Electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency spectrum fingerprints

Abstract

The invention discloses an electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency spectrum fingerprints, and relates to the field of electrochemical analysis, and the method comprises the following steps: initializing an electronic tongue system; joint judgment is carried out through constant-temperature stability, angle in-place and electrode stability criteria; outputting a segmented multi-frequency pulse volt-ampere excitation signal to the working electrode; a complete current response sequence is obtained through an FIFO cache period pulling mechanism; sequentially gating a plurality of working electrode channels to obtain a current response sequence of each channel; constructing a two-dimensional Mel spectrum fingerprint; based on the time-frequency spectrum feature-attention fusion model and the time-frequency spectrum feature-regression model, a qualitative recognition result and an index quantitative prediction result of the sample are obtained; and generating a contribution degree heat map in combination with a gradient weighted class activation mapping strategy, and outputting an interpretability analysis result. Through closed-loop stable scheduling, hardware channel multiplexing calibration and time-frequency spectrum fingerprint deep learning analysis, the accuracy and result credibility of the electronic tongue in complex sample detection are improved.

Description

Technical Field

[0001] This invention relates to the field of electrochemical analysis, and more specifically to an electronic tongue analysis method based on multi-frequency pulse voltammetric time-frequency fingerprinting. Background Technology

[0002] In the fields of quality control, flavor evaluation, and safety monitoring in the food industry, electronic tongues are a rapid analysis technology based on sensor arrays. By simulating the human taste perception mechanism, they can achieve qualitative differentiation and quantitative prediction of the physicochemical indicators and sensory attributes of liquid samples.

[0003] However, the measurement process of existing electronic tongue systems relies heavily on preset fixed timing sequences or the operator's experience and judgment. The output of excitation signals, the start of data acquisition, and the switching between multiple channels are often based on simple delay control, which cannot sense and respond to the actual state of the system in real time. This leads to hasty data acquisition before the temperature is balanced, the mechanical movement is completely still, or the electrode interface is stable, which introduces random errors and makes it difficult to guarantee the reliability of the measurement results when dealing with food systems with complex compositions.

[0004] Furthermore, to achieve parallel detection of multi-electrode arrays, existing technical solutions often configure independent front-end measurement circuits for each working electrode or adopt manual switching methods. The inherent differences in hardware parameters between channels, such as operational amplifier gain, zero-point drift, and wiring distribution parameters, make it difficult to achieve complete consistency. Such differences will result in the response signals of different electrode channels for the same sample not being at the same measurement standard, which undermines the foundation of multi-sensor data fusion analysis and limits the improvement of the overall array performance.

[0005] Furthermore, at the signal processing and analysis level, traditional methods typically utilize time-domain feature points or simple transformation features of the current-time response curve directly, failing to fully explore the rich dynamic information contained in multi-band pulse volt-ampere excitation. At the same time, conventional machine learning models have limited ability to model deep correlations and global contextual dependencies between features, and their decision-making process lacks intuitive physical explanations, making it difficult to form a credible chain of evidence in quality control scenarios that require high credibility.

[0006] Therefore, how to design an electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprint that can achieve intelligent closed-loop scheduling, ensure consistent measurement across multiple channels, and possess strong interpretable analysis capabilities is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides an electronic tongue analysis method based on multi-frequency pulse voltammetry time-spectrum fingerprinting, which aims to solve the problems of poor acquisition consistency, insufficient comparability of multi-channel data, and limited feature expression ability of existing electronic tongues in complex food systems; by integrating closed-loop scheduling control, calibrable multiplexed measurement hardware, and a deep learning analysis model based on time-spectrum fingerprinting, the method improves the ability to stably and reliably identify and predict the quality of complex samples.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides an electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprint, comprising the following steps: S1. Initialize the electronic tongue system, establish communication with the lower-level hardware, and complete parameter loading and channel mapping; S2. Jointly judge the stability criteria of constant temperature, angle positioning and electrode stability. When all stability criteria are met at the same time, open the signal acquisition window. S3. Within the signal acquisition window, a segmented multi-frequency pulse volt-ampere excitation signal is output to the working electrode; S4. Acquire the response current signal of the working electrode in continuous acquisition mode, and obtain the complete current response sequence through the FIFO buffer periodic pull mechanism; S5. Select multiple working electrode channels in sequence, repeat steps S3 to S4 for each channel, and set a waiting period after channel switching to obtain the current response sequence of each channel. S6. Based on the current response sequence of each channel, construct the corresponding two-dimensional Mel-frequency fingerprint; S7. Input the two-dimensional Mel-frequency fingerprint features into the pre-trained temporal-frequency feature-attention fusion model and temporal-frequency feature-regression model to obtain the qualitative identification results and quantitative prediction results of the sample, respectively. S8. Combining the gradient-weighted class activation mapping strategy, generate corresponding contribution heatmaps for the qualitative identification results and quantitative prediction results of the indicators, and output interpretability analysis results.

