An AI psychological counseling system with emotional intelligence and privacy computing dual engines

Abstract

This invention belongs to the technical field of artificial intelligence and relates to an AI psychological counseling system with dual engines of emotional intelligence and privacy computing. The system comprises: a data acquisition and fusion module that generates a spatiotemporally aligned fused input stream; an emotional state analysis module that generates a dynamic emotional state matrix; a privacy policy generation module that triggers the privacy computing engine to construct a differential privacy domain based on the dynamic emotional state matrix and generate a hierarchical data grid; a privacy recoding module that performs privacy recoding on the fused input stream based on the hierarchical data grid to generate a ciphertext package that is resistant to reverse engineering; an adaptive policy generation module that generates an adaptive counseling policy tree; and a closed-loop intervention execution module that drives terminal devices to execute feedback from the adaptive counseling policy tree to generate a closed-loop psychological intervention flow. This invention addresses the problem that most interactive modes are one-way outputs and lack a real-time capture and response mechanism for user feedback.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence and relates to an AI psychological counseling system with dual engines of emotional intelligence and privacy computing. Background Art

[0002] With the prevalence of digital lifestyles, people are increasingly demanding intelligent mental health support and emotional interaction services. However, the core challenge in achieving precise and effective digital intervention lies in accurately and comprehensively perceiving and understanding users' complex and ever-changing real-time emotional states. Human emotional expression is multi-channel and non-linear. Relying solely on single-modal data sources such as text, voice, or emojis makes it difficult to capture the nuances and underlying dynamics of emotion. This results in insufficient accuracy and reliability in emotion recognition, making it impossible to provide truly personalized feedback to users.

[0003] The commonly adopted solution in the industry is to upload user data to cloud servers for centralized processing, leveraging powerful computing resources to run complex AI models for sentiment analysis. Some wearable devices or mobile applications use simplified on-device models to analyze physiological indicators such as heart rate and step count, or perform simple facial expression recognition. While these solutions enable the collection and analysis of sentiment data to a certain extent, the data is often isolated and lacks effective fusion of multimodal information. Temporal synchronization and correlation analysis between different data streams are often overlooked.

[0004] In response to the above problems, traditional methods have significant drawbacks. Existing interaction modes are mostly one-way outputs and lack real-time capture and response mechanisms for user feedback, which greatly weakens the actual effect of the intervention. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides an AI psychological counseling system with dual engines of emotional intelligence and privacy computing.

[0006] An AI psychological counseling system with dual engines of emotional intelligence and privacy computing, including:

[0007] The data acquisition and fusion module is used to collect original user interaction data, perform real-time signal fusion at the edge through a multimodal sensor group, and generate a spatiotemporally aligned fusion input stream;

[0008] The emotional state analysis module performs multi-level emotional analysis based on the fused input stream through the emotional intelligence engine to generate a dynamic emotional state matrix;

[0009] The privacy policy generation module triggers the privacy calculation engine to build a differential privacy domain based on the dynamic emotional state matrix and generate a hierarchical data grid;

[0010] A privacy recoding module is used to perform privacy recoding on the fused input stream based on the hierarchical data grid to generate a ciphertext package that is resistant to reverse engineering;

[0011] Adaptive strategy generation module, used to synchronize the dynamic emotional state matrix and the ciphertext package to generate an adaptive tutoring strategy tree;

[0012] The closed-loop intervention execution module drives the terminal device to execute the adaptive counseling strategy tree feedback and generate a closed-loop psychological intervention flow.

[0013] A further embodiment of the present invention generates a spatiotemporally aligned fused input stream, comprising the following steps:

[0014] The microphone array of the terminal device is used to collect voice waveform data, and the camera is simultaneously activated to capture the facial expression video stream;

[0015] Use the heart rate sensor of the wearable device to obtain physiological signals, mark the timestamp and align them with the speech waveform data and facial expression video stream;

[0016] Physiological signals are segmented based on dynamic thresholds, and speech spectrum features and facial micro-movement trajectories are fused to generate a spatiotemporally aligned fused input stream.

[0017] A further solution of the present invention is to segment physiological signals based on dynamic thresholds and fuse speech spectrum features with facial micro-movement trajectories to generate a fused input stream with spatiotemporal labels, including the following steps:

[0018] Segmenting the physiological signal based on a dynamic threshold to generate a segmented physiological signal;

[0019] Extracting speech spectrum features from speech waveform data;

[0020] Analyze facial expression micro-movement trajectories in facial expression video streams;

[0021] The speech spectrum features, facial micro-movement trajectories and segmented physiological signals are integrated to generate a fused input stream with spatiotemporal labels.

[0022] A further embodiment of the present invention generates a dynamic emotional state matrix, comprising the following steps:

[0023] Extract the fundamental frequency jitter features from the speech waveform data in the fused input stream and identify the speech emotion polarity;

[0024] Analyze the eye muscle activity frequency and mouth corner displacement vector in the expression video stream fused into the input stream to quantify the emotion intensity value;

[0025] The voice emotion polarity, emotion intensity value and heart rate variation coefficient are weightedly fused to generate a dynamic emotion state matrix.

