average pooling layer, a fully connected layer, and an output layer connected in sequence. The input layer receives a 128-dimensional × 3-channel feature vector, with the three channels corresponding to pressure, temperature, and vibration data, respectively; the number of channels in the four residual blocks are 16, 16, 32, and 32, respectively; the output layer outputs the probability distributions of three types of operating conditions: normal operating conditions, sensor drift, and circuit or mechanical failure.

7. A transmitter system based on multi-sensor fusion and edge AI according to claim 1, characterized in that, The dual-mode communication unit includes both wired and wireless communication modes; the wired communication outputs a standard 4-20mA analog signal and is compatible with the HART protocol; the wireless communication outputs LoRa WAN.

8. A self-diagnostic method for an industrial intelligent transmitter, characterized in that, Includes the following steps: S1: Simultaneously acquire raw data from multiple sensors, perform signal conditioning and analog-to-digital conversion to obtain digital signals; S2: Denoise and filter the digital signal, and run a multi-sensor fusion algorithm for adaptive environmental compensation; S3: Extract the time-domain and frequency-domain features of the compensated data to generate multi-channel feature vectors; S4: Input the multi-channel feature vector into the trained ResNet-8 model and output the fault probability; S5: Execute local alarms or upload data based on the fault probability value.

9. The self-diagnostic method for an industrial intelligent transmitter according to claim 8, characterized in that, The training process of the ResNet-8 model includes the following steps: SA1: Collect synchronous three-channel time-series data of pressure, temperature and vibration from multiple industrial transmitters during operation, and construct training sets and validation sets containing known fault modes respectively; SA2: Perform data augmentation and channel normalization on the time-series data in the training and validation sets, and extract time-domain and frequency-domain features to generate corresponding multi-channel feature vector samples. SA3: Initialize the ResNet-8 model parameters. The weights of the convolutional layers are initialized using the He normal distribution. The scaling factor γ of the batch normalization layer is initialized to 1 and the offset factor β is initialized to 0. SA4: The ResNet-8 model is trained using the cross-entropy loss function and the Adam optimization algorithm. A gradient pruning threshold is set, and the adaptive moment estimation algorithm is used for optimization during training. An early stopping mechanism is implemented based on the validation set loss to save the optimal model.

10. The self-diagnostic method for an industrial intelligent transmitter according to claim 8, characterized in that, The trained ResNet-8 model underwent three levels of optimization before deployment: INT8 quantization, layer fusion, and channel pruning.

Images