[0010] Preferably, in step S2, the isothermal stability criterion is the difference between the current temperature T(t) and the target temperature. The difference satisfies And continuously preset duration , This is a preset temperature tolerance threshold; The stability criterion for angle positioning is the preset vibration damping waiting time after the motion mechanism enters the positioning state. ; The electrode stability criterion is the baseline drift of the current response during the pre-sampling period before excitation begins. ,variance Both the root mean square noise (RMS) and the root mean square noise (RMS) are less than their respective preset thresholds.

[0011] Preferably, in step S3, the segmented multi-frequency pulse volt-ampere excitation signal consists of a set of discrete reference potential sequences. Generation, through the reference potential sequence By setting different repetition counts or time steps in different frequency bands, segmented excitation waveforms are formed that include at least low-frequency, mid-frequency, and high-frequency bands.

[0012] Preferably, in step S4, the FIFO cache periodic fetch mechanism includes: S41. After starting the continuous acquisition mode, periodically check the number of readable samples in the FIFO buffer. ; S42. Based on the current total number of samples required. Maximum number of reads per session Determine the amount of data read this time. ; S43, Read Each sample data point is appended and stored in the response sequence cache corresponding to the current channel; S44. Before stopping the continuous acquisition operation, repeat steps S41 to S43 until all remaining data in the FIFO buffer is read.

[0013] Preferably, in step S5, multiple working electrode channels are selected sequentially, specifically by the host computer outputting a three-bit address signal (A, B, C) to an analog multiplexer. The analog multiplexer, based on the address signals (A, B, C), inputs multiple working electrodes. Select one channel and connect it to the common terminal COM, which is connected to the input terminal of the constant potential measurement circuit.

[0014] Preferably, S6 includes: The current response sequence x[n] is framed and windowed to obtain the windowed signal of the k-th frame. ; Windowed signal for each frame Perform a short-time Fourier transform to obtain the time-frequency power spectrum P(k,f), where f is the linear frequency; Through a set of trigonometric mell filters Mapping the time-frequency power spectrum P(k,f) to the Mel frequency domain yields the Mel spectrum. , where m is the Mel band index; For the Mel spectrum Logarithmic compression and normalization are performed to obtain the final two-dimensional Mel-spectral fingerprint features. ,in Let δ be the normalization function, and let δ be the zero constant.

[0015] Preferably, in S7, the time-spectrum feature-attention fusion model includes: A multi-scale feature extraction module is used to extract features under different receptive fields from the input two-dimensional Mel-spectral fingerprint; The channel attention module is used to learn and relabel the importance weights of different feature channels; The spatial attention module is used to learn and relabel the importance weights of different spatial locations in the feature map. The classification header is used to output the probability distribution of sample categories.

[0016] Preferably, in S7, the time-spectrum feature-regression model includes: The convolutional feature extraction backbone is used to extract primary feature maps from the input two-dimensional Mel-spectral fingerprint; The global dependency modeling module is used to model long-range dependencies of the primary feature map based on the Transformer encoder structure and through a self-attention mechanism. The regression prediction header is used to output continuous predicted values ​​of the target indicator.

[0017] Preferably, in step S8, the gradient-weighted class activation mapping strategy includes: For target category c or regression task, calculate the feature map output of the last convolutional layer of the network. gradient weights :

[0018] Where H and W are the spatial height and width of the feature map, k is the model's score or output value for target c, k is the feature map index, and i and j are the spatial location indices. Gradient weights With the corresponding feature map We perform a weighted summation and activate it using the ReLU function to obtain a heatmap. :

[0019] The heat map Upsample to the same size as the input fingerprint features F(k,m) to generate a visual contribution heatmap.