[0026] A further solution of the present invention is to perform weighted fusion of speech emotion polarity, emotion intensity value and heart rate variation coefficient to generate a dynamic emotion state matrix, including the following steps:

[0027] Get the preset weight coefficient;

[0028] Perform weighted fusion of the normalized value of speech emotion polarity, the normalized value of emotion intensity and the normalized value of heart rate variation coefficient;

[0029] Generate a dynamic emotional state matrix based on the weighted fusion results.

[0030] A further embodiment of the present invention generates a hierarchical data grid, comprising the following steps:

[0031] When the emotional intensity value of the dynamic emotional state matrix exceeds the preset threshold, the sensitive privacy label is activated;

[0032] Divide the speech waveform data in the fused input stream into publicly available segments and segments to be encrypted;

[0033] For the expression video stream of the fused input stream, detect the face area and bind the dynamic mosaic factor;

[0034] Generate hierarchical data grids based on sensitive privacy labels, including plaintext segments, generalized segments, encrypted segments, and permanently deleted segments.

[0035] A further solution of the present invention generates a reverse engineering resistant ciphertext packet, comprising the following steps:

[0036] Performing voiceprint stripping processing on the plaintext segments in the hierarchical data grid to generate voiceprint stripped data;

[0037] Inject random noise pixels into the generalized segments in the hierarchical data grid to generate a noisy video;

[0038] The encrypted segments in the hierarchical data grid are processed using a homomorphic encryption algorithm to generate a cryptographic hash sequence;

[0039] Combine voiceprint stripping data, noisy video, and cryptographic hash sequences;

[0040] The encapsulated combined data is a reverse engineering-resistant ciphertext package with a time window verification signature.

[0041] A further solution of the present invention generates an adaptive tutoring strategy tree, comprising the following steps:

[0042] Compare the similarity between the current dynamic emotional state matrix and the emotional labels in the anti-reverse engineering ciphertext package of the historical conversation;

[0043] When the voice emotion polarity is identified to be persistently negative, the cognitive behavioral therapy branch is called to generate a progressive counseling node;

[0044] If the emotion intensity value increases sharply, insert a crisis intervention node and associate it with the local emergency resource library coordinates.

[0045] A further solution of the present invention compares the similarity between the current dynamic emotional state matrix and the emotional labels in the anti-reverse engineering ciphertext package of the historical conversation, including the following steps:

[0046] Calculate the cosine similarity between the current sentiment label and the historical sentiment label;

[0047] When the similarity exceeds the preset threshold, the strategy generation is triggered.

[0048] A further solution of the present invention generates a closed-loop psychological intervention flow, comprising the following steps:

[0049] Convert the adaptive tutoring strategy tree into speech synthesis parameters and interface rendering instructions;

[0050] Output emotion-adapted voice through bone conduction headphones and generate avatar expression animation on the screen;

[0051] Calculate the change value of user feedback micro-expression and modify the dynamic emotional state matrix parameters in real time;

[0052] A closed-loop psychological intervention flow is generated based on the modified dynamic affective state matrix parameters.

[0053] In summary, the present invention has the following beneficial technical effects:

[0054] 1. By integrating multimodal data, we achieve a deep and comprehensive understanding of the user's emotional state, significantly improving the accuracy and reliability of emotion recognition. Rather than simply analyzing a single signal, we align and integrate speech acoustic features, facial micro-expression trajectories, and physiological heart rate signals in both time and space to construct a three-dimensional, multi-dimensional emotional portrait. This comprehensive analysis method captures subtle emotional cues that are often overlooked with a single modality, more realistically reflecting the user's internal emotional fluctuations and providing a solid and reliable data foundation for subsequent personalized interventions.

[0055] 2. A dynamic and refined privacy protection mechanism has been established, effectively addressing the security issues of sensitive personal information in intelligent interactions. The system intelligently determines the sensitivity level of data based on real-time analysis of user emotions and automatically triggers corresponding privacy protection strategies, such as homomorphic encryption of highly sensitive emotional fragments and stripping of identity features. This hierarchical and classified adaptive privacy computing model maximizes user data security while ensuring functional implementation, enhancing user trust and willingness to use.

[0056] 3. The intervention strategies provided are highly personalized and adaptive, generating unique counseling plans based on each user's long-term emotional patterns and immediate state changes. By comparing current and historical emotional data, the system can identify the user's habitual emotional reactions and draw on professional knowledge bases such as cognitive behavioral therapy to generate targeted counseling paths. If a sudden change in emotional state is detected, the system can also instantly switch to crisis intervention mode, truly providing intelligent support tailored to the individual and the situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. The drawings are used to provide a further understanding of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0058] Figure 1 It is a schematic diagram of the framework in the embodiment of the present application.

[0059] Figure 2 It is a schematic diagram of the process flow in the embodiment of this application. DETAILED DESCRIPTION

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0061] The following is combined with Figure 1-Figure 2 The preferred embodiments of the present invention are described in detail.