[0020] Secondly, the present invention provides an electronic tongue system for implementing the above method, comprising: The electrode array module includes multiple working electrodes, a reference electrode, and a counter electrode; The constant potential excitation and measurement module is used to generate and output the segmented multi-frequency pulse volt-ampere excitation signal to the electrode, and to measure the response current of the working electrode in real time and convert it into a voltage signal; The channel selection and multiplexing module includes an analog multiplexer and its driving circuit, used for switching and selecting among multiple working electrode channels; The data acquisition and control module includes one digital-to-analog converter for outputting excitation waveforms, one analog-to-digital converter for acquiring current response signals, and multiple digital input / output interfaces for controlling peripheral devices. The motion and temperature control execution module is used to control the lifting and rotating motion of the sample stage and maintain a constant solution temperature, providing positioning status and temperature feedback signals. The host computer software module is used to execute the entire process of system initialization, stability criterion judgment, acquisition window control, excitation signal output, data acquisition, fingerprint construction, model inference, heat map generation and result storage.

[0021] As can be seen from the above technical solution, compared with the prior art, the technical solution of the present invention has the following beneficial effects: 1. This scheme ensures optimal timing of data acquisition and system automation through closed-loop scheduling driven by multi-source stability criteria. Before acquisition, it synchronously detects the stability of temperature, mechanical motion, and electrode background, and automatically opens the acquisition window only when all conditions are met simultaneously. This replaces the operation mode that relies on fixed delays or manual experience, ensuring that each measurement is performed in the optimal state and improving the reliability and repeatability of data acquisition.

[0022] 2. Multiple working electrodes are sequentially connected to the constant potential excitation and current measurement circuit, and the excitation signal and measurement channel are calibrated through a calibrable conditioning network. This eliminates the problem of inconsistent gain and bias caused by using multiple independent front-end circuits, and provides a unified benchmark data for the identification and prediction of multi-electrode fusion signals.

[0023] 3. The current response sequence is constructed as a two-dimensional Mel-time spectrum, which can integrate time-frequency coupling information under multi-band excitation. Then, qualitative classification is performed by a convolutional network with an embedded attention mechanism, and quantitative regression is performed by modeling global dependencies using the Transformer module. This allows the model to more fully explore the deep patterns in the fingerprint. In addition, an interpretable heatmap is generated by using gradient-weighted class activation mapping, which provides an intuitive basis for analysis and decision-making, enhancing the credibility of the results and the practicality of the system. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 A flowchart of an electronic tongue analysis method based on multi-frequency pulse voltammetric time-frequency fingerprinting provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the segmented multi-frequency pulse volt-ampere excitation waveform provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of two-dimensional Mel-spectral fingerprint construction provided in an embodiment of the present invention; Figure 4 A framework diagram of an electronic tongue system provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the host computer status display interface provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1; like Figure 1 As shown, this embodiment provides an electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprint, including the following steps: S1. Initialize the electronic tongue system, establish communication with the lower-level hardware, and complete parameter loading and channel mapping; S2. Jointly judge the stability criteria of constant temperature, angle positioning and electrode stability. When all stability criteria are met at the same time, open the signal acquisition window. S3. Within the signal acquisition window, a segmented multi-frequency pulse volt-ampere excitation signal is output to the working electrode; S4. Acquire the response current signal of the working electrode in continuous acquisition mode, and obtain the complete current response sequence through the FIFO buffer periodic pull mechanism; S5. Select multiple working electrode channels in sequence, repeat steps S3 to S4 for each channel, and set a waiting period after channel switching to obtain the current response sequence of each channel. S6. Based on the current response sequence of each channel, construct the corresponding two-dimensional Mel-frequency fingerprint; S7. Input the two-dimensional Mel-frequency fingerprint features into the pre-trained temporal-frequency feature-attention fusion model and temporal-frequency feature-regression model to obtain the qualitative identification results and quantitative prediction results of the sample, respectively. S8. Combining the gradient-weighted class activation mapping strategy, generate corresponding contribution heatmaps for the qualitative identification results and quantitative prediction results of the indicators, and output interpretability analysis results.

[0028] This method optimizes the timing of data acquisition through closed-loop scheduling based on multi-source stability criteria, ensures the consistency and comparability of multi-electrode data by using single measurement link multiplexing and calibrable hardware, and converts the one-dimensional response into a two-dimensional Mel-spectral fingerprint before inputting it into a dedicated deep learning model for analysis. This improves the repeatability of data acquisition and the accuracy of qualitative and quantitative analysis in complex food systems using electronic tongues, while enhancing the credibility of the results through interpretable output.