[0062] Refer to the attached Figure 1-Figure 2 , this paper proposes an AI psychological counseling system with dual engines of emotional intelligence and privacy computing, including the following modules:

[0063] The data acquisition and fusion module is used to collect original user interaction data, perform real-time signal fusion at the edge through a multimodal sensor group, and generate a spatiotemporally aligned fusion input stream;

[0064] The emotional state analysis module performs multi-level emotional analysis based on the fused input stream through the emotional intelligence engine to generate a dynamic emotional state matrix;

[0065] The privacy policy generation module triggers the privacy calculation engine to build a differential privacy domain based on the dynamic emotional state matrix and generate a hierarchical data grid;

[0066] A privacy recoding module is used to perform privacy recoding on the fused input stream based on the hierarchical data grid to generate a ciphertext package that is resistant to reverse engineering;

[0067] Adaptive strategy generation module, used to synchronize the dynamic emotional state matrix and the ciphertext package to generate an adaptive tutoring strategy tree;

[0068] The closed-loop intervention execution module drives the terminal device to execute the adaptive counseling strategy tree feedback and generate a closed-loop psychological intervention flow.

[0069] In one embodiment of the present invention, generating a spatiotemporally aligned fused input stream includes the following steps:

[0070] The microphone array of the terminal device is called to collect voice waveform data, and the camera is synchronously activated to capture the facial expression video stream; the heart rate sensor of the wearable device is used to obtain physiological signals, mark the timestamp and align it with the voice waveform and expression video stream; the physiological signals are segmented based on dynamic thresholds, and the voice spectrum features and expression micro-movement trajectories are integrated to generate a spatiotemporally aligned fusion input stream.

[0071] Multimodal sensor groups are sensors integrated into terminal devices and wearable devices, including microphone arrays for collecting voice waveforms, cameras for capturing video images, and heart rate sensors for monitoring physiological signals.

[0072] Specifically, the microphone array of the terminal device is called to collect the user's voice waveform data, the camera is synchronously activated to capture the user's facial expression video stream, and the heart rate sensor of the wearable device is used to collect the user's physiological signal data, generating the user's original interaction data consisting of voice waveform data, facial expression video stream and physiological signal data. The user's physiological signal data is marked with precise timestamps during collection. These timestamps are aligned with the time axis of the voice waveform data and facial expression video stream to ensure that all data streams are aligned in time and space, based on the millisecond-level accuracy of the device system clock. The physiological signal data is segmented based on a preset dynamic threshold. The dynamic threshold is the critical value for segmenting physiological signal data. It is calibrated based on 100 groups of healthy volunteer experimental data. For example, the critical value for segmenting physiological signal data is 100 heartbeats per minute. At the same time, it is calibrated based on the average heart rate distribution of the population to segment key physiological events.

[0073] Speech spectral features are extracted from speech waveform data. Speech spectral features are frequency and amplitude distribution parameters extracted from speech waveform data. Facial micro-movement trajectories are analyzed from facial expression video streams. Facial micro-movement trajectories are quantized sequences consisting of eye muscle activity frequencies and mouth corner displacement paths detected from facial expression video streams. Speech spectral features, facial micro-movement trajectories, and segmented physiological signal data are fused to generate a spatiotemporally aligned fused input stream. The fused input stream represents a composite data stream containing time tags and spatial coordinates.

[0074] Among them, real-time signal fusion at the edge refers to the data integration process performed in real time on the local processor of the user device to ensure low-latency processing.

[0075] For example, when a user terminal device performs data collection, it uses the terminal device's microphone array to collect speech waveform data, such as an amplitude of 0.5V. It also activates the camera to capture a facial expression video stream, such as at 30 frames per second. The heart rate sensor also captures physiological signals, such as a heart rate of 110 beats per minute. These timestamps are aligned with the timelines of the speech waveform data and facial expression video streams, such as the speech waveform start time 2024-10-01T14:00:00.000, which is aligned with the video stream start time. A dynamic threshold of 100 beats per minute is set, and the physiological signal is segmented based on the dynamic threshold of 100 beats per minute, identifying events where the heart rate exceeds the threshold.

[0076] Extract speech spectral features, such as a fundamental frequency of 200 Hz, and facial micro-movement trajectories, such as a mouth corner displacement vector of 0.2 mm / s. Fusion generates a fused input stream with spatiotemporal labels, such as the time label 2024-10-01T14:00:00.500 and the location label device coordinates (40.7128, -74.0060), to directly verify the segmentation effectiveness of the dynamic threshold and the synchronization accuracy of the spatiotemporal alignment.

[0077] In one embodiment of the present invention, generating a dynamic emotional state matrix includes the following steps:

[0078] The fundamental frequency jitter characteristics of speech waveform data are extracted to identify the speech emotion polarity as positive, neutral or negative; the eye area muscle activity frequency and mouth corner displacement vector of the expression video stream are analyzed to quantify the emotion intensity value; the speech emotion polarity, emotion intensity value and heart rate variability coefficient are weighted and fused to generate a three-dimensional dynamic emotion state matrix.