[0029] The following provides a further detailed explanation of each step in the above method; In this embodiment, S1, the electronic tongue system is initialized, communication with the lower-level hardware is established, and parameter loading and channel mapping are completed. System initialization not only establishes basic communication connections but also requires precise hardware parameter loading and channel mapping calibration. In this embodiment, the host computer communicates with the lower-level data acquisition card via USB protocol. The acquisition card needs to provide at least two 16-bit DA channels, one 16-bit AD channel, multiple digital I / O channels, and PWM outputs. The initialization process includes loading the calibration parameters of each channel and configuring the address lines (A, B, C) and working electrode channels of the analog multiplexer. The mapping relationship is used to ensure that the subsequent software control can accurately select the target electrode; in addition, a brief device self-test is performed during the initialization phase, such as checking whether the DA / AD path is unobstructed and whether the digital IO status is normal, laying the foundation for stable data acquisition in the future.

[0030] In this embodiment S2, a joint judgment is made by the constant temperature stability criterion, the angle positioning stability criterion and the electrode stability criterion. When all stability criteria are met at the same time, the signal acquisition window is opened. The isothermal stability criterion is the difference between the current temperature T(t) and the target temperature. The difference satisfies And continuously preset duration , This is a preset temperature tolerance threshold; The stability criterion for angle positioning is the preset vibration damping waiting time after the motion mechanism enters the positioning state. ; The electrode stability criterion is the baseline drift of the current response during the pre-sampling period before excitation begins. ,variance Both the root mean square noise (RMS) and the root mean square noise (RMS) are less than their respective preset thresholds.

[0031] By combining these three criteria, intelligent closed-loop scheduling of data acquisition timing is achieved. The system will only adaptively open the acquisition window when the three conditions of thermodynamic equilibrium, mechanical static state, and electrochemical steady state are met simultaneously. This avoids the problem of inconsistent initial conditions for signal acquisition caused by environmental fluctuations or unstable states, and improves the comparability of data from multiple measurements or different channels in complex food systems.

[0032] In this embodiment, S3, within the signal acquisition window, a segmented multi-frequency pulse volt-ampere excitation signal is output to the working electrode; The segmented multi-frequency pulse volt-ampere excitation signal consists of a set of discrete reference potential sequences. Generation, through the reference potential sequence By setting different repetition counts or time steps in different frequency bands, segmented excitation waveforms are formed that include at least low-frequency, mid-frequency, and high-frequency bands.

[0033] like Figure 2 As shown, the key design of the segmented multi-frequency pulsed voltammetric excitation signal lies in segmentation and multi-frequency. The excitation waveform can be generated from the same set of discrete reference potential sequences, such as 21 potential points, by changing the number of repetitions to generate segments of different frequencies: the low-frequency segment (1Hz) is repeated 100 times for each potential point, the mid-frequency segment (10Hz) is repeated 10 times, and the high-frequency segment (100Hz) is repeated once, thus forming a composite waveform with a total length of 2331 points. This allows for the simultaneous acquisition of the sample's response information to electrochemical processes at different time scales (fast, medium, and slow) with a single excitation. The excitation signal output is driven by a host computer timer, which converts the digital waveform value into an analog voltage output point by point using the Set_DA_Single function with a fixed time step. At the same time, the output progress of different frequency bands is monitored by index thresholds to ensure the timing accuracy of the excitation waveform. Furthermore, since it is generated based on the same reference sequence, there is good phase coherence between the frequency bands, which provides an ideal signal source for the subsequent construction of the time-spectrum fingerprint.

[0034] In this embodiment, S4, the response current signal of the working electrode is acquired in continuous acquisition mode, and the complete current response sequence is obtained through the FIFO buffer periodic pull mechanism. The FIFO cache periodic fetch mechanism includes: S41. After starting the continuous acquisition mode, periodically check the number of readable samples in the FIFO buffer. ; S42. Based on the current total number of samples required. Maximum number of reads per session Determine the amount of data read this time. ; S43, Read Each sample data point is appended and stored in the response sequence cache corresponding to the current channel; S44. Before stopping the continuous acquisition operation, repeat steps S41 to S43 until all remaining data in the FIFO buffer is read.

[0035] To fully capture the response generated by the multi-frequency excitation and avoid potential data loss during high-speed continuous acquisition, a zero-loss mechanism of continuous acquisition and periodic FIFO pull is adopted. When acquisition starts, the AD channel is configured to enter the continuous sampling mode triggered by the internal clock. During acquisition, the host computer periodically queries the number of readable samples in the hardware FIFO buffer through another timer, and dynamically calculates and reads a batch of data based on the current required number of samples and the upper limit of a single read, and appends it to the data buffer array on the software side. The core of this mechanism lies in the management of the acquisition lifecycle: before the stop acquisition command is issued, the cache query and read operations must be executed repeatedly to ensure that all cached data in the FIFO is read back before executing AD_Continu_Stop. This avoids data loss caused by clearing the hardware FIFO due to directly stopping acquisition, and ensures that the entire response sequence from the start to the end of the stimulus is completely recorded, providing a reliable data foundation for subsequent analysis.

[0036] In this embodiment, S5, multiple working electrode channels are selected sequentially, and steps S3 to S4 are repeated for each channel. After the channel is switched, a waiting period is set to obtain the current response sequence of each channel. In this process, multiple working electrode channels are selected sequentially, specifically by the host computer outputting a three-bit address signal (A,B,C) to an analog multiplexer. The analog multiplexer, based on the address signals (A, B, C), inputs multiple working electrodes. Select one channel and connect it to the common terminal COM, which is connected to the input terminal of the constant potential measurement circuit.

[0037] For sensor arrays containing multiple working electrodes, a channel polling strategy based on an analog multiplexer is used to achieve single-set measurement link multiplexing. The host computer outputs binary address codes (A, B, C) through three digital I / O channels to control the multiplexer. to One channel is connected to the common terminal COM, thereby connecting the working electrode of this channel to the constant potential measurement loop. By changing the address code in sequence, the response of each electrode can be collected sequentially. Furthermore, after selecting a new channel, excitation and acquisition do not begin immediately. Instead, a waiting period is set. This waiting period is used to absorb transient disturbances that may occur during analog switch switching and to allow the newly connected electrode to reach initial polarization equilibrium in the electrolyte. The minor fluctuations in the liquid surface caused by mechanical switching are also calmed down. This effectively reduces the spurious signal differences introduced by the channel switching itself, ensuring that the differences in multi-channel data mainly stem from the specific response of the electrode sensitive membrane material to different taste substances, rather than switching noise, thus improving the effectiveness of the array data.

[0038] In this embodiment, S6, a corresponding two-dimensional Mel-frequency fingerprint is constructed based on the current response sequence of each channel; including: The current response sequence x[n] is framed and windowed to obtain the windowed signal of the k-th frame. Specifically, this includes: the original sequence By window length With frame shift Perform frame segmentation to obtain the first frame. Frame signal:

[0039] And apply a window function to each frame. To suppress boundary effects and spectral leakage:

[0040] Windowed signal for each frame Perform a short-time Fourier transform to obtain the time-frequency power spectrum P(k,f), where f is the linear frequency; specifically, this includes performing an STFT on each frame of the signal to obtain the time-frequency complex spectrum.

[0041] And take the amplitude spectrum or power spectrum as the time-frequency energy representation:

[0042] This allows us to obtain a two-dimensional spectrum of time-linear frequency, preserving the dynamic variation characteristics of the current-voltage response at different scales; Through a set of trigonometric mell filters Mapping the time-frequency power spectrum P(k,f) to the Mel frequency domain yields the Mel spectrum. , where m is the Mel band index; this is essentially a band convergence, which compresses the detailed information of the high-frequency band while preserving and enhancing the distribution of low-frequency information that is more effective for classification and regression tasks; For the Mel spectrum Logarithmic compression and normalization are performed to obtain the final two-dimensional Mel-spectral fingerprint features. ,in Let δ be the normalization function, and let δ be the zero constant.

[0043] like Figure 3 As shown, Au (gold), Pt (platinum), and Ag (silver) represent typical working electrode materials in the electronic tongue sensor array. By demonstrating the response values ​​of these different electrode materials in the same time series, it illustrates how a multi-electrode array provides multi-dimensional and complementary electrochemical response information, ultimately forming a two-dimensional matrix that comprehensively reflects the response characteristics of different electrodes, i.e., Mel-spectral fingerprint. This achieves the natural fusion and compact expression of multi-frequency information. On the one hand, the fingerprint construction process avoids the engineering complexity and human error caused by traditional multi-segment signal segmentation, alignment, and splicing, improving the batch processing efficiency of the host computer. On the other hand, Mel-band convergence and logarithmic compression have stronger suppression capabilities for noise and transient disturbances, thereby improving the repeatability, distinguishability, and stability of the algorithm input of the fingerprint in complex food systems, making the output results of qualitative identification and quantitative regression more consistent and reliable.