[0079] Specifically, the emotional intelligence engine receives the fused input stream generated in step S1. The emotional intelligence engine is a software module that performs multi-level emotion analysis. The fused input stream contains speech waveform data with spatiotemporal labels, an expression video stream, and physiological signal data. The fundamental frequency jitter feature is extracted from the speech waveform data in the fused input stream. The fundamental frequency jitter feature represents the variance of the fundamental frequency variation. Specifically, the fundamental frequency jitter feature is precisely defined as the variance of the cycle-by-cycle frequency variation of the fundamental frequency sequence within a small time window (e.g., 30-50 milliseconds). Based on this fluctuation, statistical analysis of historical data and expert experience is used to identify the speech emotion polarity as positive, neutral, or negative.

[0080] The emotional intelligence engine analyzes the facial expression video stream in the fused input stream and calculates the frequency of eye muscle activity, representing the cyclical movement rate of the eye muscles. Specifically, this is calculated by tracking the cyclical displacement of key points on the corners of the eyes, eyelids, and eyebrows. The coordinates of these key points in consecutive frames, particularly the vertical Y coordinates, are recorded, forming a time-varying digital signal sequence that reflects the physical displacement caused by the contraction and relaxation of eye muscles. For example, the orbicularis oculi muscle, responsible for blinking and squinting, and the frontalis muscle, responsible for raising the eyebrows. Rapid blinking produces high-frequency oscillations, while slow eyebrow raising manifests as low-frequency fluctuations. The emotional intelligence engine applies a Fast Fourier Transform (FFT) to this collected coordinate time series. The FFT decomposes this seemingly complex motion signal in the temporal dimension into its constituent frequency components. The resulting spectrum clearly displays the intensity of each frequency.

[0081] By analyzing this spectrum, we can find the frequency with the strongest energy, that is, the dominant movement frequency; objectively describe the periodic movement rate of the eye muscles within a short time window, and provide a key quantitative basis for subsequent judgment of the emotional arousal (such as tension, excitement) or authenticity (such as the squeezing of the corners of the eyes in a Duchenne smile).

[0082] The mouth corner displacement vector is a two-dimensional vector representing the change in the position of the mouth corners in the video stream, which contains both magnitude and direction components. Specifically, the real-time position of the left and right mouth corners is calculated with stable reference points such as the facial midline and the nose bottom as the origin to form a two-dimensional vector. , the amplitude represents the distance the mouth corner moves, and the direction component represents the valence of the emotion, such as upward for positive and downward for negative.

[0083] By quantifying these parameters, we can get the emotion intensity value, which is calculated by weighting the eye muscle activity frequency and the mouth corner displacement vector. The range is set to 0-10 and is calibrated based on a database of 100 facial expressions. . is the standardized frequency of eye muscle activity; is the magnitude of the mouth corner displacement vector; and is a weight coefficient obtained by regression analysis and calibration using 100 sets of standard expression sequences, aiming to maximize the consistency between the calculated results and the emotional intensity annotated by human experts.

[0084] The heart rate variability coefficient is extracted from the physiological signal data of the fused input stream. The coefficient represents a statistical indicator of the change in heart beat intervals. The heart rate variability coefficient is an indicator of the standard deviation of the heart beat intervals calculated from the physiological signal.

[0085] The voice emotion polarity, emotion intensity value and heart rate variability coefficient are weighted and fused, and the three-dimensional dynamic emotion state matrix is ​​generated using the formula: . The first row representing the dynamic affective state matrix Dimensional component; is the normalized value of the speech emotion polarity, mapped to the value range of -1 to 1, with positive as 1, neutral as 0, and negative as -1; Indicates the normalized value of the emotion intensity, ranging from 0 to 1; Indicates the normalized value of heart rate variability, ranging from 0 to 1; The results were obtained through 50 user emotion calibration experiments, using emotional self-reports as a benchmark to optimize weight distribution. The output was a three-dimensional dynamic emotion state matrix.

[0086] For example, based on the fusion input stream outputted in step S1, such as the time tag And the location tag device coordinates (40.7128, -74.0060), the emotional intelligence engine extracts the fundamental frequency jitter characteristics in the voice waveform data, such as the fluctuation value of 0.15, and combines the statistics of historical data and expert experience analysis to identify the voice emotion polarity as negative.

[0087] The eye muscle activity frequency of the expression video stream is analyzed to quantify the emotional intensity value of 7, the heart rate variation coefficient of the physiological signal is extracted to be 50ms, the voice polarity normalization value is mapped to -1, the emotional intensity value is normalized to 0.7, and the heart rate variation coefficient is normalized to 0.5. , calculate the dynamic emotional state matrix as .

[0088] In one embodiment of the present invention, generating a hierarchical data grid includes the following steps:

[0089] When the emotional intensity value in the emotional state matrix exceeds a preset threshold, the sensitive privacy label is activated; the speech waveform is divided into publicly available segments and encrypted segments, and the facial area of ​​the video stream is bound to a dynamic mosaic factor; based on the sensitive privacy label, a four-level data grid is generated, including plaintext segments, generalized segments, encrypted segments, and permanently deleted segments.