[0044] In this embodiment, S7, the two-dimensional Mel-frequency fingerprint features are input into the pre-trained temporal-frequency feature-attention fusion model and temporal-frequency feature-regression model to obtain the qualitative identification results and quantitative prediction results of the sample, respectively. The time-spectral feature-attention fusion model includes: A multi-scale feature extraction module is used to extract features under different receptive fields from the input two-dimensional Mel-spectral fingerprint; The channel attention module is used to learn and relabel the importance weights of different feature channels; The spatial attention module is used to learn and relabel the importance weights of different spatial locations in the feature map. The classification header is used to output the probability distribution of sample categories.

[0045] Furthermore, the time-spectrum feature-regression model includes: The convolutional feature extraction backbone is used to extract primary feature maps from the input two-dimensional Mel-spectral fingerprint; The global dependency modeling module is used to model long-range dependencies of the primary feature map based on the Transformer encoder structure and through a self-attention mechanism. The regression prediction header is used to output continuous predicted values ​​of the target indicator.

[0046] For qualitative identification, the backbone network of the temporal-spectral feature-attention fusion model uses a lightweight inverse residual bottleneck convolution module to efficiently extract multi-scale spatiotemporal features. Subsequently, the network embeds channel attention module and spatial attention module, allowing the model to autonomously learn which frequency bands and time segments in the fingerprint image are more critical for distinguishing different sample categories and enhance these features. Finally, the class probability is output through global average pooling and fully connected layers. For quantitative prediction, the time-spectrum feature-regression model adopts a hybrid structure of convolutional backbone and Transformer encoder. The convolutional layer first extracts local features, and then the Transformer encoder uses its self-attention mechanism to model the long-range global dependencies between local features, thereby more accurately capturing the nonlinear mapping between complex indicators and global fingerprint patterns. Finally, the predicted value is output through the regression prediction head. The two models can be trained and deployed independently, and can be synchronously invoked by the host computer to achieve an efficient analysis process that outputs qualitative and quantitative results simultaneously in a single detection.

[0047] In this embodiment S8, by combining the gradient-weighted class activation mapping strategy, a corresponding contribution heatmap is generated for the qualitative identification results and the quantitative prediction results of the indicators, and the interpretability analysis results are output. The gradient-weighted activation mapping strategy includes: For target category c or regression task, calculate the feature map output of the last convolutional layer of the network. gradient weights :

[0048] Where H and W are the spatial height and width of the feature map, k is the model's score or output value for target c, k is the feature map index, and i and j are the spatial location indices. Gradient weights With the corresponding feature map We perform a weighted summation and activate it using the ReLU function to obtain a heatmap. :

[0049] The heat map Upsample to the same size as the input fingerprint features F(k,m) to generate a visual contribution heatmap.

[0050] The generated heatmap can intuitively identify, by color intensity, which time intervals and frequency components in the original current response sequence contribute most to the model's current classification or regression decision. For example, the heatmap can show that the model mainly distinguishes between two types of juice based on the response of a certain time window in the mid-frequency band. This interpretable output helps users understand and verify the model's decision-making basis, providing visual analytical evidence in scenarios such as food quality traceability and process optimization.

[0051] The electronic tongue analysis method based on multi-frequency pulse voltammetric time-frequency fingerprinting provided in this embodiment ensures the consistency of data acquisition through strict stability criteria. Combined with the construction of multi-frequency response fingerprints based on Mel-scale unified characterization, it fully explores the time-frequency information in electrochemical response. Furthermore, by introducing attention mechanisms and interpretability tools, it enhances the reliability of results while improving analytical performance, providing reliable technical support for rapid and reliable sensory evaluation of complex food systems.

[0052] Example 2; like Figure 4 As shown, this embodiment provides an electronic tongue system for implementing the method of the above embodiment, comprising: The electrode array module includes multiple working electrodes, a reference electrode, and a counter electrode; The constant potential excitation and measurement module is used to generate and output the segmented multi-frequency pulse volt-ampere excitation signal to the electrode, and to measure the response current of the working electrode in real time and convert it into a voltage signal; The channel selection and multiplexing module includes an analog multiplexer and its driving circuit, used for switching and selecting among multiple working electrode channels; The data acquisition and control module includes one digital-to-analog converter for outputting excitation waveforms, one analog-to-digital converter for acquiring current response signals, and multiple digital input / output interfaces for controlling peripheral devices. The motion and temperature control execution module is used to control the lifting and rotating motion of the sample stage and maintain a constant solution temperature, providing positioning status and temperature feedback signals. The host computer software module is used to execute the entire process of system initialization, stability criterion judgment, acquisition window control, excitation signal output, data acquisition, fingerprint construction, model inference, heat map generation and result storage.

[0053] like Figure 5As shown, the host computer status display interface displays the core monitoring area of ​​the software, which is divided into two parts: the upper part is the excitation signal form display area, which is used to draw and monitor the segmented multi-frequency pulse volt-ampere voltage waveform output to the electrode system in real time; the lower part is the signal acquisition display area, which is used to draw and display the current response voltage signal acquired from the working electrode in a synchronous manner. The two display areas share the same horizontal axis of time, which makes the excitation and response strictly aligned in time sequence, so as to intuitively present the output and feedback status of the system, making it convenient for the operator to monitor the correctness of the measurement process and the signal quality in real time.

[0054] The interface comprises the following key areas: an excitation signal waveform display area (for monitoring the output of segmented multi-frequency pulse volt-ampere signals), a real-time current response curve display area, a historical or real-time response curve overlay display area for multiple working electrode channels (sensors), and a comprehensive system status display bar (for indicating temperature, motion mechanism status, currently selected channel, acquisition progress, and stability criterion fulfillment status, etc.). This interface is the core for realizing human-machine interaction, process monitoring, and data display.

[0055] The constant potential excitation and measurement module in the system specifically includes: The voltage conditioning unit is configured to receive the raw excitation voltage signal from the digital-to-analog converter. It then performs proportional scaling and bias adjustment to output a set potential signal that meets the requirements of the constant potential circuit. ; The constant potential control unit uses the set potential signal. Potential sampled in real time with reference electrode The difference As an error signal, after amplification, it drives the counter electrode, making... track ; The transimpedance amplification measurement unit measures the current flowing through the working electrode. The signal is converted into an output voltage signal through a transimpedance feedback structure. ,in As a reference bias voltage, This is the transimpedance gain resistor.

[0056] This system serves as the physical carrier for realizing the aforementioned analytical methods, techniques, and advantages. Through modular hardware design, it achieves core functions such as excitation output, constant potential control, multiplexed measurement, temperature control, and motion execution. Furthermore, through integrated host computer software, it intelligently schedules and analyzes the entire process, ensuring full automation and high reliability from signal excitation, acquisition, processing to result output.

[0057] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0058] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprinting, characterized in that, Includes the following steps: S1. Initialize the electronic tongue system, establish communication with the lower-level hardware, and complete parameter loading and channel mapping; S2. Jointly judge the stability criteria of constant temperature, angle positioning and electrode stability. When all stability criteria are met at the same time, open the signal acquisition window. S3. Within the signal acquisition window, a segmented multi-frequency pulse volt-ampere excitation signal is output to the working electrode; S4. Acquire the response current signal of the working electrode in continuous acquisition mode, and obtain the complete current response sequence through the FIFO buffer periodic pull mechanism; S5. Select multiple working electrode channels in sequence, repeat steps S3 to S4 for each channel, and set a waiting period after channel switching to obtain the current response sequence of each channel. S6. Based on the current response sequence of each channel, construct the corresponding two-dimensional Mel-frequency fingerprint; S7. Input the two-dimensional Mel-frequency fingerprint features into the pre-trained temporal-frequency feature-attention fusion model and temporal-frequency feature-regression model to obtain the qualitative identification results and quantitative prediction results of the sample, respectively. S8. Combining the gradient-weighted class activation mapping strategy, generate corresponding contribution heatmaps for the qualitative identification results and quantitative prediction results of the indicators, and output interpretability analysis results.

2. The electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprinting according to claim 1, characterized in that, In S2, the isothermal stability criterion is the difference between the current temperature T(t) and the target temperature. The difference satisfies And continuously preset duration , This is a preset temperature tolerance threshold; The stability criterion for angle positioning is the preset vibration damping waiting time after the motion mechanism enters the positioning state. ; The electrode stability criterion is the baseline drift of the current response during the pre-sampling period before excitation begins. ,variance Both the root mean square noise (RMS) and the root mean square noise (RMS) are less than their respective preset thresholds.