[0090] Specifically, the privacy computing engine receives the dynamic emotional state matrix generated in step S2. The dynamic emotional state matrix contains numerical components along three dimensions. The privacy computing engine checks whether the emotional intensity component of the dynamic emotional state matrix exceeds a preset threshold. The preset threshold is used to determine whether the emotional intensity triggers the critical value for privacy protection. The threshold is calibrated based on 200 sets of user privacy preference experiments. The experiment determined the privacy sensitivity of high-intensity emotions through user feedback, with a typical value of 0.6. If the emotional intensity value exceeds the preset threshold, the sensitive privacy label is activated. The sensitive privacy label is a binary state marker that, when activated, indicates that the data involves sensitive user emotions.

[0091] Based on activated sensitive privacy labels, the speech waveform data in the fused input stream is divided into publicly available segments and encrypted segments. The publicly available segments correspond to non-emotionally relevant portions, based on low-fluctuation areas of the speech spectrum; the encrypted segments are portions of the speech waveform identified as emotionally revealing, based on high-fluctuation areas of the speech spectrum. For the facial expression video stream in the fused input stream, facial regions are detected and assigned a dynamic mosaic factor, which controls the intensity of the mosaic effect. The dynamic mosaic factor is a parameter that controls the degree of mosaic blurring, ranging from 0 to 1, with 0 representing no mosaic and 1 representing complete blurring. The privacy effect was optimized based on 50 sets of face blurring experiments.

[0092] A hierarchical data grid is generated based on sensitive privacy labels. The grid includes four levels of data segments: the plaintext segment stores unprocessed data, the generalized segment applies obfuscation processing, the encrypted segment performs encryption operations, and the permanent deletion segment removes data immediately. The hierarchical data grid is output as the result.

[0093] For example, the dynamic emotional state matrix [-0.4, 0.21, 0.15] output in step S2 has its emotional intensity value component normalized to 0.7. The private calculation engine checks whether the emotional intensity value 0.7 exceeds the preset threshold of 0.6. If it does, the sensitive privacy label is activated. The speech waveform data is divided, such as the original speech waveform with an amplitude of 0.5V, into publicly available segments, such as stable segments with an amplitude of 0.2V, and encrypted segments, such as fluctuating segments with an amplitude of 0.8V. For the 30 frames per second expression video stream, the facial area is detected and a dynamic mosaic factor is bound, such as 0.8 indicating high blur intensity. A hierarchical data grid is generated based on the sensitive privacy label:

[0094] The plaintext segment contains publicly available voice segments, the generalized segment contains non-face areas of the video stream, the encrypted segment contains voice segments that need to be encrypted, and the permanently deleted segment contains highly sensitive frames of the video stream.

[0095] In one embodiment of the present invention, generating an anti-reverse engineering ciphertext package includes the following steps:

[0096] Voiceprint stripping is performed on publicly available voice clips, and random noise pixels are injected into the facial area of ​​the video stream. The fragments to be encrypted are processed using a homomorphic encryption algorithm to generate an irreversible cryptographic hash sequence. The voiceprint stripping data, noise video, and hash sequence are combined and encapsulated into a ciphertext package with a time window verification signature.

[0097] Specifically, the privacy computing engine receives the hierarchical data grid generated in step S3 and the fused input stream generated in step S1. The privacy computing engine is a software module that performs the privacy recoding function. Privacy recoding is a data processing process that applies different privacy protection operations based on the hierarchical data grid. Based on the hierarchical data grid, it contains a plaintext segment that stores unprocessed publicly available speech segments, a generalized segment that stores video stream areas that need to be blurred, an encrypted segment that stores speech segments that need to be encrypted, and a permanently deleted segment that stores removed data but is not processed in this section. Voiceprint stripping is performed on the publicly available speech segments in the plaintext segment to remove personal identity features from the speech. Voiceprint stripping is the process of removing biometric features of the speech waveform, such as the fundamental frequency pattern, and the removal accuracy is optimized based on 100 groups of speech anonymity experiments.

[0098] Random noise pixel injection is applied to the face area of ​​the video stream in the generalization segment, adding random pixel values ​​to reduce recognizability. Random noise pixel injection is an image processing technique that adds random RGB values ​​to the pixels in the face area of ​​the video stream. The noise intensity ranges from 0 to 1, with 0 representing no noise and 1 representing complete randomness. The privacy protection effect is optimized based on 50 sets of image blur experiments. Homomorphic encryption algorithms are used to process the encrypted segments in the encryption segment. Homomorphic encryption algorithms are encryption techniques, such as the Paillier algorithm, used to protect voice segments. They allow calculations to be performed on encrypted data without revealing the original content, generating encrypted output. This algorithm operates on encrypted data and generates an irreversible cryptographic hash sequence as the encryption result. The cryptographic hash sequence is the output of the encryption algorithm. It uses the SHA-256 algorithm to generate a fixed-length 256-bit digest sequence. It is irreversible and cannot restore the original data.

[0099] The combined processed voiceprint stripping data, noisy video data and cryptographic hash sequence form an integrated data set. The integrated data set is the process of aligning and merging the voiceprint stripping data, noisy video data and cryptographic hash sequence according to the time and space labels.