3. The electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprinting according to claim 1, characterized in that, In S3, the segmented multi-frequency pulse volt-ampere excitation signal consists of a set of discrete reference potential sequences. Generation, through the reference potential sequence By setting different repetition counts or time steps in different frequency bands, segmented excitation waveforms are formed that include at least low-frequency, mid-frequency, and high-frequency bands.

4. The electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprinting according to claim 1, characterized in that, In S4, the FIFO cache periodic fetch mechanism includes: S41. After starting the continuous acquisition mode, periodically check the number of readable samples in the FIFO buffer. ; S42. Based on the current total number of samples required. Maximum number of reads per session Determine the amount of data read this time. ; S43, Read Each sample data point is appended and stored in the response sequence cache corresponding to the current channel; S44. Before stopping the continuous acquisition operation, repeat steps S41 to S43 until all remaining data in the FIFO buffer is read.

5. The electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprinting according to claim 1, characterized in that, In S5, multiple working electrode channels are selected sequentially, specifically by the host computer outputting a three-bit address signal (A,B,C) to the analog multiplexer. The analog multiplexer, based on the address signals (A, B, C), inputs multiple working electrodes. Select one channel and connect it to the common terminal COM, which is connected to the input terminal of the constant potential measurement circuit.

6. The electronic tongue analysis method based on multi-frequency pulse volt-ampere-time spectral fingerprinting according to claim 1, characterized in that, S6 includes: The current response sequence x[n] is framed and windowed to obtain the windowed signal of the k-th frame. ; Windowed signal for each frame Perform a short-time Fourier transform to obtain the time-frequency power spectrum P(k,f), where f is the linear frequency; Through a set of trigonometric mell filters Mapping the time-frequency power spectrum P(k,f) to the Mel frequency domain yields the Mel spectrum. , where m is the Mel band index; For the Mel spectrum Logarithmic compression and normalization are performed to obtain the final two-dimensional Mel-spectral fingerprint features. ,in Let δ be the normalization function, and let δ be the zero constant.

7. The electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprint according to claim 1, characterized in that, In S7, the time-spectrum feature-attention fusion model includes: A multi-scale feature extraction module is used to extract features from different receptive fields from the input two-dimensional Mel-spectral fingerprint; The channel attention module is used to learn and relabel the importance weights of different feature channels; The spatial attention module is used to learn and relabel the importance weights of different spatial locations in the feature map. The classification header is used to output the probability distribution of sample categories.

8. The electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprint according to claim 1, characterized in that, In S7, the time-spectrum feature-regression model includes: The convolutional feature extraction backbone is used to extract primary feature maps from the input two-dimensional Mel-spectral fingerprint; The global dependency modeling module is used to model long-range dependencies of the primary feature map based on the Transformer encoder structure and through a self-attention mechanism. The regression prediction header is used to output continuous predicted values ​​of the target indicator.

9. The electronic tongue analysis method based on multi-frequency pulse volt-ampere time-frequency fingerprint according to claim 1, characterized in that, In S8, the gradient-weighted activation mapping strategy includes: For target category c or regression task, calculate the feature map output of the last convolutional layer of the network. gradient weights : Where H and W are the spatial height and width of the feature map, k is the model's score or output value for target c, k is the feature map index, and i and j are the spatial location indices. Gradient weights With the corresponding feature map We perform a weighted summation and activate it using the ReLU function to obtain a heatmap. : The heat map Upsample to the same size as the input fingerprint features F(k,m) to generate a visual contribution heatmap.

10. An electronic tongue system for implementing the method of any one of claims 1 to 9, characterized in that, include: The electrode array module includes multiple working electrodes, a reference electrode, and a counter electrode; The constant potential excitation and measurement module is used to generate and output the segmented multi-frequency pulse volt-ampere excitation signal to the electrode, and to measure the response current of the working electrode in real time and convert it into a voltage signal; The channel selection and multiplexing module includes an analog multiplexer and its driving circuit, used for switching and selecting among multiple working electrode channels; The data acquisition and control module includes one digital-to-analog converter for outputting excitation waveforms, one analog-to-digital converter for acquiring current response signals, and multiple digital input / output interfaces for controlling peripheral devices. The motion and temperature control execution module is used to control the lifting and rotating motion of the sample stage and maintain a constant solution temperature, providing positioning status and temperature feedback signals. The host computer software module is used to execute the entire process of system initialization, stability criterion judgment, acquisition window control, excitation signal output, data acquisition, fingerprint construction, model inference, heat map generation and result storage.

Images