[0100] This dataset is packaged as a reverse-engineering-resistant ciphertext package with a time-window verification signature. The time-window verification signature is generated based on a timestamp, with a time window set to 5 seconds based on 30 sets of data integrity calibration experiments. The reverse-engineering-resistant ciphertext package is the output result, designed to be impossible to restore the original data through reverse analysis. The timestamp signature ensures data freshness and integrity.

[0101] For example, based on the hierarchical data grid output from step S3, for example, the plaintext segment contains publicly available speech segments with an amplitude of 0.2V, the generalized segment contains non-face areas of the video stream, and the encrypted segment contains speech segments requiring encryption with an amplitude of 0.8V. The privacy computing engine performs voiceprint stripping on the publicly available speech segments, such as removing fundamental frequency pattern features, and outputs voiceprint stripping data with an amplitude of 0.2V but no identity traces. Random noise pixels are injected into the facial areas of the video stream, such as adding pixel values ​​with a noise standard deviation of 10, and outputting a 70% blur effect on the noise video. The 0.8V amplitude of the segments requiring encryption is processed using a homomorphic encryption algorithm to generate a cryptographic hash sequence, such as a SHA-256 digest string.

[0102] The voiceprint stripping data, noise video, and hash sequence are combined and encapsulated into a reverse engineering-resistant ciphertext package with a time window verification signature, such as the signature timestamp 2024-10-01T14:00:05.000. This data directly verifies the effectiveness of the voiceprint stripping process to remove identity, the blurring effect of the noise pixel injection, and the irreversible nature of the cryptographic hash sequence.

[0103] In one embodiment of the present invention, generating an adaptive tutoring strategy tree includes the following steps:

[0104] The current dynamic emotional state matrix is ​​compared with the emotional labels of the ciphertext packages of historical conversations. When persistent negative polarity is identified, the cognitive behavioral therapy branch is called to generate a progressive counseling node. If the emotional intensity value increases sharply, a crisis intervention node is inserted and associated with the local emergency resource library coordinates.

[0105] Specifically, the coaching strategy generator receives the dynamic emotional state matrix output from step S2 and the anti-reverse engineering ciphertext package output from step S4. The coaching strategy generator is a software module that performs strategy construction functions. The emotional label is extracted from the current dynamic emotional state matrix, including the normalized value of the voice emotional polarity and the normalized value of the emotional intensity. The emotional label metadata attached to the anti-reverse engineering ciphertext package stored in the historical session is synchronously read. The historical session is a collection of anti-reverse engineering ciphertext packages stored in the past. Each package is attached with emotional label metadata for reading. The metadata is the unencrypted part. Calculate the similarity between the current emotional label and the historical emotional label:

[0106] in, Indicates the similarity between the current sentiment label and the historical sentiment label, Represents the current label component, Represents the historical label component, Indicates the number of label dimensions. Similarity is an indicator that quantifies label consistency. The cosine similarity formula outputs a range of -1 to 1. Identify when the speech emotion polarity is negative three times in a row and the similarity exceeds the similarity threshold. When , the cognitive behavioral therapy branch is called to generate a progressive counseling node, and the counseling steps are constructed in sequence. The experiment was calibrated based on 100 user conversations, using manually annotated emotion consistency verification. Persistent negative polarity refers to a state where the voice emotion polarity remains negative for three consecutive conversations, with a negative polarity value defined as a normalized value less than 0. The cognitive behavioral therapy branch is a pre-defined counseling strategy template that includes standardized intervention processes, such as cognitive restructuring steps. Progressive counseling nodes are sequentially generated strategy units, representing phased counseling actions.

[0107] Check whether the absolute difference between the current emotion intensity value and the previous value exceeds the steep increase threshold If it exceeds, a crisis intervention node is inserted and associated with the coordinate index stored in the local emergency resource library, and finally an adaptive counseling strategy tree is output. A sudden increase in the emotional intensity value is when the absolute change between the current emotional intensity value component and the previous value exceeds The critical state, Based on 50 sets of emotional abrupt change experiments, the experiment was based on abrupt physiological signal changes. Crisis intervention nodes are pre-set emergency response units that contain immediate response instructions. Local emergency resource library coordinates are pre-stored resource location indexes on terminal devices, such as hotline numbers.

[0108] For example, the current dynamic emotional state matrix is ​​[-0.4, 0.21, 0.15], and the extracted emotional labels are speech polarity -0.4 and emotional intensity 0.21. The historical conversation ciphertext package emotional labels are [-0.3, 0.25, 0.18], and the similarity is calculated:

[0109] molecular: ;

[0110] Denominator: ;

[0111] Similarity: .

[0112] Pre-set similarity threshold The value is 0.8. Since the similarity is 0.807, which exceeds the threshold of 0.8 and the current polarity is negative, it is judged as continuous negativity for three consecutive times. The cognitive behavioral therapy branch is called to generate a progressive counseling node such as the initial node "Guiding the Identification of Negative Thinking". The difference between the current emotional intensity of 0.21 and the previous 0.45 is 0.24. The steep increase threshold is pre-set. The output adaptive coaching strategy tree contains a single coaching node, which verifies the accuracy of the similarity formula calculation and the continuous negative judgment logic.

[0113] In one embodiment of the present invention, generating a closed-loop psychological intervention flow includes the following steps:

[0114] Convert the coaching strategy tree into speech synthesis parameters and interface rendering instructions; output emotion-adapted speech through bone conduction headphones, and generate virtual image expression animation on the screen; record the user's feedback micro-expression change values, and correct the dynamic emotion state matrix parameters in real time.

[0115] Specifically, the terminal device receives the adaptive tutoring strategy tree generated in step S5, which contains serialized tutoring nodes. Based on the adaptive tutoring strategy tree, the tutoring strategy executor parses each tutoring node and converts it into speech synthesis parameters and interface rendering instructions. The tutoring strategy executor is a software module that drives the terminal device to perform feedback functions. Conversion represents the process of mapping strategy tree nodes into executable parameters. Speech synthesis parameters represent numerical settings that control speech output, such as pitch adjustment values ​​ranging from -1 to 1, with bass being negative and treble being positive. Interface rendering instructions are command sequences that control screen animations, such as avatar expressions, including joy and sadness.

[0116] Speech synthesis parameters include pitch adjustment values ​​and speech rate control values, and interface rendering instructions include the avatar's expression type and animation frame rate. These parameters and instructions are transmitted to the bone conduction earphones and the screen display module. The bone conduction earphones output emotion-adapted speech audio through temporal bone vibrations, and the screen display module renders the avatar's expression animation. Bone conduction earphones are audio output devices that transmit sound directly to the user's inner ear through mechanical vibrations, avoiding air conduction. Emotion-adapted speech audio is speech output with adjusted pitch and speech rate, optimizing the auditory experience based on emotional state. Avatar expression animation is a sequence of digital character expression movements displayed on the screen, such as eye and mouth corner animation.

[0117] The camera continuously captures the user's facial video stream and calculates the user's feedback micro-expression change value, which is based on the amplitude change of the mouth corner displacement vector. The user's feedback micro-expression change value is obtained by analyzing the amplitude change of the mouth corner displacement vector in the camera video stream, normalized to a range of 0 to 1, and calibrated based on a database of 100 facial expressions. The dynamic emotional state matrix parameters are corrected in real time using the formula: the new parameter is equal to the original parameter plus the correction factor multiplied by the change value, that is, . represents the modified dynamic emotional state matrix parameters, represents the original parameters, Represents the user feedback micro-expression change value, is a correction factor, calibrated through 30 user feedback experiments using emotional self-reports to optimize micro-expression changes. The entire process generates a closed-loop psychological intervention flow as output, integrating voice output, animation display, and parameter correction feedback loops. The closed-loop psychological intervention flow is the output data stream, representing the complete cycle from strategy execution to feedback correction.

[0118] Exemplarily, the adaptive counseling strategy tree based on the output of step S5 includes a single counseling node such as guiding the identification of negative thinking. The counseling strategy executor converts the node into speech synthesis parameters such as a pitch adjustment value of -0.5 (indicating a low tone) and interface rendering instructions such as a virtual image sad expression animation at 24 frames per second. The speech audio is output through bone conduction headphones, such as a frequency of 200Hz, and the screen generates a virtual image expression animation such as a drooping mouth corner animation. The camera records the user's feedback micro-expression change value, such as the change in the amplitude of the mouth corner displacement vector from 0.3mm to 0.4mm, and the normalized change value The original dynamic emotional state matrix parameters are such as the voice emotional polarity component -0.4, and the modified formula is applied. The real-time correction parameter is -0.39. A closed-loop psychological intervention flow is generated, such as a speech output log, an animation frame sequence, and a corrected parameter record.

[0119] It should be noted that the formulas described above, through the principle of dimensional consistency and mathematical standardization (e.g., normalization, dimensionless parameter conversion, or unified unit system), can translate physical quantities of different attributes into unitless standard values ​​or homogeneous, superimposable parameters. This eliminates the interference of different dimensions on operational logic, ensuring that the formulas retain the distribution characteristics of the original data while maintaining mathematical rationality and adaptability to objective laws. These are merely exemplary embodiments of the present invention and are not intended to limit the scope of the invention.

[0120] The modules can be implemented in whole or in part through software, hardware, or a combination thereof, supporting hardware embedded in or independent of a processor in a computer device, and also supporting software stored in a memory in a computer device, so that the processor can call and execute operations corresponding to the modules.

[0121] It should be noted that the human body information (including but not limited to human device information and personal information, etc.) and data (including but not limited to data used for analysis, stored data and displayed data, etc.) involved in the present invention are all information and data authorized by the human body or fully authorized by all parties. The collection, use and processing of relevant data require relevant legal standards.

[0122] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. An AI psychological counseling system with dual engines of emotional intelligence and privacy computing, characterized by: include: The data acquisition and fusion module is used to collect original user interaction data, perform real-time signal fusion at the edge through a multimodal sensor group, and generate a spatiotemporally aligned fusion input stream; The emotional state analysis module performs multi-level emotional analysis based on the fused input stream through the emotional intelligence engine to generate a dynamic emotional state matrix; Extract the fundamental frequency jitter features from the speech waveform data in the fused input stream and identify the speech emotion polarity; Analyze the eye muscle activity frequency and mouth corner displacement vector in the expression video stream fused into the input stream to quantify the emotion intensity value; The voice emotion polarity, emotion intensity value and heart rate variation coefficient are weighted and fused to generate a dynamic emotion state matrix; The privacy policy generation module triggers the privacy calculation engine to build a differential privacy domain based on the dynamic emotional state matrix and generate a hierarchical data grid; When the emotional intensity value of the dynamic emotional state matrix exceeds the preset threshold, the sensitive privacy label is activated; The speech waveform data in the fused input stream is divided into publicly available segments and encrypted segments. For the expression video stream in the fused input stream, the facial region is detected and a dynamic mosaic factor is bound. Based on the sensitive privacy labels, a hierarchical data grid is generated, including plaintext segments, generalized segments, encrypted segments, and permanently deleted segments. A privacy recoding module is used to perform privacy recoding on the fused input stream based on the hierarchical data grid to generate a ciphertext package that is resistant to reverse engineering; Adaptive strategy generation module, used to synchronize the dynamic emotional state matrix and the ciphertext package to generate an adaptive tutoring strategy tree; The closed-loop intervention execution module drives the terminal device to execute the adaptive counseling strategy tree feedback and generate a closed-loop psychological intervention flow.

2. The AI ​​psychological counseling system with emotional intelligence and privacy computing dual engines according to claim 1 is characterized in that: Generating a spatiotemporally aligned fused input stream involves the following steps: The microphone array of the terminal device is used to collect voice waveform data, and the camera is simultaneously activated to capture the facial expression video stream; Use the heart rate sensor of the wearable device to obtain physiological signals, mark the timestamp and align them with the speech waveform data and facial expression video stream; Physiological signals are segmented based on dynamic thresholds, and speech spectrum features and facial micro-movement trajectories are fused to generate a spatiotemporally aligned fused input stream.

3. The AI ​​psychological counseling system with emotional intelligence and privacy computing dual engines according to claim 2 is characterized in that: Physiological signals are segmented based on dynamic thresholds, and speech spectrum features and facial micro-movement trajectories are integrated to generate a fused input stream with spatiotemporal labels. The process includes the following steps: Segmenting the physiological signal based on a dynamic threshold to generate a segmented physiological signal; Extracting speech spectrum features from speech waveform data; Analyze facial expression micro-movement trajectories in facial expression video streams; The speech spectrum features, facial micro-movement trajectories and segmented physiological signals are integrated to generate a fused input stream with spatiotemporal labels.

4. The AI ​​psychological counseling system with emotional intelligence and privacy computing dual engines according to claim 1 is characterized in that: The speech emotion polarity, emotion intensity value and heart rate variation coefficient are weightedly fused to generate a dynamic emotion state matrix, including the following steps: Get the preset weight coefficient; Perform weighted fusion of the normalized value of speech emotion polarity, the normalized value of emotion intensity and the normalized value of heart rate variation coefficient; Generate a dynamic emotional state matrix based on the weighted fusion results.

5. The AI ​​psychological counseling system with emotional intelligence and privacy computing dual engines according to claim 1 is characterized in that: Generating an anti-reverse engineering ciphertext package includes the following steps: Performing voiceprint stripping processing on the plaintext segments in the hierarchical data grid to generate voiceprint stripped data; Inject random noise pixels into the generalized segments in the hierarchical data grid to generate a noisy video; The encrypted segments in the hierarchical data grid are processed using a homomorphic encryption algorithm to generate a cryptographic hash sequence; Combine voiceprint stripping data, noisy video, and cryptographic hash sequences; The encapsulated combined data is a reverse engineering-resistant ciphertext package with a time window verification signature.

6. The AI ​​psychological counseling system with emotional intelligence and privacy computing dual engines according to claim 1 is characterized in that: Generating an adaptive tutoring strategy tree includes the following steps: Compare the similarity between the current dynamic emotional state matrix and the emotional labels in the anti-reverse engineering ciphertext package of the historical conversation; When the voice emotion polarity is identified to be persistently negative, the cognitive behavioral therapy branch is called to generate a progressive counseling node; If the emotion intensity value increases sharply, insert a crisis intervention node and associate it with the local emergency resource library coordinates.

7. The AI ​​psychological counseling system with emotional intelligence and privacy computing dual engines according to claim 6 is characterized in that: Comparing the similarity between the current dynamic emotional state matrix and the emotional labels in the anti-reverse engineering ciphertext package of the historical session includes the following steps: Calculate the cosine similarity between the current sentiment label and the historical sentiment label; When the similarity exceeds the preset threshold, the strategy generation is triggered.

8. The AI ​​psychological counseling system with emotional intelligence and privacy computing dual engines according to claim 1 is characterized in that: Generate a closed-loop psychological intervention flow, including the following steps: Convert the adaptive tutoring strategy tree into speech synthesis parameters and interface rendering instructions; Output emotion-adapted voice through bone conduction headphones and generate avatar expression animation on the screen; Calculate the change value of user feedback micro-expression and modify the dynamic emotional state matrix parameters in real time; A closed-loop psychological intervention flow is generated based on the modified dynamic affective state matrix parameters.

Images